claim_id
large_stringlengths
20
20
paper_id
large_stringlengths
17
20
year
int32
2.02k
2.03k
atomic_claim
large_stringlengths
18
387
claim_index
int32
0
8
extractor_model
large_stringclasses
1 value
prompt_hash
large_stringclasses
1 value
run_id
large_stringclasses
2 values
created_at
large_stringclasses
2 values
clm_fdb0d9de496390e7
2025.emnlp-main.14
2,025
GA-DMS adaptively masks noisy textual tokens using gradient-attention similarity scores for cross-modal alignment.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_4cb53f5da8093908
2025.emnlp-main.14
2,025
GA-DMS uses masked token prediction objectives that compel the model to predict informative text tokens, enhancing fine-grained semantic representation learning.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_3c1339247a7352a9
2025.emnlp-main.14
2,025
GA-DMS achieves state-of-the-art performance on multiple text-based person retrieval benchmarks.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_5b03f70d85e91582
2025.emnlp-main.597
2,025
The perception bottleneck and token overload are key challenges in extending image-based vision-language models to the video domain.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_60924a3d49c529ca
2025.emnlp-main.597
2,025
D-CoDe is a training-free framework for extending image-pretrained vision-language models to video.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_cd331fda1a0e81d4
2025.emnlp-main.597
2,025
Dynamic compression alleviates the perception bottleneck by adaptively selecting representative frames and aggregating spatial tokens in a content-aware manner.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_9f6d2448eaa402fb
2025.emnlp-main.597
2,025
Question decomposition mitigates token overload by reformulating queries into sub-questions that guide the model to focus on distinct aspects of the video.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_b707f8e47dedcf24
2025.emnlp-main.1216
2,025
Audio-Reasoner performs deep reasoning on audio.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_4d05335fb20335b8
2025.emnlp-main.1216
2,025
CoTA is a high-quality, large-scale, multi-task dataset for audio reasoning.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_2ce54db16790d6d5
2025.emnlp-main.1216
2,025
Structured chain-of-thought training is essential for advancing audio reasoning.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_63bfc734b904c91a
2025.emnlp-main.837
2,025
Med-PRM verifies each reasoning step against established medical knowledge bases using retrieval-augmented generation.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_83c596df4488b295
2025.emnlp-main.837
2,025
Med-PRM achieves state-of-the-art performance on medical QA benchmarks and diagnostic tasks.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_b2fd29f642e0d6bc
2025.emnlp-main.837
2,025
Med-PRM integrates in a plug-and-play fashion with policy models.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_69694d12f6d94a6c
2025.emnlp-main.837
2,025
Med-PRM enables small-scale models to achieve state-of-the-art accuracy on MedQA.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_4e25a7c5519f6592
2025.emnlp-main.770
2,025
MuLA-F uses singular value decomposition to disentangle language-specific knowledge in LoRA blocks.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_2d969eede3635408
2025.emnlp-main.770
2,025
MuLA-F uses a local orthogonal tuning strategy.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_fb7f718dfb062371
2025.emnlp-main.770
2,025
MuLA-F reduces multilingual knowledge conflicts in federated social media content anomaly detection.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_4681945097ca5639
2025.emnlp-main.770
2,025
MuLA-F reduces the number of communication rounds in federated social media content anomaly detection.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_eb123ed15ab2ec9a
2025.emnlp-main.535
2,025
A taxonomy categorizes Machine Translation errors in the healthcare domain for a publicly available MT system.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_5250a78645e62a6c
2025.emnlp-main.535
2,025
An evaluation dataset for Machine Translation in healthcare for Amharic and Tigrinya is derived from pre-existing medical datasets.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_98491d0698463cf5
2025.emnlp-main.535
2,025
Machine Translation errors in healthcare most commonly occur when the source sentence includes medical terminology and procedure descriptions, synonyms, figurative language, and word order differences.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_a188cb6e6f795333
2025.emnlp-main.535
2,025
Pre-translation interventions are not effective in reducing clinical risk when the base translation model performs poorly.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_0dbd9c95d1b9c434
2025.emnlp-main.319
2,025
Glider uses a semantic global router that leverages LLMs' semantic reasoning to generate task-specific instructions from the input query for expert selection across all layers.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_6b57f65b693dfe95
2025.emnlp-main.319
2,025
Glider uses a learned local router that facilitates token-level routing decisions within each module.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_cc07592d06c38d8a
2025.emnlp-main.319
2,025
Glider retrieves the correct expert for held-in tasks.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_ec072800005d1238
2025.emnlp-main.319
2,025
Glider demonstrates compositional capabilities for held-out tasks.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_415c17f0cb9b9fe8
2025.emnlp-main.1263
2,025
Balcony freezes the pretrained large language model and inserts additional transformer layers at selected exit points, enabling depth-based dynamic inference with minimal performance degradation.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_3fdd4a0b2621aed8
2025.emnlp-main.1263
2,025
Balcony trains the additional layers using a self-distillation loss that aligns sub-model outputs with the full model outputs.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_e3d8e1c33c1c610b
2025.emnlp-main.1263
2,025
Balcony requires fewer training tokens and tunable parameters than prior dynamic inference methods.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_dd011b9ecbe0a3b7
2025.emnlp-main.1263
2,025
Balcony outperforms state-of-the-art dynamic inference methods and compression techniques on multiple models and benchmarks.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_3e5df230062b5962
2025.emnlp-main.1389
2,025
CogDual is a role-playing language agent that uses a cognize-then-respond reasoning paradigm.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_b3ec205606183535
2025.emnlp-main.1389
2,025
CogDual jointly models external situational awareness and internal self-awareness.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_57b867cd9c8eb093
2025.emnlp-main.1389
2,025
CogDual generates responses with improved character consistency and contextual alignment.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f0a238c89260c40b
2025.emnlp-main.1389
2,025
CogDual uses reinforcement learning with two general-purpose reward schemes designed for open-domain text generation.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_6d0d324e03d3fa58
2025.emnlp-main.1389
2,025
CogDual outperforms existing baselines on the CoSER, Cross-MR, and LifeChoice benchmarks.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_0e02ce390391a478
2025.emnlp-main.1389
2,025
CogDual generalizes across diverse role-playing tasks.
5
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_4f281a6ff879cea6
2025.emnlp-main.1728
2,025
LiTEx is a linguistically-informed taxonomy for categorizing free-text explanations in English.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_b150f2a26a4623b6
2025.emnlp-main.1728
2,025
A subset of the e-SNLI dataset is annotated using LiTEx.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_8f16f2b9a0a13ccb
2025.emnlp-main.1728
2,025
LiTEx is reliable.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_49fd1e06e3943068
2025.emnlp-main.1728
2,025
LiTEx aligns with Natural Language Inference labels, highlighted spans, and free-text explanations.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f655f4ac83d4e2ca
2025.emnlp-main.1728
2,025
Conditioning explanation generation on LiTEx produces explanations that are linguistically closer to human explanations than conditioning on labels or highlighted spans alone.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_2324f311c49e4a87
2025.emnlp-main.438
2,025
VELA evaluates long image captions using an LLM-Hybrid-as-a-Judge framework.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_29737e4ef7e6ce31
2025.emnlp-main.438
2,025
LongCap-Arena evaluates metrics for long image captions with human judgments on Descriptiveness, Relevance, and Fluency.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_d1b7c2eff51401d3
2025.emnlp-main.438
2,025
VELA outperforms existing automatic metrics for long image caption evaluation on LongCap-Arena.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_937752c58868bd11
2025.emnlp-main.438
2,025
VELA achieves superhuman performance on LongCap-Arena.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f9202ec711b178f6
2025.emnlp-main.205
2,025
A multimodal event extraction strategy uses a multi-level redundant feature selection mechanism.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_3c12a7c7833b3e49
2025.emnlp-main.205
2,025
A multimodal event extraction strategy leverages knowledge editing techniques to enhance event understanding in multimodal large language models.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f2c2c1524cd106dd
2025.emnlp-main.205
2,025
A multimodal event extraction strategy requires no additional parameter optimization work.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_99ab8bbc708d7f3f
2025.emnlp-main.205
2,025
A multimodal event extraction strategy that employs multi-level redundant feature selection and knowledge editing outperforms state-of-the-art baselines on the M2E2 benchmark.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_ed8f04adf12525e0
2025.emnlp-main.1392
2,025
MrGuard uses reasoning to classify multilingual prompts for safety.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_3edd8b9cecf9f8b8
2025.emnlp-main.1392
2,025
Synthetic multilingual data generation for safety classification includes culturally and linguistically nuanced variants.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_90927d2d9a16d241
2025.emnlp-main.1392
2,025
Group Relative Policy Optimization (GRPO) is a curriculum-based framework for multilingual safety classification.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_a566076dc58cabcd
2025.emnlp-main.1392
2,025
MrGuard consistently outperforms recent baselines in multilingual safety classification across in-domain and out-of-domain languages.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_6fbfae615f1e945a
2025.emnlp-main.1392
2,025
MrGuard preserves safety judgments under multilingual variations including code-switching and low-resource language distractors.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f46dfca807589acf
2025.emnlp-main.1392
2,025
MrGuard generates explanations for safety judgments that help understand language-specific risks and ambiguities in multilingual content.
5
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_c9677612627d98d1
2025.emnlp-main.472
2,025
For hierarchical text classification, HYDRA treats each level in the hierarchy as a separate classification task with its own label space.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_8044a4097e166d73
2025.emnlp-main.472
2,025
For hierarchical text classification, HYDRA uses parameter sharing and level-specific parameterization to incorporate hierarchical awareness without architectural complexity.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_31f465a067618a43
2025.emnlp-main.472
2,025
For hierarchical text classification, HYDRA consistently outperforms flat models.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_e8ea09fd50ec112a
2025.emnlp-main.472
2,025
For hierarchical text classification, HYDRA matches or exceeds the performance of complex state-of-the-art methods.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_fc364d735b37b498
2025.emnlp-main.472
2,025
In hierarchical text classification, complex components such as graph encoders, label semantics, and autoregressive decoders are often unnecessary.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_341fcc81651ed345
2025.emnlp-main.1093
2,025
Inverse confidence weighting selects examples across tasks for label-efficient supervised finetuning of LLMs.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_b38546bfc24318f4
2025.emnlp-main.1093
2,025
Inverse confidence weighting is easier to implement and less computationally intensive than more complex sampling procedures for label-efficient supervised finetuning of LLMs.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_b6a1b483238857b4
2025.emnlp-main.1093
2,025
Inverse confidence weighting achieves better accuracy than training on the complete dataset in label-efficient supervised finetuning of LLMs.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_d4232e6f62eeeafc
2025.emnlp-main.1093
2,025
Inverse confidence weighting reduces annotation costs while maintaining or improving performance compared to the best existing methods in label-efficient supervised finetuning of LLMs.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_3ab804b78d812c69
2025.emnlp-main.345
2,025
CoPriva evaluates adherence of large language models to contextual non-disclosure policies in question answering using direct and indirect attacks.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_270d899160ec0ea0
2025.emnlp-main.345
2,025
Many large language models violate user-defined policies and leak sensitive information in question answering.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_7f1254ea1a18ab5d
2025.emnlp-main.345
2,025
Violations of user-defined policies by large language models are particularly severe against indirect attacks in question answering.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_d8626a68c6396d6c
2025.emnlp-main.345
2,025
Large language models struggle to incorporate policy constraints during generation in question answering.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_2e9ce0e8f7e0b25c
2025.emnlp-main.345
2,025
Large language models exhibit a partial ability to revise outputs when explicitly prompted in question answering.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f77f7c7862ce30df
2025.emnlp-main.1578
2,025
Waste-Bench evaluates the robustness and accuracy of VLLMs on waste classification tasks in cluttered real-world environments with complex scenes and deformed objects.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_cf2ec307c99e7f83
2025.emnlp-main.614
2,025
Guard models are vulnerable to jailbreak attacks due to representation shifts in the embedding layer that cascade through the Transformer stack.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_2054b66db2c83c80
2025.emnlp-main.614
2,025
Gamma-Guard uses lightweight residual adapters inserted after the embeddings and at sparse intervals in the model.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_38b70f245163144a
2025.emnlp-main.614
2,025
Gamma-Guard adapters are initialized with zero-scaled gates to retain original behavior and then fine-tuned adversarially to denoise embeddings and refocus attention.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_3e004cd5a322ebf8
2025.emnlp-main.614
2,025
Gamma-Guard achieves high accuracy on jailbreak attack examples with minimal degradation in accuracy on clean examples.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_23363f34f235e904
2025.emnlp-main.614
2,025
Gamma-Guard's robustness improvements persist across different layer placements and model sizes.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_98e98cc13f7a9e3d
2025.emnlp-main.240
2,025
QSPEC decouples efficiency from quality by integrating low-precision joint quantization for fast drafting and high-precision weight-only quantization for accurate verification via speculative decoding.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_5619a5f9b8824503
2025.emnlp-main.240
2,025
QSPEC reuses weights and KV cache across stages to enable near-zero-cost switching without retraining or auxiliary models.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_a95be5449a372ace
2025.emnlp-main.240
2,025
QSPEC achieves speedup without quality degradation compared to high-precision baselines.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_62340f0f117292f2
2025.emnlp-main.240
2,025
QSPEC outperforms state-of-the-art speculative decoding methods in batched settings.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_a6a2fa4d38a6db6e
2025.emnlp-main.240
2,025
QSPEC supports plug-and-play deployment.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_a14173f8a7ed5437
2025.emnlp-main.240
2,025
QSPEC generalizes across model scales, quantization methods, and workloads.
5
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_c8303a1f4bbf3b95
2025.emnlp-main.1110
2,025
ArgCMV is a key point extraction dataset for long-context online discussions from actual human debates.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_1926d2d2b3ceade4
2025.emnlp-main.1110
2,025
ArgCMV exhibits higher complexity than ArgKP21 for key point extraction.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_956b9d8d507320e2
2025.emnlp-main.1110
2,025
Existing key point extraction methods do not adapt well to ArgCMV.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f863ab6c94d553a6
2025.emnlp-main.1110
2,025
ArgCMV is curated using state-of-the-art large language models.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_32750422af9cb72c
2025.emnlp-main.760
2,025
The MUCAR multilingual dataset contains visual contexts that uniquely resolve ambiguous textual expressions.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_a3a5a2dad7cc4a21
2025.emnlp-main.760
2,025
The MUCAR dual-ambiguity dataset contains pairs of ambiguous images and ambiguous textual contexts that produce a single clear interpretation through mutual disambiguation.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_7c8e3ce17d47c839
2025.emnlp-main.760
2,025
Evaluations of multimodal models on MUCAR reveal substantial performance gaps compared to human-level performance.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f9f901895779a49a
2025.emnlp-main.918
2,025
Within a hierarchical text classification task, different prompting strategies yield substantially different results from large language models.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_98b677aa32c9cb2e
2025.emnlp-main.918
2,025
Across different hierarchical text classification tasks, the relative effectiveness of a prompting strategy varies significantly.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_f3984ac6a3d5b334
2025.emnlp-main.918
2,025
HiEPS integrates the results of diverse hierarchical text classification prompting strategies for zero-shot hierarchical text classification.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_8a24ddc58c79561f
2025.emnlp-main.918
2,025
The path-valid voting mechanism selects a valid hierarchical text classification result with the highest path frequency score.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00
clm_5db0fafc530e6d0f
2025.emnlp-main.918
2,025
HiEPS outperforms single prompting strategies for zero-shot hierarchical text classification.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T094754Z
2026-05-24T09:47:55.807293+00:00