Daily Papers

How do multimodal models solve visual spatial tasks -- through genuine planning, or through brute-force search in token space? We introduce MazeBench, a benchmark of 110 procedurally generated maze images across nine controlled groups, and evaluate 16 model configurations from OpenAI, Anthropic, Google, and Alibaba. GPT-5.4 solves 91\% and Gemini 3.1 Pro 79\%, but these scores are misleading: models typically translate images into text grids and then enumerate paths step by step, consuming 1,710--22,818 tokens per solve for a task humans do quickly. Without added reasoning budgets, all configurations score only 2--12\%; on 20times20 ultra-hard mazes, they hit token limits and fail. Qualitative traces reveal a common two-stage strategy: image-to-grid translation followed by token-level search, effectively BFS in prose. A text-grid ablation shows Claude Sonnet 4.6 rising from 6\% on images to 80\% when given the correct grid, isolating weak visual extraction from downstream search. When explicitly instructed not to construct a grid or perform graph search, models still revert to the same enumeration strategy. MazeBench therefore shows that high accuracy on visual planning tasks does not imply human-like spatial understanding.

Scientific discovery is an inherently creative and uncertain process, requiring reasoning beyond the recall of known knowledge. While many benchmarks have been proposed to evaluate large language model (LLM) performance on deep research tasks via multi-hop retrieval, their innovative reasoning abilities essential for true scientific discovery remain largely untested. We introduce a benchmark framework for evaluating model performance in scientific discovery and reasoning, building up from a raw problem to the classical null hypothesis test. In our framework, models initially receive only the topic and research question from a recent paper, with technical details progressively revealed. At each stage of information disclosure, the model is tasked with generating hypotheses that address the research question, which is compared with the conclusions from the original paper and evaluated via automated semantic similarity of constituent atomic claims. This progressive evaluation of semantic divergence from ground-truth conclusions enables assessment of a model's innovativeness (under minimal information) to grounded reasoning capabilities (under full experimental details), both critical for using LLMs for scientific discovery purposes. Our framework provides a foundation for systematically evaluating scientific reasoning and discovery capabilities in LLMs, crucial for advancing the development of next-generation AI scientist/co-scientist systems. Specifically, here we evaluate GPT-5, GPT-5.4, Gemini 2.5 pro, and Gemini 3.1 pro preview across 45 papers spanning bioactive materials, mechanical materials, and nanomaterials. We find that GPT-5.4 and Gemini 3.1 pro outperform their previous generation counterparts as expected, and GPT-5.4 in particular maintains 0.7 F1 score alignment with ground truth conclusions even under minimal context.

LLM agents are evolving from conversational chatbots to operational tools in real-world workspaces. In local agentic harnesses, an LLM can read and write files, call tools, and reuse workspace state across sessions. While such capabilities enhance utility, they also expose a new attack surface for attackers. Attackers can embed a prompt injection within a file or tool output. Agents may read this hidden instruction, store it, and execute it later. In this multi-step trojan attack paradigm, no individual step appears malicious on its own, but these steps can collectively turn untrusted text into persistent control content. However, existing defenses often inspect each step in isolation. As a result, they can block a clear harmful action, but fail to detect the earlier write operation that plants the backdoor. To reveal this threat, we introduce ClawTrojan, a benchmark designed to identify multi-step trojan attacks in local agentic harnesses. In an OpenClaw-style simulated workspace with GPT-5.4, ClawTrojan reaches a 95.5% attack success rate (ASR), while existing single-turn prompt-injection attacks produce near-zero ASR on the same model. To address this threat, we propose DASGuard, which scans control-like text in sensitive local files, traces its origin, and removes control content that does not originate from a trusted source. Our results show that DASGuard achieves strong dynamic defense by combining runtime attack blocking with sanitized commits to the workspace.

LLM agents handle user requests on behalf of organizations through tool calls and must follow the company policies stated in their system prompts. Prior work approaches this as a safeguarding problem -- external checks that block non-compliant agent actions. We argue that policy adherence is a broader problem: real workflows unfold across many turns, require explicit user confirmation and prerequisite reads, and hinge on the content of the dialogue rather than on any single argument value. Meeting this bar requires (i) full conversation context, (ii) self-reasoning over the policy and the current dialogue, and (iii) conversation-specific remediation that guides the agent's next turn -- three capabilities that prior safeguard work has often underestimated. We introduce POLICYGUARD, a sub-agent verifier that shares the agent's view of the dialogue, reasons over the policy in context, and provides actionable feedback for the agent's next turn. On tau^2-BENCH airline across three vendors (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro) with four trials per setting, POLICYGUARD improves PASS4 by +12.0 / +6.0 / +12.0 pp. Per-call analyses show POLICYGUARD achieves higher policy-violation recall while blocking roughly half as often as argument-level guards.

Most reported gains on agentic-LLM clinical benchmarks are often attributed to prompt engineering, yet our results suggest that larger improvements can come from architectural and engine-level design. We present MDIA, a Multi-agent Diagnostic Intelligence Agent implemented as a 7-node specialty-routed clinical reasoning graph, on the full HealthBench Professional benchmark (n = 525), on a non-fine-tuned LLM. MDIA achieves 0.6272 under OpenAI's GPT-5.4-2026-03-05, which is +3.72 pp above the performance of OpenAI's ChatGPT for Clinicians. The experimental work shows that performance lift is attributable to system architecture: specialty routing, multi-turn context preservation, drug-state safety gating, site-filtered search, length-aware synthesis, and engine-level reliability. These findings support the view that agentic clinical benchmark performance is shaped both by the underlying foundation model and the orchestration architecture. Nevertheless, we also noticed notable differences when using other models as a grader; in particular, when using Gemini 2.5 Pro, MDIA scored 0.6585, which suggests that the choice of grader is a source of variability. Robust evaluation of LLMs would therefore require assessment across several independent grader models.

Forecasting when AI systems will become capable of meaningfully accelerating AI research is a central challenge for AI safety. Existing benchmarks measure broad capability growth, but may not provide ample early warning signals for recursive self-improvement. We propose measuring AI's capability to autonomously implement end-to-end machine learning pipelines from past AI research breakthroughs, given a minimal task description. By providing a concise task description instead of the full prior work as reference, we hope to better elicit emerging AI research taste. We introduce a proof-of-concept benchmark in which frontier coding agents autonomously implement an AlphaZero-style machine learning pipeline for Connect Four on consumer hardware within a three-hour budget, and we evaluate the resulting game AIs in a round-robin tournament anchored to the Pascal Pons Connect Four solver. Across four agents with eight trials each, we find substantial differentiation: Claude Opus 4.7 won as first-mover against Pons in seven of eight trials, statistically significantly better than the other agents tested, none of which exceeded two of eight. The task, which no frontier agent could reliably complete when we began development in January of 2026, is now near-saturation. Our evaluation also surfaced anomalous behavior in GPT-5.4, which consistently used far less of its allocated time budget than other agents. A follow-up 16-trial probe using shorter, less evaluation-coded prompts substantially increased GPT-5.4's time-budget usage, consistent with but not diagnostic of sandbagging; Bradley-Terry ratings across probe conditions showed only directional differences, despite significant differences in time-budget usage. We release our data, code, and prompts to support reproduction and extension.

Computer Use Agents (CUAs) fundamentally rely on graphical user interface (GUI) grounding to translate language instructions into executable screen actions, but editing-level grounding in dense coding interfaces, where sub-pixel accuracy is required to interact with dense IDE elements, remains underexplored. Existing approaches typically rely on single-shot coordinate prediction, which lacks a mechanism for error correction and often fails in high-density interfaces. In this technical report, we conduct an empirical study of pixel-precise cursor localization in coding environments. Instead of a single-step execution, our agent engages in an iterative refinement process, utilizing visual feedback from previous attempts to reach the target element. This closed-loop grounding mechanism allows the agent to self-correct displacement errors and adapt to dynamic UI changes. We evaluate our approach across GPT-5.4, Claude, and Qwen on a suite of complex coding benchmarks, demonstrating that multi-turn refinement significantly outperforms state-of-the-art single-shot models in both click precision and overall task success rate. Our results suggest that iterative visual reasoning is a critical component for the next generation of reliable software engineering agents. Code: https://github.com/microsoft/precision-cua-bench.

Multimodal large language models (MLLMs) are increasingly deployed as embodied planners in egocentric environments, where task success requires not only achieving instructed goals but also acting in socially appropriate ways. While explicit goals may render certain actions optimal, implicit social norms often impose hidden constraints. Existing evaluations typically focus on explicit goal achievement or direct norm knowledge, seldom assessing whether planners can infer and apply these hidden constraints within action sequences. We introduce NormAct, a benchmark for embodied social-norm interactions that evaluates plans on Goal Achievement, Norm Compliance, and overall Task Success. NormAct uniquely embeds hidden norms within ordinary tasks, testing whether models can realize them without explicit instruction. Experiments with state-of-the-art MLLMs (GPT-5.4, Claude Opus 4.7, Gemini 3 Pro) reveal a significant gap: models achieve explicit goals in 67.3\% of cases, but comply with hidden norms in only 26.4\%. Cue-condition experiments indicate that this gap stems not from a lack of general social knowledge, but from challenges in activating and grounding relevant norms in context. To address this, we propose NormPerceptor, a context-conditioned cue generator that infers scene-relevant norms prior to planning, increasing Task Success from 24.2\% to 46.7\%. Our results underscore the importance of enabling embodied agents to proactively detect hidden norms, ground them in visual evidence, and integrate them as action-planning constraints. Our benchmark is publicly available at https://huggingface.co/datasets/Caleb196x/NormAct.

Production LLM deployments increasingly maintain heterogeneous model pools spanning order-of-magnitude cost differences. Existing routers make binary strong-vs-weak decisions and couple learned parameters to specific model identities, requiring retraining whenever the catalog changes. We present HyDRA (Hybrid Dynamic Routing Architecture), a framework that predicts fine-grained, multi-dimensional capability requirements per query and matches them against configuration-defined model profiles via shortfall matching. A ModernBERT encoder with K=4 independent sigmoid heads scores each query along reasoning, code generation, debugging, and tool use; a shortfall-matching algorithm then selects the cheapest model whose capabilities meet the predicted requirements. The deployed predictor runs at 86 ms median CPU inference latency in production, and is fully decoupled from the model catalog -- adding or removing models requires only a configuration change, with zero retraining. On SWE-Bench Verified (5-model pool: GPT-5.4-mini, Claude Haiku 4.5, GPT-5.3 Codex, Claude Sonnet 4.6, GPT-5.4), HyDRA's tunable shortfall threshold spans three regimes: peak-quality exceeds the always-strong Claude Sonnet 4.6 baseline (75.4% vs. 74.2% resolution) at 12.9% cost savings; iso-quality matches Sonnet at 54.1% cost savings, a 6x improvement over our prior in-house binary router at 9.1%; aggressive pushes savings to 72.5% for a 3.2-point quality trade. Results generalize across LiveCodeBench, BigCodeBench, and tau-bench. HyDRA is deployed to all users in GitHub Copilot's VS Code Chat auto-mode and -- to our knowledge for the first time in the LLM routing literature -- demonstrates language-invariant routing across CJK, European, and other script families.

We give a computer-assisted counterexample to the open question, posed by Rudin, Schapire, and Daubechies in COLT 2012, of whether exhaustive AdaBoost always converges to a finite cycle. The construction is based on a block-product gadget whose two factors share an exact period-2 orbit for their 5-step branch maps, but whose linearized return maps have dominant eigenvalues with an irrational logarithmic ratio. This irrationality forces the burst-winner sequence to have an irrational asymptotic frequency, precluding eventual periodicity. All assertions are certified by exact rational arithmetic. This work was developed in collaboration with GPT-5.4 Pro and Claude Opus 4.6.

As large language models (LLMs) become the engine behind conversational systems, their ability to reason about the intentions and states of their dialogue partners (i.e., form and use a theory-of-mind, or ToM) becomes increasingly critical for safe interaction with potentially adversarial partners. We propose a novel privacy-themed ToM challenge, ToM for Steering Beliefs (ToM-SB), in which a defender must act as a Double Agent to steer the beliefs of an attacker with partial prior knowledge within a shared universe. To succeed on ToM-SB, the defender must engage with and form a ToM of the attacker, with a goal of fooling the attacker into believing they have succeeded in extracting sensitive information. We find that strong frontier models like Gemini3-Pro and GPT-5.4 struggle on ToM-SB, often failing to fool attackers in hard scenarios with partial attacker prior knowledge, even when prompted to reason about the attacker's beliefs (ToM prompting). To close this gap, we train models on ToM-SB to act as AI Double Agents using reinforcement learning, testing both fooling and ToM rewards. Notably, we find a bidirectionally emergent relationship between ToM and attacker-fooling: rewarding fooling success alone improves ToM, and rewarding ToM alone improves fooling. Across four attackers with different strengths, six defender methods, and both in-distribution and out-of-distribution (OOD) evaluation, we find that gains in ToM and attacker-fooling are well-correlated, highlighting belief modeling as a key driver of success on ToM-SB. AI Double Agents that combine both ToM and fooling rewards yield the strongest fooling and ToM performance, outperforming Gemini3-Pro and GPT-5.4 with ToM prompting on hard scenarios. We also show that ToM-SB and AI Double Agents can be extended to stronger attackers, demonstrating generalization to OOD settings and the upgradability of our task.

As recommender systems transition toward agentic, multi-turn conversational interfaces, evaluation paradigms have struggled to keep pace. Current benchmarks often rely on "LLM-as-a-judge" evaluations, which introduce subjectivity, high costs and inconsistency. We present τ-Rec, a benchmark for agentic recommender systems that replaces subjective evaluation with verifiable rewards and a reveal-tagged elicitation (RTE) mechanism that controls how task constraints surface during dialogue. By testing agents against structured catalog predicates and employing a pass^k reliability metric, τ-Rec provides a systematic test for consistent reasoning. Our evaluation of nine configurations across five model families -- GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Flash, DeepSeek V4 Flash, Qwen3-32B and GPT-5 mini -- reveals a steep reliability cliff, where even the best model achieves only ~57% at pass^1 and ~38% at pass^4, highlighting a critical gap in current conversational agent deployment. All code and data are publicly available at https://github.com/nbharaths/tau-rec.

Large language models show promise for automated CUDA programming, however even the strongest coding models (e.g., Claude-Opus-4.6) may still fall short of expert-level, architecture-aware optimization. We introduce CUDAHercules, a benchmark that evaluates generated CUDA against end-to-end human-expert SOTA systems. It spans single kernels, module-level operators, full applications, and unsolved challenge tasks across Ampere, Hopper, and Blackwell GPUs, with end-to-end tasks gated by domain-specific semantic validators. Evaluating models such as Claude-Opus-4.6 and GPT-5.4 shows a large gap between runnable CUDA and expert CUDA engineering: models often compile and pass tests, but rarely recover the optimization strategies needed to match expert performance. Application semantics further reduce success, and iterative or tool-augmented feedback can improve correctness while drifting toward slow fallback implementations. These results show that automated CUDA programming remains far from fully solved and requires stronger hardware reasoning, better tool use, and training objectives that connect code understanding to hardware architecture-grounded intelligence.

This report presents MARS, short for Multimodal Agentic Reasoning with Source selection, our system for the CASTLE Challenge at EgoVis 2026. Participants must answer 185 closed-form questions over the CASTLE 2024 dataset. In contrast to prior single-video egocentric benchmarks, CASTLE requires reasoning over four days of activity, 15 synchronized perspectives, official transcripts, and multiple auxiliary modalities, including personal photos, auxiliary videos, gaze, thermal imagery, and heartrate measurements. MARS therefore treats the task as an agentic evidence-selection problem over multimodal sources rather than a purely text-only pipeline. MARS first follows the official CASTLE directory organization to build evidence memories from two primary sources, videos and transcripts, and four auxiliary sources, gaze, heartrate, photos, and thermal imagery. Long videos are converted into captions and DeepSeek-based summaries only because CASTLE videos are too long to fit directly into the model context for every question; this step compresses temporal evidence while keeping photos and other auxiliary media available as source-specific evidence. At inference time, a GPT-5.4 decision agent repeatedly chooses whether to continue reasoning, request a specific missing modality, produce an answer, or fall back to a random option when the evidence remains insufficient. The resulting system achieved second place on the final CASTLE Challenge leaderboard. Our codes are available at https://github.com/Hyu-Zhang/MARS.

LLM agents increasingly operate in large tool ecosystems, where real-world tasks require discovering relevant tools, inferring implicit sub-goals, and adapting to dynamic environments over long horizons. However, existing benchmarks rarely evaluate planning under retrieval-limited tool visibility. To address this gap, we introduce PlanBench-XL, an interactive benchmark of 327 retail tasks over 1,665 tools that tests whether agents can iteratively retrieve usable tools, invoke them to uncover intermediate evidence for subsequent calls toward the final goal. PlanBench-XL further features an optional blocking mechanism that simulates real-world unpredictability through missing, failing, or distracting tool functions, forcing agents to detect disrupted paths and adapt at runtime. Experiments on ten leading LLMs show that massive-tool planning remains challenging: while GPT-5.4 achieves 51.90% accuracy in block-free settings, it collapses to 11.36% under the most severe blocking condition. Further analysis shows that agents are especially vulnerable when failures lack explicit error signals or when recovery requires longer alternative tool-use paths. These results establish PlanBench-XL as a testbed for diagnosing agentic planning failures and highlight the need for robust adaptive planning in long-horizon tasks with large, imperfect tool environments.

Accurately understanding the intent behind speech, conversation, and writing is crucial to the development of helpful Large Language Model (LLM) assistants. This paper introduces IntentGrasp, a comprehensive benchmark for evaluating the intent understanding capability of LLMs. Derived from 49 high-quality, open-licensed corpora spanning 12 diverse domains, IntentGrasp is constructed through source datasets curation, intent label contextualization, and task format unification. IntentGrasp contains a large-scale training set of 262,759 instances and two evaluation sets: an All Set of 12,909 test cases and a more balanced and challenging Gem Set of 470 cases. Extensive evaluations on 20 LLMs across 7 families (including frontier models such as GPT-5.4, Gemini-3.1-Pro, and Claude-Opus-4.7) demonstrate unsatisfactory performance, with scores below 60% on All Set and below 25% on Gem set. Notably, 17 out of 20 tested models perform worse than a random-guess baseline (15.2%) on Gem Set, while the estimated human performance is ~81.1%, showing substantial room for improvement. To enhance such ability, this paper proposes Intentional Fine-Tuning (IFT), which fine-tunes the models on the training set in IntentGrasp, yielding significant gains of 30+ F1 points on All Set and 20+ points on Gem Set. Tellingly, the leave-one-domain-out (Lodo) experiments further demonstrate the strong cross-domain generalizability of IFT, verifying that it is a promising approach to substantially enhancing the intent understanding of LLMs. Overall, by benchmarking and boosting intent understanding ability, this study sheds light on a promising path towards more intentional, capable, and safe AI assistants for human benefits and social good.

Real-world professional desktop workflows in specialized creative and engineering software unfold over long horizons and often require human-in-the-loop coordination, where agents proactively seek necessary information and users provide additional instructions, clarifications, feedback, or corrections as the task progresses. Yet existing desktop GUI benchmarks mostly reduce this setting to short, simplified tasks with all user instructions provided upfront. To address this issue, we introduce DeskCraft, a desktop GUI benchmark targeting long horizon creative and engineering workflows and proactive human-agent collaboration. DeskCraft organizes tasks into a multilevel difficulty taxonomy, with long horizon tasks requiring over 50 execution steps, and covers professional creative software across design, video, audio, and 3D creation. Furthermore, DeskCraft formalizes human-agent collaboration into an interaction protocol covering mid-turn and post-turn exchanges. Mid-turn interaction captures both agent-initiated clarification under uncertainty and user-initiated interruption during execution, while post-turn interaction accommodates user-driven feedback after the agent signals completion, together spanning the full space of realistic collaboration patterns. We evaluate 18 proprietary and open source agents on 538 tasks and find that GPT-5.4 reaches 31.6% on standard tasks and 27.6% on interactive tasks. Further analyses reveal persistent failures in long horizon workflow delivery and proactive clarification. We will open-source all evaluation codes, tasks, and data at https://github.com/mrwwk/DeskCraft.

We introduce OfficeQA Pro, a benchmark for evaluating AI agents on grounded, multi-document reasoning over a large and heterogeneous document corpus. The corpus consists of U.S. Treasury Bulletins spanning nearly 100 years, comprising 89,000 pages and over 26 million numerical values. OfficeQA Pro consists of 133 questions that require precise document parsing, retrieval, and analytical reasoning across both unstructured text and tabular data. Frontier LLMs including Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro Preview achieve less than 5% accuracy on OfficeQA Pro when relying on parametric knowledge, and less than 12% with additional access to the web. When provided directly with the document corpus, frontier agents still struggle on over half of questions, scoring 34.1% on average. We find that providing agents with a structured document representation produced by Databricks' ai_parse_document yields a 16.1% average relative performance gain across agents. We conduct additional ablations to study the effects of model selection, table representation, retrieval strategy, and test-time scaling on performance. Despite these improvements, significant headroom remains before agents can be considered reliable at enterprise-grade grounded reasoning.

LLM agents acting in structured environments fail in operational rather than conversational ways, and reliability depends on procedural knowledge of the environment. Prior self-improvement methods accumulate natural-language guidance without checking that each new item preserves previously correct behavior, so a note that fixes one trajectory can silently regress another. We introduce GRASP (Gated Regression-Aware Skill Proposer), which treats agent improvement as a sequence of edits to a bounded skill library, admitting each candidate only if it produces a net improvement on a balanced held-out probe under a hard regression budget. We evaluate GRASP across five base models (gpt-oss-120b, DeepSeek V4 Flash, Gemini 3.1 Flash Lite, GPT-4.1, GPT-5.4) on two FHIR-based clinical benchmarks. On MedAgentBench, GRASP lifts gpt-oss-120b from 40.6% to 88.8%, exceeds the strongest of five self-improvement baselines by 21.0 points, and improves every other base model by 17.2 to 40.3 points. Ablations attribute the gain to comparative proposal generation, the acceptance gate, and the hard regression budget rather than to skill writing itself, which without validation is no better than using no skills. The mechanism generalizes beyond the clinical domain, improving agents on three of four non-clinical environments and remaining flat only where the action space is open-ended. Frozen libraries transfer across models, where skills from a stronger model improve weaker executors beyond what they learn for themselves while the reverse does not, an asymmetry that no ungated baseline reproduces.

Humans learn social norms and behaviors from verbal feedback (e.g., a parent saying "that was rude" or a friend explaining "here's why that hurt"). Yet, learning from feedback for LLMs has largely focused on domains like code and math, where RL rewards are directly verifiable and condensed into scalar values. As LLMs are increasingly used to simulate human behavior, e.g., standing in for users, patients, students, and other personas, there is a pressing need to make them more human-like, which requires embracing a fundamentally different kind of signal: feedback that is verbal, subjective, and multi-faceted. We present DITTO, a model trained by treating verbal feedback as a first-class signal in reinforcement learning. After each rollout, DITTO receives verbal feedback and generates a feedback-conditioned improved rollout; both outputs are jointly optimized with GRPO, distilling verbal guidance into the base policy without requiring feedback at test time. We also introduce SOUL (Simulation gym Of hUman-Like behavior), a unified benchmark and training data suite spanning 10 tasks across six categories: Theory of Mind, character role play, social skill, learner simulation, user simulation, and persona simulation. DITTO achieves an average 36% improvement over the base model and exceeds GPT-5.4 on 6 of 10 SOUL benchmarks, demonstrating that RL with verbal feedback is a promising direction for training LLMs to simulate human behavior.

Retrieval for search agents is still inherited from non-agentic information retrieval: a retriever ranks the corpus and the agent reads a small set of returned documents. Recent direct corpus interaction (DCI) work shows that agents can instead interact with the raw corpus through shell tools such as grep and file reads. But unbounded interaction does not scale: every broad shell command is a scan over the whole corpus, and latency degrades sharply as the corpus grows. We argue that the role of retrieval for agentic search is not just to select documents that fit in the LLM context window, but to construct an interaction space: a bounded subset of the corpus the agent can explore with associated tools. Two design consequences follow. The space needs a boundary supplied by retrieval, and the objects within it should be processed for interaction. As a proof of concept, we propose RISE (Retrieving Interaction SpacE): we use BM25 to construct the interaction space; meanwhile, its documents are processed during indexing for shell-style navigation. On BrowseComp-Plus, RISE matches the pure-shell DCI baseline at 78% accuracy with gpt-5.4-mini at roughly one quarter of the per-query cost. At 1M documents, RISE-BM25 reaches 81% on gpt-5.4-mini, whereas DCI on gpt-5.4-nano degrades to 60% with 33 of 100 wall-clock failures.

Computer-use agents (CUAs) automate on-screen work, as illustrated by GPT-5.4 and Claude. Yet their reliability on complex, low-frequency interactions is still poor, limiting user trust. Our analysis of failure cases from advanced models suggests a long-tail pattern in GUI operations, where a relatively small fraction of complex and diverse interactions accounts for a disproportionate share of task failures. We hypothesize that this issue largely stems from the scarcity of data for complex interactions. To address this problem, we propose a new benchmark CUActSpot for evaluating models' capabilities on complex interactions across five modalities: GUI, text, table, canvas, and natural image, as well as a variety of actions (click, drag, draw, etc.), covering a broader range of interaction types than prior click-centric benchmarks that focus mainly on GUI widgets. We also design a renderer-based data-synthesis pipeline: scenes are automatically generated for each modality, screenshots and element coordinates are recorded, and an LLM produces matching instructions and action traces. After training on this corpus, our Phi-Ground-Any-4B outperforms open-source models with fewer than 32B parameters. We will release our benchmark, data, code, and models at https://github.com/microsoft/Phi-Ground.git

One-shot Program-of-Thought (PoT) emits a Python program that prints a primitive-action plan; a single invalid action silently invalidates the trajectory. We introduce RePoT (Recoverable PoT): a deterministic verified replay that walks the plan through the environment to its first invalid transition, then one LLM call that resumes from the verified prefix. RePoT costs at most one extra LLM call on the ~14% of problems where PoT fails. RePoT beats PoT by +3 to +11pp across four closed-model configurations on PuzzleZoo-775 and peaks at 96.9% vs 86.3% on gpt-5.4-mini-medium; against the matched-budget PoT-retry baseline, RePoT wins decisively on Gemini (+3.8pp, 95% CI [+2.2,+5.4]), is within sampling noise on GPT-medium and Claude, and loses on GPT-mini -- a capability-scaling pattern we begin to address with Adaptive RePoT, a rule-based dispatcher that routes between suffix repair and a fresh PoT retry based on verified-prefix length (preliminary). We replicate on PlanBench Blocksworld (+1.1 to +11.4pp) and on four open-weights models (+3.3 to +20.0pp on three of four). On Derail-550, our controlled recovery benchmark, every condition with access to checkpoint information clears >=30% on GPT-medium and >=70% on Gemini, vs <=3.1% for error-only feedback -- showing that checkpoint information, not the specific verified-prefix tail, is the load-bearing recovery signal.

Can AI make progress on important, unsolved mathematical problems? Large language models are now capable of sophisticated mathematical and scientific reasoning, but whether they can perform novel research is still widely debated and underexplored. We introduce HorizonMath, a benchmark of over 100 predominantly unsolved problems spanning 8 domains in computational and applied mathematics, paired with an open-source evaluation framework for automated verification. Our benchmark targets a class of problems where discovery is hard, requiring meaningful mathematical insight, but verification is computationally efficient and simple. Because these solutions are unknown, HorizonMath is immune to data contamination, and most state-of-the-art models score near 0%. Existing research-level benchmarks instead rely on formal proof verification or manual review, both of which are expensive to scale. Using this platform, we find two problems for which GPT 5.4 Pro proposes solutions that improve on the best-known published results, representing potential novel contributions (pending expert review). We release HorizonMath as an open challenge and a growing community resource, where correct solutions to problems in the unsolved problem classes could constitute novel results in the mathematical literature.

Frontier coding agents are increasingly used in workflows where users supervise progress primarily through repeated improvement of a public score, namely the reported score on a public evaluation file with labels in the workspace, rather than through direct inspection of the agent's intermediate outputs. We study whether multi-round user pressure to improve that score induces public score exploitation: behavior that raises the public score through shortcuts without improving hidden private evaluation. We begin with a preliminary single-script tabular classification task, where GPT-5.4 and Claude Opus 4.6 both exploit label information within 10 rounds of user-agent interaction. We then build AgentPressureBench, a 34-task machine-learning repository benchmark spanning three input modalities, and collect 1326 multi-round trajectories from 13 coding agents. On our benchmark, we observe 403 exploitative runs, spanning across all tasks. We also find that stronger models have higher exploitation rates, supported by a significant Spearman rank correlation of 0.77. Our ablation experiments show that higher user pressure leads to earlier exploitation, reducing the average first exploit round by 15.6 rounds (i.e., 19.67 to 4.08). As a mitigation, adding explicit anti-exploit wordings in prompt mostly eliminates exploitation (100% to 8.3%). We hope that our work can bring attention to more careful use of coding agents workflow, and developing more robust coding agents under user pressure. Our project page is at https://ucsc-vlaa.github.io/AgentPressureBench .

Trust in an AI system is often anchored by explanations of how it works, which one then uses to forecast its behavior on new inputs. For large reasoning models (LRMs), this conventional route is particularly difficult to follow: explanation methods for single token generations do not naturally generalize to long trajectories, and the trajectories themselves are often not faithful when read as natural language. We propose an alternative that bypasses the explanation step: treat behavior forecasting as a learnable task and train Behavior Forecasters that operates on a single reasoning trajectory to make the same forecasts one would typically seek from an explanation. The forecaster's training data is obtained by querying the LRM with no human annotation, and its inference is done in a single forward pass. We instantiate this approach on two tasks: how likely the LRM is to repeat its answer on re-runs, and how removing parts of the input changes its answer. We evaluate this approach on both tasks across three diverse reasoning datasets and find that trained Behavior Forecasters are more accurate than GPT-5.4 and Claude Opus-4.6 reading the same trajectories as naive readers, at a small fraction of their inference cost. We find that fine-tuning the backbone end-to-end and initializing it from the target LRM are each necessary for strong performance. These results show that the reasoning trajectory carries information about the LRM's future behavior that goes beyond what naive reading conveys.

Can VLMs predict how each camera move changes the view, and plan many such moves ahead? We call this capability view planning, requiring (1)understanding how a single action transforms the view, and (2)composing many such transformations across multi-turn plans to identify a target view. We probe both abilities in our proposed ViewSuite, a 3D point-cloud environment on real ScanNet scenes. Across 13 frontier VLMs, a critical planning gap emerges: they possess basic view-action knowledge but fail to compose it across multi-turn plans, with the gap widening as viewpoint distance grows. To close this gap, we propose an iterative framework that alternates self-exploration with view graph distillation. The key insight is that all exploration trajectories, regardless of their outcome, collectively form a view graph that compactly captures how viewpoints connect across a scene. Distilling this graph into diverse supervised tasks reshapes the policy distribution and overcomes the sparse rewards that stall pure RL. This improves Qwen2.5-VL-7B from 2.5% to 47.8% on interactive view planning, surpassing GPT-5.4 Pro (18.5%) and Gemini 3.1 Pro (21.4%). Self-exploration emerges as a promising path toward VLMs that can actively reason and plan in 3D space. Code and Data are at https://viewsuite.github.io.

The dominance of colonial languages in African education and scientific communication limits how hundreds of millions of speakers of African languages access and produce scientific knowledge. A core obstacle is the lack of established scientific terminology in these languages. We introduce AfriScience-MT, a parallel corpus covering six African languages (Amharic, Hausa, Luganda, Northern Sotho, Yorùbá, and isiZulu) across 11 scientific domains. Professional translators, working with expert science communicators, translated plain-language summaries of scientific papers into each target language and created new terms where none existed. We benchmark machine translation systems and large language models in zero-shot, few-shot, and fine-tuned settings. Our results show that closed-source models outperform all open-source models at both the sentence and document levels: GPT-5.4 and Gemini-3.1-Flash-Lite lead with average sentence-level COMET scores of 68.3 and 68.0, respectively, and tie at an average document-level COMET of 48.3. Among open systems, fine-tuned NLLB-1.3B reaches 67.3 at the sentence level, and TranslateGemma-12B reaches 44.0 at the document level with 1-shot in-context learning. We release AfriScience-MT to support benchmarking and document-level scientific MT for African languages.

Inference-time scaling for LLM reasoning has focused on three axes: chain depth, sample breadth, and learned step-scorers (PRMs). We introduce a fourth axis, granularity of external verbal supervision, via Verbal Process Supervision (VPS), a training-free framework that uses structured natural-language critique from a stronger supervisor to guide an iterative generate-critique-refine loop up to a round budget R. Across GPQA Diamond, AIME 2025, and LiveCodeBench V6 (covering both closed and open models), VPS yields three key results. First, on GPQA Diamond, GPT-5.4 (High) | GPT-5.4 (Low) reaches 94.9% at R=4, surpassing the 94.1% state of the art without gradient updates. Second, on AIME 2025, VPS enables strong weak-actor rescue, boosting scores from 11.7-26.7% to 63.3-90.0% (up to +63.3 points). Third, at matched compute, VPS outperforms Reflexion by +8.5 to +12.1 points and Self-Consistency@5 by +5.0 pp (GPQA) and +8.3 pp (LiveCodeBench), isolating critique granularity as the key driver. Performance scales with the supervisor-actor capability gap (Pearson r=0.90) and degrades when errors are not linguistically expressible (e.g., code synthesis), motivating hybrid verbal-executable methods. These results establish critique granularity as a new axis of inference-time scaling.

We introduce a cross-domain behavioural assay of monitoring-control coupling in LLMs, grounded in the Nelson and Narens (1990) metacognitive framework and applying human psychometric methodology to LLM evaluation. The battery comprises 524 items across six cognitive domains (learning, metacognitive calibration, social cognition, attention, executive function, prospective regulation), each grounded in an established experimental paradigm. Tasks T1-T5 were pre-registered on OSF prior to data collection; T6 was added as an exploratory extension. After every forced-choice response, dual probes adapted from Koriat and Goldsmith (1996) ask the model to KEEP or WITHDRAW its answer and to BET or decline. The critical metric is the withdraw delta: the difference in withdrawal rate between incorrect and correct items. Applied to 20 frontier LLMs (10,480 evaluations), the battery discriminates three profiles consistent with the Nelson-Narens architecture: blanket confidence, blanket withdrawal, and selective sensitivity. Accuracy rank and metacognitive sensitivity rank are largely inverted. Retrospective monitoring and prospective regulation appear dissociable (r = .17, 95% CI wide given n=20; exemplar-based evidence is the primary support). Scaling on metacognitive calibration is architecture-dependent: monotonically decreasing (Qwen), monotonically increasing (GPT-5.4), or flat (Gemma). Behavioural findings converge structurally with an independent Type-2 SDT approach, providing preliminary cross-method construct validity. All items, data, and code: https://github.com/synthiumjp/metacognitive-monitoring-battery.

Existing AI benchmarks lack the fidelity to assess economically meaningful progress on professional workflows. To evaluate frontier AI agents in a high-value, labor-intensive profession, we introduce BankerToolBench (BTB): an open-source benchmark of end-to-end analytical workflows routinely performed by junior investment bankers. To develop an ecologically valid benchmark grounded in representative work environments, we collaborated with 502 investment bankers from leading firms. BTB requires agents to execute senior banker requests by navigating data rooms, using industry tools (market data platform, SEC filings database), and generating multi-file deliverables--including Excel financial models, PowerPoint pitch decks, and PDF/Word reports. Completing a BTB task takes bankers up to 21 hours, underscoring the economic stakes of successfully delegating this work to AI. BTB enables automated evaluation of any LLM or agent, scoring deliverables against 100+ rubric criteria defined by veteran investment bankers to capture stakeholder utility. Testing 9 frontier models, we find that even the best-performing model (GPT-5.4) fails nearly half of the rubric criteria and bankers rate 0% of its outputs as client-ready. Our failure analysis reveals key obstacles (such as breakdowns in cross-artifact consistency) and improvement directions for agentic AI in high-stakes professional workflows.

Learning from past experience benefits from two complementary forms of memory: episodic traces -- raw trajectories of what happened -- and consolidated abstractions distilled across many episodes into reusable, schema-like lessons. Recent agentic-memory systems pursue the consolidated form: an LLM rewrites past trajectories into a textual memory bank that it continuously updates with new interactions, promising self-improving agents without parameter updates. Yet we find that such consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As consolidation proceeds, memory utility first rises, then degrades, and can fall below the no-memory baseline. More surprisingly, even when consolidating from ground-truth solutions, GPT-5.4 fails on 54% of a set of ARC-AGI problems it had previously solved without memory. We trace the regression to the consolidation step rather than the underlying experience: the same trajectories yield qualitatively different memories under different update schedules, and an episodic-only control that simply retains those trajectories remains competitive with the consolidators we test. In a controlled ARC-AGI Stream environment that exposes Retain, Delete, and Consolidate actions, agents preserve raw episodes by default and double the accuracy of their forced-consolidation counterparts; disabling consolidation entirely (episodic management only) matches this auto regime. Practically, robust agent memory should treat raw episodes as first-class evidence and gate consolidation explicitly rather than firing it after every interaction. Looking forward, reliable agentic memory will require LLMs that can consolidate without overwriting the evidence they depend on.

Large language models (LLMs) achieve strong relation extraction (RE), but their computational demands and reliance on proprietary APIs limit deployment in resource-constrained or privacy-sensitive settings. We investigate how far small language models (SLMs) can close this gap across general-domain and literary text. We evaluate five models from 360M to 3B parameters under three domain-composition regimes and two prompt-conditioned tuning styles (30 configurations), comparing them with zero-shot frontier LLMs and a discriminative RoBERTa baseline. Across nine benchmarks, the best sub-billion model, Qwen2.5-0.5B fine-tuned on pooled general-domain data, achieves a general-domain positive-class micro-F1 of 0.83, versus 0.69 for GPT-5.4 and 0.66 for Claude Sonnet 4.6 evaluated zero-shot. This does not imply that SLMs are intrinsically stronger; rather, targeted task adaptation enables 4-bit models deployable on a single consumer GPU to outperform general-purpose frontier systems under this protocol. An in-domain RoBERTa baseline also exceeds both frontier models, indicating that the gain stems from task adaptation rather than generative decoding. On literary RE, tuned SLMs reach 0.92 on the human-annotated Biographical benchmark versus 0.83 for GPT-5.4, and 0.833 versus 0.578 on the two-benchmark literary average. A targeted domain-adaptive pretraining case study yields no practically meaningful gain over supervised fine-tuning, while the cleanest within-family scale comparison shows only marginal improvement. These results show that, when task-specific data are available, compact task-adapted models can provide accurate, private, and hardware-efficient RE.

Currently, evaluating vision-language models (VLMs) in medical imaging tasks oversimplifies clinical reality by relying on pre-selected 2D images that demand significant manual labor to curate. This setup misses the core challenge of realworld diagnostics: a true clinical agent must actively navigate full 3D volumes across multiple sequences or modalities to gather evidence and ultimately support a final decision. To address this, we propose MEDOPENCLAW, an auditable runtime designed to let VLMs operate dynamically within standard medical tools or viewers (e.g., 3D Slicer). On top of this runtime, we introduce MEDFLOWBENCH, a full-study medical imaging benchmark covering multi-sequence brain MRI and lung CT/PET. It systematically evaluates medical agentic capabilities across viewer-only, tool-use, and open-method tracks. Initial results reveal a critical insight: while state-of-the-art LLMs/VLMs (e.g., Gemini 3.1 Pro and GPT-5.4) can successfully navigate the viewer to solve basic study-level tasks, their performance paradoxically degrades when given access to professional support tools due to a lack of precise spatial grounding. By bridging the gap between static-image perception and interactive clinical workflows, MEDOPENCLAW and MEDFLOWBENCH establish a reproducible foundation for developing auditable, full-study medical imaging agents.

Multi-agent reasoning systems adopt a "generate-then-transfer" paradigm that forces end-to-end latency to scale linearly with pipeline depth. We introduce StreamMA, a multi-agent reasoning system that streams each reasoning step to downstream agents as soon as it is generated, pipelining adjacent agents and thus reducing latency. Surprisingly, this pipelining also improves effectiveness: because multi-step reasoning quality is non-uniform and early steps are more reliable than later ones, working with these reliable early steps instead of the full chain prevents error-prone late steps from misleading downstream agents. We formalize both advantages with the first closed-form joint analysis of stream, serial, and single protocols, deriving the effectiveness ordering, speedup upper bound, and cost ratio. Across eight reasoning benchmarks spanning mathematics, science, and code, two frontier LLMs (Claude Opus 4.6 and GPT-5.4), and three topologies (Chain, Tree, Graph), StreamMA outperforms both baselines (avg. +7.3 pp, max +22.4 pp on HMMT 2026; Claude Opus 4.6-high). Beyond these contributions, we discover a "step-level scaling law": increasing per-agent steps consistently improves both effectiveness and efficiency, a new scaling dimension orthogonal to and composable with agent-count scaling.

OpenClaw, the most widely deployed personal AI agent in early 2026, operates with full local system access and integrates with sensitive services such as Gmail, Stripe, and the filesystem. While these broad privileges enable high levels of automation and powerful personalization, they also expose a substantial attack surface that existing sandboxed evaluations fail to capture. To address this gap, we present the first real-world safety evaluation of OpenClaw and introduce the CIK taxonomy, which unifies an agent's persistent state into three dimensions, i.e., Capability, Identity, and Knowledge, for safety analysis. Our evaluations cover 12 attack scenarios on a live OpenClaw instance across four backbone models (Claude Sonnet 4.5, Opus 4.6, Gemini 3.1 Pro, and GPT-5.4). The results show that poisoning any single CIK dimension increases the average attack success rate from 24.6% to 64-74%, with even the most robust model exhibiting more than a threefold increase over its baseline vulnerability. We further assess three CIK-aligned defense strategies alongside a file-protection mechanism; however, the strongest defense still yields a 63.8% success rate under Capability-targeted attacks, while file protection blocks 97% of malicious injections but also prevents legitimate updates. Taken together, these findings show that the vulnerabilities are inherent to the agent architecture, necessitating more systematic safeguards to secure personal AI agents. Our project page is https://ucsc-vlaa.github.io/CIK-Bench.

We present DEI: Diversity in Evolutionary Inference, a distributed Quality-Diversity (QD) search framework that assigns heterogeneous large language models (LLMs) as mutation operators across peer nodes communicating with non-blocking collective operations. Unlike homogeneous parallel search, which replicates a single model's inductive biases across all workers, DEI treats each LLM's distinct creative prior as a complementary source of behavioral novelty. Extending the Digital Red Queen framework with DEI, nodes share local optimal solutions at the end of each round to seed the next round's population. This creates cross-model adversarial pressure that drives robustness beyond intra-model self-play. Evaluated on the Core War domain, a competitive programming benchmark in which Redcode warrior programs battle inside a simulated machine, a four-node heterogeneous ensemble (GPT-5.4-mini, Claude Sonnet 4.6, GPT-5.2, and Claude Haiku 4.5) achieves 124 percent higher merged-archive QD-Score (45.90 vs. 20.46) and 28 percent higher coverage (80.6 percent vs. 63.0 percent of cells) than a single-node baseline at equal total LLM-call budget. The heterogeneous ensemble also outperforms an equally-budgeted homogeneous ensemble on QD-Score, coverage, and held-out solution generality across all four model families. These results provide the first empirical evidence that model diversity, not merely parallelism, is the key driver of gain in distributed LLM-based QD search.

We introduce CurveBench, a benchmark for hierarchical topological reasoning from visual input. CurveBench consists of 756 images of pairwise non-intersecting Jordan curves across easy, polygonal, topographic-inspired, maze-like, and dense counting configurations. Each image is annotated with a rooted tree encoding the containment relations between planar regions. We formulate the task as structured prediction: given an image, a model must recover the full rooted containment tree induced by the curves. Despite the visual simplicity of the task, the strongest evaluated model, Gemini 3.1 Pro, achieves only 71.1\% tree-generation accuracy on CurveBench-Easy and 19.1\% on CurveBench-Hard. We further demonstrate benchmark utility through RLVR-style fine-tuning of open-weight vision-language models. Our trained Qwen3-VL-8B model improves over Qwen-3-VL-8B-Thinking from 2.8\% to 33.3\% tree-generation accuracy on CurveBench-Easy, exceeding GPT-5.4 and Claude Opus 4.5 under our evaluation protocol. The remaining gap, especially on CurveBench-Hard, shows that exact topology-aware visual reasoning remains far from solved.

Training language models (LMs) remains a highly human-intensive process, even as frontier language model agents become increasingly capable at software engineering and other long-horizon tasks. A central challenge is that autonomous post-training is not just a coding problem: it requires the agent to repeatedly plan iterations, construct benchmark-aligned data, run stable training jobs, evaluate checkpoints, and preserve experiment state across many hours of interaction. We present AutoTrainess, a LM agent that exposes these operations as a repository of agent-computer interfaces for planning, data preparation, training, evaluation, and logging. Rather than leaving the agent to operate in a raw CLI environment with an underspecified action space, AutoTrainess externalizes prior human experience as explicit workflows, rules, and execution constraints that guide the agent toward effective and reliable training behavior. On PostTrainBench, AutoTrainess consistently outperforms CLI-only baselines, achieving 26.94 average score with GPT-5.4 (Codex) versus 23.21 for CLI-only. It also generalizes across models and harnesses, improving DeepSeek-V4-Flash (OpenCode) from 12.13 to 19.58.

Motivated by the underspecified, multi-hop nature of search queries and the multimodal, heterogeneous, and often conflicting nature of real-world web results, we introduce MERRIN (Multimodal Evidence Retrieval and Reasoning in Noisy Web Environments), a human-annotated benchmark for evaluating search-augmented agents. MERRIN measures AI agents' ability to identify relevant modalities, retrieve multimodal evidence, and perform multi-hop reasoning over noisy web sources. It differs from prior work in three important aspects: (1) using natural language queries without explicit modality cues, (2) incorporating underexplored modalities such as video and audio, and (3) requiring the retrieval of complex, often noisy or conflicting multimodal evidence during web search. We evaluate diverse search agents powered by ten models, including strong closed-source models (e.g., GPT-5.4-mini, Gemini 3/3.1 Flash/Pro) and open-weight models (Qwen3-4B/30B/235B), across three search settings (no search, native search, and agentic search). Our results show that MERRIN is highly challenging: the average accuracy across all agents is 22.3%, with the best-performing agent reaching only 40.1%. We further observe that while stronger agents like Gemini Deep Research achieve higher performance, gains are modest due to over-exploration; they take more steps and use more tools, but are often distracted by conflicting or partially relevant web content, leading to incorrect answers. Compared to humans, these agents consume more resources yet achieve lower accuracy, largely due to inefficient source selection and an overreliance on text modalities. These findings highlight the need for search agents capable of robust search and reasoning across diverse modalities in noisy web environments, making MERRIN a valuable testbed for evaluating such capabilities.

Large Language Models (LLMs) are poised to disrupt knowledge work, with the emergence of delegated work as a new interaction paradigm (e.g., vibe coding). Delegation requires trust - the expectation that the LLM will faithfully execute the task without introducing errors into documents. We introduce DELEGATE-52 to study the readiness of AI systems in delegated workflows. DELEGATE-52 simulates long delegated workflows that require in-depth document editing across 52 professional domains, such as coding, crystallography, and music notation. Our large-scale experiment with 19 LLMs reveals that current models degrade documents during delegation: even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows, with other models failing more severely. Additional experiments reveal that agentic tool use does not improve performance on DELEGATE-52, and that degradation severity is exacerbated by document size, length of interaction, or presence of distractor files. Our analysis shows that current LLMs are unreliable delegates: they introduce sparse but severe errors that silently corrupt documents, compounding over long interaction.

Recent auto-research systems can produce complete papers, but feasibility is not the same as quality, and the field still lacks a systematic study of how good agent-generated papers actually are. We introduce ResearchArena, a minimal scaffold that lets off-the-shelf agents (Claude Code using Opus 4.6, Codex using GPT-5.4, and Kimi Code using K2.5) carry out the full research loop themselves (ideation, experimentation, paper writing, self-refinement) under only lightweight guidance. Across 13 computer science seeds and 3 trials per agent-domain pair, ResearchArena yields 117 agent-generated papers, each evaluated under three complementary lenses: a manuscript-only reviewer (SAR), an artifact-aware peer review (PR) in which agents inspect the workspace alongside the manuscript, and an human conducted meta-review. Under SAR alone the picture is optimistic: Claude Code obtains the highest score, outperforms Analemma's FARS, and matches the weighted-average human ICLR 2025 submission, suggesting that minimally scaffolded agents can produce papers that look competitive on manuscript-only review. Manual inspection, however, reveals this picture is overstated: SAR scores are poorly aligned with its actual acceptance decisions and reward plausible framing without verifying experimental substance. Under artifact-aware PR scores drop sharply, and manual auditing identifies experimental rigor as the major bottleneck, decomposing into three failure modes (fabricated results, underpowered experiments, and plan/execution mismatch) that are highly agent-dependent: Codex 5%/8% paper-vs-artifact mismatch / fabricated references versus Kimi Code 77%/72%, a sim15times spread that tracks distinct research personas the agents develop. None of the 117 agent-generated papers reaches the acceptance bar of a top-tier venue. This suggests that we are still gapped from the true auto-research.

We release Terminal Wrench, a subset of 331 terminal-agent benchmark environments, copied from the popular open benchmarks that are demonstrably reward-hackable. The data set includes 3,632 hack trajectories and 2,352 legitimate baseline trajectories across three frontier models (Claude Opus 4.6, Gemini 3.1 Pro, GPT-5.4). Each entry preserves the original task definition alongside full attack trajectories that show how the verifier was bypassed. It also includes cases where the task was not solved as intended. The tasks span system administration, machine learning, software engineering, and security challenges; the exploits range from simple output spoofing to stack-frame introspection, standard-library patching, and rootkit-style binary hijacking. Crucially, these exploits are specific to each task, rather than the evaluation harness, making them harder to patch. We also present a monitorability study in which hack trajectories are sanitized or stripped of reasoning traces and then scored by an LLM judge, showing that detection degrades meaningfully when chain-of-thought is removed (AUC drops from 0.97 to 0.92). The data set is publicly available at https://github.com/few-sh/terminal-wrench.

Foundation models succeed when they learn in the native structure of a modality, whether morphology-respecting tokens in language or pixels in vision. Wireless packet traces deserve the same treatment: meaning emerges from layered headers, typed fields, timing gaps, and cross-packet state machines, not flat strings. We present Plume (Protocol Language Understanding Model for Exchanges), a compact 140M-parameter foundation model for 802.11 traces that learns from structured PDML dissections. A protocol-aware tokenizer splits along the dissector field tree, emits gap tokens for timing, and normalizes identifiers, yielding 6.2x shorter sequences than BPE with higher per token information density. Trained on a curated corpus, Plume achieves 74-97% next-packet token accuracy across five real-world failure categories and AUROC >= 0.99 for zero-shot anomaly detection. On the same prediction task, frontier LLMs (Claude Opus 4.6, GPT-5.4) score comparably despite receiving identical protocol context, yet Plume does so with > 600x fewer parameters, fitting on a single GPU at effectively zero marginal cost vs. cloud API pricing, enabling on-prem, privacy-preserving root cause analysis.

Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised. Acting correctly requires using the current value of a fact and discarding values that have been superseded. We isolate this ability on real conversational data and show that it is a distinct, unsolved failure. On the knowledge-update subset of LongMemEval, replacing an agent's full context with a bounded, self-maintained memory drops accuracy from 92% to 77% even on a frontier model (gpt-5.4), a gap that is statistically significant (paired McNemar p<0.005) and persists across model scale while full-context accuracy saturates near 92%. The bottleneck is therefore memory maintenance, not comprehension, and is not closed by a stronger model. We then ask whether this is merely an undersized memory, and find it is not: as the conversation grows 24x, accuracy falls further (from 68% to 28%), and granting the agent proportionally more memory yields no detectable recovery (28% to 28%, n=25). The failure scales with the length of the conversation, not the compression ratio. We release Supersede, an open reinforcement-learning environment (on the verifiers / prime-rl stack) that turns this measurement into a training signal: agents are rewarded for answering from the current value and penalized for stale ones. Finally, we close the loop and show the gap is trainable: GRPO fine-tuning a small open model (Qwen2.5-3B) on this environment nearly doubles its held-out supersession accuracy on real, unseen conversations (9.0% to 16.7%, a single run), along a monotonic checkpoint curve indicating the learned policy, not the harness, carries the gain. To our knowledge this is the first trainable environment whose reward targets temporal fact-currency, and the first evidence the supersession gap can be trained down, not only measured.

Collecting computer use data from human demonstrations is expensive and slow, motivating the need for scalable generation strategies. This requires two key ingredients: environments in which agents can act and verifiers that can judge whether their demonstrations succeeded. We introduce FaraGen1.5, a scalable data pipeline for computer use agents composed of three modular components: environments, solvers, and verifiers. FaraGen1.5 uses both live websites and synthetic environments that faithfully simulate domains gated by authentication or that require irreversible actions. It employs a solver harness that can be powered by multiple models, including strong frontier models such as GPT-5.4, and also incorporates a user simulator to enable multi-turn rollouts. Finally, FaraGen1.5 scores the resulting trajectories with three complementary verifiers covering task correctness, efficiency, and critical-point adherence. Using data produced by this pipeline, we train Fara1.5, a family of native computer use agents (CUAs) at three scales built on Qwen3.5 (4B, 9B, and 27B). To train these models, we employ a supervised finetuning (SFT) recipe that carefully balances data from FaraGen1.5 for broad coverage, specific high-value tasks, and target model deficiencies in an iterative approach. Each model sets a new state of the art for its size class on browser-use benchmarks: Fara1.5-9B reaches 63.4% on Online-Mind2Web and 86.6% on WebVoyager, while Fara1.5-27B achieves 72.3% on Online-Mind2Web, which is competitive with much larger proprietary systems.

AI is increasingly used to support scientific peer review, from manuscript screening, reviewer assistance to editorial triage. Although such systems promise to reduce reviewer burden and accelerate publication, their robustness to strategic manipulation remains poorly understood. Here we show that AI-mediated peer review is vulnerable to a simple, low-cost manipulation: superficial rephrasing of the manuscript abstract. Without changing the underlying scientific content and communication, and even without knowledge of the reviewing model, adversarially rewritten abstracts substantially improve AI review outcomes. We see this across disciplines and publication venues, for both human-written and AI-generated papers. Our strongest attack achieves an attack-success-rate of about 38%, increasing acceptance ratings by +1.31 for Gemini 3 Flash reviewers and by +0.88 for GPT 5.4 Mini reviewers on a 10-point scale. When the original AI review suggests 'reject', the success rate rises to more than 50%. This effect extends beyond overall score inflation, increasing review confidence and scores on core scientific criteria such as soundness, significance and perceived contribution. The attack is practical, requiring only about 5 minutes and $1 for a 10-page AI conference submission, and is hard to distinguish from ordinary scientific editing. Inflated AI reviews could bias downstream human decision-making, shifting editorial recommendations from rejection towards acceptance. These findings reveal a general vulnerability in AI-assisted scientific evaluation: when AI-generated review influence editorial decisions, authors may be incentivized to optimize manuscripts for AI judgment rather than scientific merit. Our results suggest that AI tools should not be treated as neutral evaluators in high-stakes peer review without systematic robustness testing, transparent safeguards and careful human oversight.

As language models are increasingly deployed as autonomous agents, they must coordinate with others over long horizons in open-ended interactive tasks. Yet existing evaluations rarely test these demands together, instead emphasising single-agent tasks, short interactions, or highly structured multi-agent settings. We introduce alem, a JAX-based benchmark for open-ended multi-agent coordination built on Craftax-like dynamics. Alem embeds procedurally generated coordination tasks, soft specialisation, communication, and controllable coordination difficulty into a long-horizon survival world with exploration, crafting, trading, and combat. We evaluate 13 modern LLMs zero-shot within homogeneous teams, with trained MARL agents as reference points. Current LLM agents remain far from solving alem, averaging only ~6% normalised return, but their failures are not uniform. On the hardest coordination setting, zero-shot Gemini-3.1-Pro-High approaches MARL agents trained for one billion steps, while GPT-5.4-High achieves strong base-task reward but much lower coordination reward. This contrast shows that individual task competence does not imply coordination competence. Ablations show that communication is the largest contributor to coordination, while memory and reasoning help when used to maintain multi-step plans. Overall, our results identify coordination as a distinct bottleneck for frontier LLM agents, separate from single-agent capabilities. Alem makes this bottleneck measurable and provides a controlled testbed for developing agents that communicate, allocate roles, and execute shared plans. Code is available at https://github.com/alem-world/alem-env.

We investigate whether large language models produce different medical triage recommendations for identical neurological symptoms when only the patient's stated gender and age vary. Using three model families--Gemini 3.5 Flash, Claude Sonnet 4.6, and GPT-5.4-mini--we present a standardized symptom profile (persistent headache, blurred vision, morning nausea, visual disturbances) across seven demographic conditions: three age groups (25, 38, 65) x two genders (male, female), plus a gender-unspecified baseline (n = 30 per condition per model, 630 total trials). We find a stark, systemic gender-dependent triage disparity: young women receive significantly lower emergency room (ER) referral rates than age-matched men (Gemini: 0% vs. 23.3%; Claude: 6.7% vs. 96.7%; GPT: 6.7% vs. 66.7%, all p < 0.001). The disparity disappears at age 65 for all models. The primary mechanism is diagnostic substitution: the models anchor on a gender-associated diagnosis, preferentially classifying young women with Idiopathic Intracranial Hypertension (IIH)--a condition epidemiologically linked to women of childbearing age--while diagnosing men with generic increased intracranial pressure with space-occupying lesions in the differential. This diagnostic closure routes female patients to lower-urgency care (outpatient doctor appointments) despite comparable severity ratings (7-9/10). Our findings demonstrate that clinical LLMs replicate documented human clinical biases by using epidemiological priors to suppress triage urgency, suggesting that AI triage engines must decouple urgency assessment from probabilistic diagnostic priors. We release all code, prompts, and raw results.

We present E3, an automated review assistant that augments reviewers and engineering teams by identifying decision-relevant technical concerns in research papers. For each concern, E3 reports its nature, its location, its bearing on the contribution, and the analysis or evidence that would resolve it, covering unsupported claims, missing ablations, weak baselines, hidden assumptions, threats to validity, and leakage risks. To evaluate E3 without contamination confounds we adopt an issue-level backtesting protocol: the corpus is restricted to papers postdating the training cutoff of every automated source, and for each paper a meta-judge that observes only anonymised reviews labels every issue-source pair as Caught, Partial, or Missed. Applied to 100 ICLR 2026 papers and 4598 judged issue rows, comparing E3 against the ICLR human reviews and two prompt-matched LLM baselines built on gpt-5.4 from OpenAI and claude-opus-4-6 from Anthropic, with meta-judge gpt-5.5, E3 attains the highest recall on every aggregate metric. Partial-inclusive recall reaches 90.2 percent, which is 15.5 points over GPT, 17.1 points over Claude, and 29.2 points over the human reviews, and strict recall preserves the ordering at 65.8 percent. On concerns raised by the human reviewers, E3 recovers 89.6 percent; on concerns the human reviewers missed it surfaces 1635 additional rows admitted into the judged union, 406 above the next-best source. Corpus, baseline prompts, judge prompt template, and evaluation code are released.

Prior research has established that instruction-tuned large language models exhibit left-of-center political bias, measured exclusively through abstract political questionnaires. We show that this finding does not generalize to concrete policy decisions. We introduce a dual-instrument methodology grounded in Swiss democratic reality. The Smartvote questionnaire (75 abstract policy questions) is administered to 66 LLMs from 27 model families and compared to 184 elected members of the Swiss National Council, replicating the established leftward convergence (Cohen's d = 3.64, p = 0.0002). Then, novel to this work, 9 flagship LLMs are confronted with 48 real federal referenda (Volksabstimmungen) in four national languages (German, French, Italian, Romansh) under three information conditions, comparing votes to actual outcomes and party recommendations (Parolen). Three findings challenge the prevailing narrative. (1) Abstract questionnaires do not predict concrete behavior: the left-to-right agreement gradient on Smartvote shifts from left-peaked to center-peaked on Volksabstimmungen, where models align most with centrist Die Mitte and FDP rather than leftist SP and Gruene (Wilcoxon p = 0.008). (2) For some models, the language of a political question changes the answer more than the political content does: cross-linguistic consistency ranges from 50% (Mistral) to 98% (GPT-5.4). (3) Two models exhibit systematic change-aversion rather than political bias, voting Nein on 83-94% of referenda regardless of direction (binomial p < 0.0001). What prior work measured as "leftward bias" may not generalize beyond abstract instruments. On concrete policy decisions, LLMs behave less like coalition partners of the left and more like cautious civil servants: centrist, status-quo-favoring, and inconsistent across languages.

Forecasting has become a natural benchmark for reasoning under uncertainty. Yet existing evaluations of large language models remain limited to judgmental tasks in simple formats, such as binary or multiple-choice questions. In practice, however, forecasting spans a far broader scope. Across domains such as economics, public health, and social demographics, decisions hinge on numerical estimates over continuous quantities, a capability that current benchmarks do not capture. Evaluating such estimates requires a format that makes uncertainty explicit and testable. We propose prediction intervals as a natural and rigorous interface for this purpose. They demand scale awareness, internal consistency across confidence levels, and calibration over a continuum of outcomes, making them a more suitable evaluation format than point estimates for numerical forecasting. To assess this capability, we introduce a new benchmark QuantSightBench, and evaluate frontier models under multiple settings, assessing both empirical coverage and interval sharpness. Our results show that none of the 11 evaluated frontier and open-weight models achieves the 90\% coverage target, with the top performers Gemini 3.1 Pro (79.1\%), Grok 4 (76.4\%), and GPT-5.4 (75.3\%) all falling at least 10 percentage points short. Calibration degrades sharply at extreme magnitudes, revealing systematic overconfidence across all evaluated models.

Healthcare administration accounts for over $1 trillion in annual spending, making it a promising target for LLM-based computer-use agents (CUAs). While clinical applications of LLMs have received significant attention, no benchmark exists for evaluating CUAs on end-to-end administrative workflows. To address this gap, we introduce HealthAdminBench, a benchmark comprising four realistic GUI environments: an EHR, two payer portals, and a fax system, and 135 expert-defined tasks spanning three administrative task types: Prior Authorization, Appeals and Denials Management, and Durable Medical Equipment (DME) Order Processing. Each task is decomposed into fine-grained, verifiable subtasks, yielding 1,698 evaluation points. We evaluate seven agent configurations under multiple prompting and observation settings and find that, despite strong subtask performance, end-to-end reliability remains low: the best-performing agent (Claude Opus 4.6 CUA) achieves only 36.3 percent task success, while GPT-5.4 CUA attains the highest subtask success rate (82.8 percent). These results reveal a substantial gap between current agent capabilities and the demands of real-world administrative workflows. HealthAdminBench provides a rigorous foundation for evaluating progress toward safe and reliable automation of healthcare administrative workflows.

We introduce Embodied-R1.5, a unified Embodied Foundation Model (EFM) that integrates comprehensive embodied reasoning capabilities, spanning embodied cognition, task planning, correction, and pointing, within a single architecture toward general physical intelligence. Leveraging three automated data construction pipelines to significantly expand the data coverage of critical capabilities, we build a large-scale data system of over 15B tokens, and design a multi-task balanced RL recipe to alleviate heterogeneous task conflicts. We further introduce a Planner-Grounder-Corrector (PGC) closed-loop framework that enables a single model to autonomously execute and self-correct over long-horizon tasks. With only 8B parameters, Embodied-R1.5 achieves SOTA on 16 out of 24 embodied VLM benchmarks, surpassing leading models like Gemini-Robotics-ER-1.5 and GPT-5.4. Benefiting from the internalized embodied capabilities, Embodied-R1.5 can be fine-tuned into a VLA with only a small amount of data, outperforming leading VLA models like π_{0.5} across 4 popular manipulation benchmark suites. We further conduct extensive zero-shot real-robot experiments, validating performance in instruction following, affordance grounding, articulated object manipulation, and long-horizon complex tasks, demonstrating strong generalization to the physical world. We open-source model weights, datasets, training code, and EmbodiedEvalKit, an evaluation framework tailored for embodied tasks, to facilitate future research in EFMs.

The effectiveness of Direct Preference Optimization (DPO) depends on preference data that reflect the quality differences that matter in multimodal tasks. Existing pipelines often rely on off-policy perturbations or coarse outcome-based signals, which are not well suited to fine-grained visual reasoning. We propose rDPO, a preference optimization framework based on instance-specific rubrics. For each image-instruction pair, we create a checklist-style rubric of essential and additional criteria to score responses from any possible policies. The instruction-rubric pool is built offline and reused during the construction of on-policy data. On public reward modeling benchmarks, rubric-based prompting massively improves a 30B-A3B judge and brings it close to GPT-5.4. On public downstream benchmarks, rubric-based filtering raises the macro average to 82.69, whereas outcome-based filtering drops it to 75.82 from 81.14. When evaluating scalability on a comprehensive benchmark, rDPO achieves 61.01, markedly outperforming the style-constrained baseline (52.36) and surpassing the 59.48 base model. Together, these results show that visual preference optimization benefits from combining on-policy data construction with instance-specific criterion-level feedback.

The viability of chain-of-thought (CoT) monitoring hinges on models being unable to reason effectively in their latent representations. Yet little is known about the limits of such latent reasoning in LLMs. We test these limits by studying whether models can discover multi-step planning strategies without supervision on intermediate steps and execute them latently, within a single forward pass. Using graph path-finding tasks that precisely control the number of required latent planning steps, we uncover a striking limitation unresolved by massive scaling: tiny transformers trained from scratch discover strategies requiring up to three latent steps, fine-tuned GPT-4o and Qwen3-32B reach five, and GPT-5.4 attains seven under few-shot prompting. Although the maximum latent planning depth models can learn during training is five, the discovered strategy generalizes up to eight latent steps at test-time. This reveals a dissociation between the ability to discover a latent strategy under final-answer supervision alone and the ability to execute it once discovered. If similar limits hold more broadly, strategies requiring multiple coordinated latent planning steps may need to be explicitly taught or externalized, lending credence to CoT monitoring.

Modern open-world agents such as OpenClaw exhibit powerful cross-environment execution capabilities yet introduce broad new safety risk sources. Meanwhile, advanced frontier AI models drastically lower attack barriers, rendering current agent alignment frameworks inadequate for real-world deployment. To tackle these emerging threats, we propose a lightweight and scalable agent safety alignment framework. Specifically, we update the agent safety taxonomy to accommodate emergent risks from Codex and OpenClaw execution scenarios. We further build a taxonomy-guided data engine with influence-function purification to train lightweight AgentDoG 1.5 variants (0.8B, 2B, 4B, and 8B parameters) using only around 1k samples, achieving comparable performance with leading closed-source models (e.g., GPT-5.4). Based on AgentDoG 1.5, we construct a highly efficient agentic safety SFT and RL training environment, which reduces deployment overhead in Docker-level environments by two orders of magnitude. Finally, we deploy AgentDoG 1.5 as a training-free online guardrail for real-time safety moderation. Extensive experimental results indicate that AgentDoG 1.5 achieves state-of-the-art performance in diverse and complex interactive agentic scenarios. All models and datasets are openly released.

Real-world spatial intelligence requires reasoning over a continuous and evolving 3D world, yet existing VLMs and tool-augmented agents largely remain tied to static, stateless inference from isolated visual observations. We introduce \textsc{S-Agent}, a spatial tool-use agentic paradigm for understanding and reasoning over continuous multi-view images and videos. By formulating spatial reasoning as spatio-temporal evidence accumulation rather than isolated frame-level prediction, S-Agent reshapes spatial perception into scene-centric understanding beyond frame-centric recognition. Specifically, S-Agent casts the VLM as a semantic planner that decides what evidence is needed, while a hierarchy of spatial tools and experts grounds objects in 2D, lifts them into 3D geometric evidence, and aggregates this evidence into high-level spatial knowledge (e.g., counting, measurement, orientation, and relative position). Additionally, a temporal memory mechanism, including Scene Memory for maintaining the evolving scene state and Agent Memory for accumulating reasoning context, enables evidence integration across frames and reasoning steps. Comprehensive experiments on multi-view and video spatial reasoning benchmarks show that S-Agent consistently improves both open-source and closed-source VLMs in a training-free manner. Beyond inference-time augmentation, supervised fine-tuning (SFT) on S-Agent-generated spatial trajectories S-300K yields S-Agent-8B, a compact spatial agent that significantly surpasses similar-scale baselines (e.g., Qwen3-VL-8B) and performs comparably to advanced closed-source models (e.g., GPT-5.4 and Gemini 3).

In industrial procurement, an LLM answer is useful only if it survives a standards check: recommended material must match operating condition, every parameter must respect a regulated threshold, and no procedure may contradict a safety clause. Partial correctness can mask safety-critical contradictions that aggregate LLM benchmarks rarely capture. We introduce IndustryBench, a 2,049-item benchmark for industrial procurement QA in Chinese, grounded in Chinese national standards (GB/T) and structured industrial product records, organized by seven capability dimensions, ten industry categories, and panel-derived difficulty tiers, with item-aligned English, Russian, and Vietnamese renderings. Our construction pipeline rejects 70.3% of LLM-generated candidates at a search-based external-verification stage, calibrating how unreliable industrial QA remains after LLM-only filtering.Our evaluation decouples raw correctness, scored by a Qwen3-Max judge validated at κ_w = 0.798 against a domain expert, from a separate safety-violation (SV) check against source texts. Across 17 models in Chinese and an 8-model intersection over four languages, we find: (i) the best system reaches only 2.083 on the 0--3 rubric, leaving substantial headroom; (ii) Standards & Terminology is the most persistent capability weakness and survives item-aligned translation; (iii) extended reasoning lowers safety-adjusted scores for 12 of 13 models, primarily by introducing unsupported safety-critical details into longer final answers; and (iv) safety-violation rates reshuffle the leaderboard -- GPT-5.4 climbs from rank 6 to rank 3 after SV adjustment, while Kimi-k2.5-1T-A32B drops seven positions.Industrial LLM evaluation therefore requires source-grounded, safety-aware diagnosis rather than aggregate accuracy. We release IndustryBench with all prompts, scoring scripts, and dataset documentation.

The rapid growth of online video platforms and AI-generated content has made reliable video guardrails a key challenge for safety and real-world deployment. While most videos can be screened through fast pattern recognition, a small subset requires deeper reasoning over temporally complex content and nuanced policy constraints. Existing approaches typically rely on large vision-language models applied uniformly across all inputs, resulting in high inference costs and inefficient allocation of computation. We propose SafeLens, a video guardrail framework that introduces a fast-and-slow inference architecture for efficient and accurate content moderation with variable computational cost across inputs. Additionally, we construct a high-quality dataset by applying influence-guided filtering to the SafeWatch Dataset, retaining only 2.4% of the original data. To further address limitations of training-time scaling, we enable test-time reasoning by augmenting the filtered data with structured Chain-of-Thought traces. Across real-world and AI-generated video benchmarks, SafeLens achieves state-of-the-art performance, outperforming strong open-source video guardrails (e.g., SafeWatch-8B, OmniGuard-7B) and closed-source models (e.g., GPT-5.4, Gemini-3.1-pro) while significantly reducing inference cost, demonstrating that efficient design serves to be more effective than scaling data or model size alone.

Current approaches to lifelong personalization operationalize relevance through semantic proximity, causing them to miss essential user information from topically unrelated interactions. To address this gap, we introduce LUCid, a benchmark designed to measure situational user-centric relevance in personalization. The benchmark consists of 1,936 realistic queries paired with interaction histories from up to 500 sessions. Across multiple architectures, our experiments show significant performance collapse when relevant context must be surfaced from semantically distant history: retrieval recall drops to near zero on the hardest instances, and response alignment remains near 50% even for state-of-the-art models such as Gemini-3-Flash, GPT-5.4, and Claude Haiku. These results expose a fundamental mismatch between the notion of relevance encoded by current systems and the situational relevance required for personalization, with direct implications for robustness and safety when critical user attributes remain undetected. LUCid enables the systematic evaluation of whether current models can surface situationally-relevant user information from previous interactions, and serves as a step toward realigning personalization with user-centered relevance.

Transformer-based language models (LMs) create hidden representations of their inputs at every layer, but only use final-layer representations for prediction. This obscures the internal decision-making process of the model and the utility of its intermediate representations. One way to elucidate this is to cast the hidden representations as final representations, bypassing the transformer computation in-between. In this work, we suggest a simple method for such casting, by using linear transformations. We show that our approach produces more accurate approximations than the prevailing practice of inspecting hidden representations from all layers in the space of the final layer. Moreover, in the context of language modeling, our method allows "peeking" into early layer representations of GPT-2 and BERT, showing that often LMs already predict the final output in early layers. We then demonstrate the practicality of our method to recent early exit strategies, showing that when aiming, for example, at retention of 95% accuracy, our approach saves additional 7.9% layers for GPT-2 and 5.4% layers for BERT, on top of the savings of the original approach. Last, we extend our method to linearly approximate sub-modules, finding that attention is most tolerant to this change.

Large-scale generative language models such as GPT-3 are competitive few-shot learners. While these models are known to be able to jointly represent many different languages, their training data is dominated by English, potentially limiting their cross-lingual generalization. In this work, we train multilingual generative language models on a corpus covering a diverse set of languages, and study their few- and zero-shot learning capabilities in a wide range of tasks. Our largest model with 7.5 billion parameters sets new state of the art in few-shot learning in more than 20 representative languages, outperforming GPT-3 of comparable size in multilingual commonsense reasoning (with +7.4% absolute accuracy improvement in 0-shot settings and +9.4% in 4-shot settings) and natural language inference (+5.4% in each of 0-shot and 4-shot settings). On the FLORES-101 machine translation benchmark, our model outperforms GPT-3 on 171 out of 182 directions with 32 training examples, while surpassing the official supervised baseline in 45 directions. We conduct an in-depth analysis of different multilingual prompting approaches, showing in particular that strong few-shot learning performance across languages can be achieved via cross-lingual transfer through both templates and demonstration examples. Finally, we evaluate our models in social value tasks such as hate speech detection in five languages and find it has limitations similar to comparable sized GPT-3 models.

Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model's internal knowledge. However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. In this paper, we present \modelname, which flares out the narrow conditioning bottleneck of DFlash through a lightweight layer-wise fusion mechanism: each draft layer attends to its own learnable combination of a broad set of target layers at negligible overhead, simultaneously injecting richer target knowledge and providing every draft layer with a distinct input. This enhanced per-layer expressiveness enables scaling the draft model to deeper architectures with consistent gains. We further scale training data from 800K to 2.4M samples to fully exploit the enlarged capacity. On six benchmarks spanning mathematical reasoning, code generation, and conversation, \modelname attains average wall-clock speedups of 5.52x on Qwen3-4B, 5.46x on Qwen3-8B, and 3.91x on GPT-OSS-20B, improving over DFlash by roughly 11\%, 8\%, and 5\% respectively. Our code is available at https://github.com/Tencent/AngelSlim.

LLM agents have incredible potential for scientific discovery applications. However, the performance of LLM agents on real-world, small molecule drug design (SMDD) tasks across diverse chemistries and targets is unclear. Current evaluation methods are either ad hoc, too simple for real-world discovery, limited in scale, or restricted to single-turn question answering. In effort to standardize the evaluation of LLM agents on small molecule design, we introduce SMDD-Bench, a challenging, multi-turn, long-horizon agentic benchmark consisting of 502 guaranteed-solvable task instances spanning 5 task types: 2D Pharmacophore Identification, Interaction Point Discovery, Scaffold Hopping, Lead Optimization, and Fragment Assembly. SMDD-Bench tasks span a wide region of chemical space and involve 102 unique protein targets. Completely solving the benchmark would require having strong chemical and biological reasoning and 3D intuition, understanding specialized tool use, and displaying planning expertise over a limited number of oracle calls. We benchmark 7 frontier open and closed source LLMs and find even the most performant LLM, GPT5.4, solves only 40.2\% of tasks. We hope SMDD-Bench provides a standardized testbed to invigorate the field towards training and evaluating LLM agents for fully autonomous computational drug design. We host a public leaderboard at smddbench.com .