File size: 9,298 Bytes
452fc52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
{
  "artifact_schema_version": 1,
  "project": {
    "name": "rlm-local",
    "description": "Owner-measured local RLM long-context QA validation on RTX PRO 4000 Blackwell desktop GPUs.",
    "source_results_summary": "RESULTS.md",
    "source_note": "NOTE-rtx-pro-4000-rlm.md",
    "source_commit_before_publish_bundle": "17cc623"
  },
  "date": "2026-06-20",
  "hardware": {
    "host": "P620",
    "gpu": {
      "name": "NVIDIA RTX PRO 4000 Blackwell",
      "variant": "desktop",
      "count": 2,
      "vram_gb_each": 24,
      "vram_gb_total": 48,
      "memory_type": "GDDR7 ECC",
      "nvlink": false
    }
  },
  "model": {
    "name": "Qwen36-35B-A3B-Q4_K_M",
    "family": "Qwen3.6-35B-A3B",
    "parameters_b": 35,
    "quantization": "GGUF Q4_K_M",
    "serving_backend": "llama.cpp OpenAI-compatible server",
    "context_window_tokens": 131072
  },
  "technique": {
    "name": "Recursive Language Models",
    "abbreviation": "RLM",
    "reference_repo": "https://github.com/alexzhang13/rlm",
    "reference_paper": "arXiv 2512.24601",
    "rlms_version": "0.1.2",
    "root_model": "local quantized 35B",
    "in_window_max_depth": 1,
    "recursion_probe_max_depth": 2,
    "docker_compat_patch": "host-network-proxy-localhost"
  },
  "environment": {
    "rlm_repl": "Docker python:3.11-slim sandbox",
    "execution": "single-stream",
    "endpoint": "local OpenAI-compatible endpoint; network address omitted",
    "service_handling": "production local service used in place and not stopped"
  },
  "task": {
    "benchmark": "oolongbench/oolong-synth",
    "task_type": "long-context QA",
    "comparison": "plain one-call baseline versus same model wrapped with RLM",
    "slice_policy": "first 50 eligible validation rows plus next 50 eligible validation rows with zero ID overlap",
    "pooled_n": 100,
    "slices": [
      {
        "name": "validation-n50",
        "n": 50,
        "items_path_private": "data/items-validation-n50.jsonl",
        "max_context_words": 17517,
        "item_ids": [
          "oolong-115010016",
          "oolong-115010017",
          "oolong-115010018",
          "oolong-115010019",
          "oolong-115010020",
          "oolong-115010021",
          "oolong-115010022",
          "oolong-115010023",
          "oolong-115010024",
          "oolong-115010025",
          "oolong-115010026",
          "oolong-115010027",
          "oolong-115010028",
          "oolong-115010029",
          "oolong-115010030",
          "oolong-115010031",
          "oolong-115010032",
          "oolong-115010033",
          "oolong-115010034",
          "oolong-115010035",
          "oolong-115010036",
          "oolong-115010037",
          "oolong-115010038",
          "oolong-115010039",
          "oolong-115010040",
          "oolong-115010041",
          "oolong-115010042",
          "oolong-115010043",
          "oolong-115010044",
          "oolong-115010045",
          "oolong-115010046",
          "oolong-115010047",
          "oolong-115010048",
          "oolong-115010049",
          "oolong-115010050",
          "oolong-115010051",
          "oolong-115010052",
          "oolong-115010053",
          "oolong-115010054",
          "oolong-115010055",
          "oolong-115010056",
          "oolong-115010057",
          "oolong-115010058",
          "oolong-115010059",
          "oolong-115010060",
          "oolong-115010061",
          "oolong-115010062",
          "oolong-115010063",
          "oolong-115010064",
          "oolong-115010065"
        ]
      },
      {
        "name": "disjoint-n50",
        "n": 50,
        "items_path_private": "data/items-validation-disjoint-n50.jsonl",
        "max_context_words": 34377,
        "item_ids": [
          "oolong-116010200",
          "oolong-116010201",
          "oolong-116010202",
          "oolong-116010203",
          "oolong-116010204",
          "oolong-116010205",
          "oolong-116010206",
          "oolong-116010207",
          "oolong-116010208",
          "oolong-116010209",
          "oolong-116010210",
          "oolong-116010211",
          "oolong-116010212",
          "oolong-116010213",
          "oolong-116010214",
          "oolong-116010215",
          "oolong-116010216",
          "oolong-116010217",
          "oolong-116010218",
          "oolong-116010219",
          "oolong-116010220",
          "oolong-116010221",
          "oolong-116010222",
          "oolong-116010223",
          "oolong-116010224",
          "oolong-116010225",
          "oolong-116010226",
          "oolong-116010227",
          "oolong-116010228",
          "oolong-116010229",
          "oolong-116010230",
          "oolong-116010231",
          "oolong-116010232",
          "oolong-116010233",
          "oolong-116010234",
          "oolong-116010235",
          "oolong-116010236",
          "oolong-116010237",
          "oolong-116010238",
          "oolong-116010239",
          "oolong-116010240",
          "oolong-116010241",
          "oolong-116010242",
          "oolong-116010243",
          "oolong-116010244",
          "oolong-116010245",
          "oolong-116010246",
          "oolong-116010247",
          "oolong-116010248",
          "oolong-116010249"
        ]
      }
    ]
  },
  "grader": {
    "name": "boundary_v2",
    "method": "deterministic exact or boundary-aware answer containment, with digit boundaries and date aliases",
    "audit_summary": "All six Phase-1 vanilla misses were audited and a sample of Phase-1 RLM-right OOLONG answers was checked; no padded incorrect RLM positives were found in that sample."
  },
  "results": {
    "in_window": {
      "pooled": {
        "n": 100,
        "vanilla_correct": 37,
        "vanilla_accuracy": 0.37,
        "rlm_correct": 95,
        "rlm_accuracy": 0.95,
        "rlm_minus_vanilla_accuracy": 0.58,
        "total_vanilla_wall_seconds": 359.67,
        "total_rlm_wall_seconds": 1156.06,
        "average_vanilla_wall_seconds": 3.5967,
        "average_rlm_wall_seconds": 11.5606,
        "rlm_repl_iterations": 327,
        "rlm_sub_calls": 0,
        "max_context_words": 34377
      },
      "slices": [
        {
          "name": "validation-n50",
          "n": 50,
          "vanilla_correct": 19,
          "vanilla_accuracy": 0.38,
          "rlm_correct": 47,
          "rlm_accuracy": 0.94,
          "rlm_minus_vanilla_accuracy": 0.56,
          "total_vanilla_wall_seconds": 107.76,
          "total_rlm_wall_seconds": 538.08,
          "rlm_repl_iterations": 154,
          "rlm_sub_calls": 0,
          "max_context_words": 17517,
          "run_timestamp_utc": "2026-06-20T02:22:42Z"
        },
        {
          "name": "disjoint-n50",
          "n": 50,
          "vanilla_correct": 18,
          "vanilla_accuracy": 0.36,
          "rlm_correct": 48,
          "rlm_accuracy": 0.96,
          "rlm_minus_vanilla_accuracy": 0.60,
          "total_vanilla_wall_seconds": 251.91,
          "total_rlm_wall_seconds": 617.98,
          "rlm_repl_iterations": 173,
          "rlm_sub_calls": 0,
          "max_context_words": 34377,
          "run_timestamp_utc": "2026-06-20T03:33:20Z"
        }
      ],
      "interpretation": "The validated OOLONG uplift came from REPL-over-context behavior, not recursive sub-LM delegation."
    }
  },
  "beyond_window": {
    "kind": "owned synthetic stress item",
    "word_count": 165730,
    "token_estimate": {
      "method": "whitespace lower bound",
      "minimum_tokens": 165730,
      "actual_tokenizer_count_measured": false,
      "model_context_window_tokens": 131072
    },
    "vanilla": "deliberately truncated before the answer and failed",
    "rlm": "solved through the Docker REPL over the full document",
    "included_in_fair_oolong_aggregate": false
  },
  "recursion_probe": {
    "kind": "owned adversarial probe",
    "n": 5,
    "vanilla_correct": 3,
    "vanilla_accuracy": 0.60,
    "rlm_correct": 1,
    "rlm_accuracy": 0.20,
    "total_vanilla_wall_seconds": 91.84,
    "total_rlm_wall_seconds": 77.43,
    "rlm_repl_iterations": 18,
    "rlm_sub_calls": 15,
    "max_depth_configured": 2,
    "interpretation": "Sub-LM delegation fired under provocation, exactly three calls per item, but this local Q4 35B setup did not use it competently on the probe."
  },
  "caveats": [
    "This is one model on one owner-operated hardware setup, not a full benchmark.",
    "OOLONG-synth is public, so contamination is possible; it would affect both the plain baseline and RLM condition.",
    "The recursion probe has only five owned adversarial items and is directional.",
    "The beyond-window synthetic item is a capability demo and is excluded from the fair OOLONG aggregate.",
    "Raw prompts, raw contexts, per-item answers, and RLM trajectories are intentionally excluded from this public artifact."
  ],
  "privacy": {
    "public_included": [
      "aggregate accuracy",
      "aggregate latency",
      "context size statistics",
      "RLM iteration and sub-call counts",
      "OOLONG item IDs",
      "hardware/model/technique metadata"
    ],
    "private_excluded": [
      "raw prompts",
      "raw contexts",
      "per-item gold answers",
      "per-item model answers",
      "RLM trajectories",
      "local endpoint network address"
    ]
  }
}