The dataset viewer is not available for this subset.
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
N1M19 (test_1018) — memory × tool sweep over 10 vision-language models (8 open, 2 API)
Full trajectories of vision-language-model agents on N1M19, an embodied search task in AI2-THOR. Each instance gives the agent 7 requirements and a pool of 20 ProcTHOR houses; exactly one house (the golden one) meets all 7. The agent walks through the houses — it sees a first-person camera view and a minimap every step — and has to submit the golden house. Random choice scores 1 in 20.
All 10 models are complete.
Sweep
- Instances: 50 — 25 per track. Explicit requirements state the property ("The home must have exactly 3 bedrooms."); implicit ones describe a need it implies ("I need a separate space for me and one for overnight guests."). Each instance has three tiers by how close the 19 distractor houses come: easy (each meets 0–1 of the 7 requirements), medium (2–3), hard (4–5). So every agent config runs 2 tracks × 3 tiers × 25 instances = 150 sessions.
- Agent configs (12) = memory × tools:
- memory — none (only the current step); text (Recent-50): the last 50 turns of the agent's own observations, capped at 10,500 characters; multimodal (Recent-50): the same plus 128×96 thumbnails of the last ≤ 10 camera views;
- tools — none; sketchpad: a persistent 1,500-character notepad the agent edits itself; retriever: dense search (all-MiniLM-L6-v2, top 3) over the session's observations, results shown next turn; both.
- Budget: 60 steps per house, 1,200 per session. Actions: move/turn/look, switch house, submit, tool call.
- Inference: vLLM 0.32.0, temperature 0, thinking disabled, ≤ 768 output tokens, 32k context, ≤ 16 images per prompt. FP8 weights for GLM-4.6V and Qwen3.8-Flash-Next, BF16 for the rest. A reply is parsed as its first complete JSON object; a bare verb counts as a navigation action.
- API models (GPT-6.1 Sol, Claude Opus 5.5): served by Amazon Bedrock —
us.openai.gpt-6.1-solthrough Bedrock's OpenAI-compatible endpoint,anthropic.claude-opus-5-5through the Anthropic SDK — with the same prompts, agent configs, budgets and reply parsing as the open models. What differs: each provider's default reasoning (no effort setting sent; the reasoning itself is not returned, soresponse.raw_textholds only the reply); temperature at each API's fixed default (GPT-6.1 Sol accepts only 1, Claude Opus 5.5 takes no sampling parameters), so these runs are not greedy; ≤ 16,384 output tokens, because thinking counts against the limit. Claude Opus 5.5'sreasoning_extractionsafeguard refuses some requests — mostly in text-memory configs, whose prompt carries the agent's own past observations and thoughts; a refused request was sent again up to 3 times to the same model (never to another model), and a step still refused counts as a parse error. - Reliability: the servers were watched for stalls and restarted when one stopped serving; the agents retried patiently (up to 12 times with backoff) through each restart, and any session that still had to take a fallback step was re-run from scratch — the released sessions contain no fallback steps. vLLM batching is not bitwise deterministic, so re-runs at temperature 0 can differ in detail.
Results — correct out of 150
Each config is 150 sessions: 2 tracks (explicit, implicit) × 3 tiers (easy, medium, hard) × 25 instances. A session explores up to 20 candidate houses and submits one; exactly one meets all 7 requirements. Random choice: 7.5 / 150.
| model | memory | none | sketchpad | retriever | sketchpad + retriever |
|---|---|---|---|---|---|
| Qwen3.8-Flash-Next (FP8) | none | 27 | 80 | 42 | 94 |
| Qwen3.8-Flash-Next (FP8) | text (Recent-50) | 84 | 90 | 86 | 91 |
| Qwen3.8-Flash-Next (FP8) | multimodal (Recent-50) | 80 | 87 | 82 | 84 |
| GLM-4.6V (FP8) | none | 13 | 16 | 13 | 14 |
| GLM-4.6V (FP8) | text (Recent-50) | 33 | 33 | 27 | 25 |
| GLM-4.6V (FP8) | multimodal (Recent-50) | 37 | 30 | 33 | 30 |
| Gemma-4-31B | none | 12 | 53 | 43 | 57 |
| Gemma-4-31B | text (Recent-50) | 43 | 68 | 62 | 69 |
| Gemma-4-31B | multimodal (Recent-50) | 46 | 64 | 65 | 83 |
| Qwen3.8-27B | none | 22 | 60 | 48 | 73 |
| Qwen3.8-27B | text (Recent-50) | 89 | 93 | 90 | 96 |
| Qwen3.8-27B | multimodal (Recent-50) | 92 | 90 | 95 | 86 |
| GLM-4.6V-Flash | none | 24 | 25 | 20 | 16 |
| GLM-4.6V-Flash | text (Recent-50) | 28 | 25 | 30 | 21 |
| GLM-4.6V-Flash | multimodal (Recent-50) | 27 | 23 | 23 | 25 |
| Qwen3-VL-8B | none | 24 | 23 | 20 | 20 |
| Qwen3-VL-8B | text (Recent-50) | 21 | 25 | 22 | 19 |
| Qwen3-VL-8B | multimodal (Recent-50) | 24 | 28 | 28 | 27 |
| Qwen3.5-9B | none | 15 | 27 | 16 | 36 |
| Qwen3.5-9B | text (Recent-50) | 42 | 36 | 46 | 42 |
| Qwen3.5-9B | multimodal (Recent-50) | 44 | 51 | 58 | 48 |
| Gemma-4-26B-A4B | none | 23 | 45 | 18 | 33 |
| Gemma-4-26B-A4B | text (Recent-50) | 35 | 30 | 35 | 41 |
| Gemma-4-26B-A4B | multimodal (Recent-50) | 31 | 27 | 33 | 25 |
| GPT-6.1 Sol | none | 42 | 113 | 32 | 111 |
| GPT-6.1 Sol | text (Recent-50) | 121 | 123 | 120 | 120 |
| GPT-6.1 Sol | multimodal (Recent-50) | 122 | 119 | 124 | 117 |
| Claude Opus 5.5 | none | 63 | 120 | 95 | 119 |
| Claude Opus 5.5 | text (Recent-50) | 120 | 122 | 119 | 124 |
| Claude Opus 5.5 | multimodal (Recent-50) | 121 | 126 | 124 | 126 |
Per-cell numbers (accuracy, mean satisfaction of the chosen house, steps, houses visited, tool calls,
parse errors) are in results_by_cell.csv.
Integrity
| model | model id | sessions | turns | parse errors | tool calls | length stops | fallback steps | thinking turns |
|---|---|---|---|---|---|---|---|---|
| Qwen3.8-Flash-Next (FP8) | Qwen/Qwen3.8-Flash-Next-FP8 |
1,800 | 525,991 | 0.23% | 4,360 | 165 | 0 | 0 |
| GLM-4.6V (FP8) | zai-org/GLM-4.6V-FP8 |
1,800 | 1,248,285 | 0.03% | 1,005 | 40 | 0 | 0 |
| Gemma-4-31B | google/gemma-4-31B-it |
1,800 | 1,054,780 | 0.06% | 7,925 | 15 | 0 | 0 |
| Qwen3.8-27B | Qwen/Qwen3.8-27B |
1,800 | 419,563 | 0.09% | 3,329 | 84 | 0 | 0 |
| GLM-4.6V-Flash | zai-org/GLM-4.6V-Flash |
1,800 | 307,937 | 1.23% | 3,404 | 21 | 0 | 0 |
| Qwen3-VL-8B | Qwen/Qwen3-VL-8B-Instruct |
1,800 | 1,416,406 | 0.78% | 3,489 | 48 | 0 | 0 |
| Qwen3.5-9B | Qwen/Qwen3.5-9B |
1,800 | 583,953 | 0.27% | 6,139 | 313 | 0 | 0 |
| Gemma-4-26B-A4B | google/gemma-4-26B-A4B-it |
1,800 | 1,573,961 | 6.54% | 7,279 | 105 | 0 | 0 |
| GPT-6.1 Sol | us.openai.gpt-6.1-sol (Amazon Bedrock) |
1,800 | 936,488 | 0.003% | 138,378 | 15 | 0 | — |
| Claude Opus 5.5 | anthropic.claude-opus-5-5 (Amazon Bedrock) |
1,800 | 293,418 | 0.65% | 22,430 | 0 | 0 | — |
A parse error is a reply without a complete JSON object; that step passes with no action. Gemma-4-26B-A4B's
rate is concentrated in implicit-track runs with multimodal memory (24–39% of their turns): there it often
writes the fields in the prompt's "field" — description layout, or as "observation": …, "thoughts": …
without the enclosing braces. Length stops are replies cut at the 768-token limit (16,384 for the API models). Claude Opus 5.5's parse errors are steps it refused even after the retries (above). Thinking turns are not counted for the API models (—): their reasoning is not returned.
MeMento: learned memory vs memory baselines (memento/)
memento/ adds 4,650 sessions (150 per row, on the instances of the sweep): MeMento (the backbone reads each house the agent leaves; an adapter turns that into one soft verdict token per requirement), its controls, memory baselines (whole log, summaries, text records, retrieval, Recent-50, none) and reader-free ablations (streaming belief network, student, free tokens), on Qwen3.8-27B, Qwen3-VL-8B and Gemma-4-26B-A4B. Layout, rows and files: memento/README.md; full report: memento/RESULTS.md.
Correct out of 150 (same policy, budgets and instances in a column):
| memory | Qwen3.8-27B | Qwen3-VL-8B | Gemma-4-26B-A4B |
|---|---|---|---|
| MeMento | 107 | 32 | 90 |
| MeMento, no revisit sentence | 103 | 31 | 95 |
| per-house summary | 100 | — | — |
| whole log | 98 | — | — |
| retrieval | 98 | — | — |
| text verdict records | 97 | — | — |
| Text Recent-50 | 89 | 23 | 30 |
| Multimodal Recent-50 | 81 | 26 | 27 |
| no memory | 63 | 18 | 12 |
| ablation: streaming MeMento (no reader), run 1 | 95 | — | — |
| ceiling: true verdicts | 144 | — | — |
These are separate runs from the sweep above (the MeMento experiment re-ran its own baselines with its client).
Files
| file | entries | size | sha256 |
|---|---|---|---|
qwen38_flash_next_fp8/trajectories.zip |
3,807 | 0.16 GB | 08f0a2945f20… |
qwen38_flash_next_fp8/images/explicit_easy.zip |
86,506 | 1.81 GB | d55abfe22a53… |
qwen38_flash_next_fp8/images/explicit_hard.zip |
248,842 | 5.80 GB | 85bee8f06d99… |
qwen38_flash_next_fp8/images/explicit_medium.zip |
116,542 | 2.59 GB | 1a3e26e9aa27… |
qwen38_flash_next_fp8/images/implicit_easy.zip |
118,686 | 2.40 GB | 3bd358de8443… |
qwen38_flash_next_fp8/images/implicit_hard.zip |
312,350 | 7.17 GB | 14500b211584… |
qwen38_flash_next_fp8/images/implicit_medium.zip |
169,056 | 3.64 GB | c06d0fca3b91… |
glm46v_fp8/trajectories.zip |
3,829 | 0.39 GB | 97dc0b7d0015… |
glm46v_fp8/images/explicit_easy.zip |
381,230 | 7.36 GB | f00c24858f6f… |
glm46v_fp8/images/explicit_hard.zip |
471,486 | 10.55 GB | 95332d1d5f81… |
glm46v_fp8/images/explicit_medium.zip |
387,590 | 8.25 GB | 4fc5c75d8799… |
glm46v_fp8/images/implicit_easy.zip |
327,786 | 6.19 GB | 36f6ab173041… |
glm46v_fp8/images/implicit_hard.zip |
502,296 | 11.19 GB | f9c88ba6e8cb… |
glm46v_fp8/images/implicit_medium.zip |
426,182 | 8.54 GB | 58bf6914728e… |
gemma4_31b/trajectories.zip |
3,846 | 0.31 GB | 651386c53e01… |
gemma4_31b/images/explicit_easy.zip |
243,928 | 4.47 GB | d4e57d86aa41… |
gemma4_31b/images/explicit_hard.zip |
279,976 | 6.39 GB | f63d07eb1293… |
gemma4_31b/images/explicit_medium.zip |
199,998 | 4.08 GB | 109ceb1ff9fb… |
gemma4_31b/images/implicit_easy.zip |
501,252 | 9.10 GB | 1cd63979a5db… |
gemma4_31b/images/implicit_hard.zip |
449,462 | 10.15 GB | 0bc08fdd8e2e… |
gemma4_31b/images/implicit_medium.zip |
434,944 | 8.60 GB | 29705b05732f… |
qwen38_27b/trajectories.zip |
3,779 | 0.14 GB | 00b2e7bab179… |
qwen38_27b/images/explicit_easy.zip |
77,084 | 1.55 GB | 4cc2d8f2fe3f… |
qwen38_27b/images/explicit_hard.zip |
254,484 | 5.88 GB | 99ac6b2a1d5a… |
qwen38_27b/images/explicit_medium.zip |
98,932 | 2.10 GB | 8cf49d8e5559… |
qwen38_27b/images/implicit_easy.zip |
140,662 | 2.58 GB | d022de2443b4… |
qwen38_27b/images/implicit_hard.zip |
180,632 | 4.14 GB | ba76eaac9965… |
qwen38_27b/images/implicit_medium.zip |
87,332 | 1.81 GB | 012b4f674d1d… |
glm46v_flash/trajectories.zip |
3,812 | 0.09 GB | e1080e24f209… |
glm46v_flash/images/explicit_easy.zip |
132,784 | 2.55 GB | bd7d00882292… |
glm46v_flash/images/explicit_hard.zip |
97,390 | 2.24 GB | 47f55b52622e… |
glm46v_flash/images/explicit_medium.zip |
111,222 | 2.37 GB | ccfc4f8e3a1f… |
glm46v_flash/images/implicit_easy.zip |
68,924 | 1.31 GB | 3c9dca020230… |
glm46v_flash/images/implicit_hard.zip |
105,502 | 2.38 GB | 17ad0fe3b408… |
glm46v_flash/images/implicit_medium.zip |
100,052 | 2.07 GB | bb93f7e409da… |
qwen3vl8b/trajectories.zip |
3,802 | 0.36 GB | d73235f81870… |
qwen3vl8b/images/explicit_easy.zip |
484,618 | 9.15 GB | 2465ac59e8f0… |
qwen3vl8b/images/explicit_hard.zip |
380,954 | 8.47 GB | f9e40065c0c9… |
qwen3vl8b/images/explicit_medium.zip |
421,960 | 8.83 GB | b45f5fa4c6c0… |
qwen3vl8b/images/implicit_easy.zip |
506,326 | 9.32 GB | a9c5d451a45d… |
qwen3vl8b/images/implicit_hard.zip |
515,618 | 11.25 GB | 9a8986ae994a… |
qwen3vl8b/images/implicit_medium.zip |
523,336 | 10.33 GB | ee871085f305… |
qwen35_9b/trajectories.zip |
3,790 | 0.14 GB | 94d884861031… |
qwen35_9b/images/explicit_easy.zip |
180,052 | 3.45 GB | dbf94a80f4cd… |
qwen35_9b/images/explicit_hard.zip |
221,316 | 4.93 GB | 12a211db83bb… |
qwen35_9b/images/explicit_medium.zip |
184,754 | 4.01 GB | 3f94c2c3964d… |
qwen35_9b/images/implicit_easy.zip |
158,110 | 2.95 GB | 6c851916e6fd… |
qwen35_9b/images/implicit_hard.zip |
235,464 | 5.58 GB | 0853cb5cea72… |
qwen35_9b/images/implicit_medium.zip |
188,210 | 3.71 GB | 93c1b827d54d… |
gemma4_26b_a4b/trajectories.zip |
3,900 | 0.44 GB | d31a7cea5462… |
gemma4_26b_a4b/images/explicit_easy.zip |
473,786 | 9.01 GB | cb0e1b92213c… |
gemma4_26b_a4b/images/explicit_hard.zip |
480,316 | 10.86 GB | 4d2ec559e2c9… |
gemma4_26b_a4b/images/explicit_medium.zip |
451,718 | 9.38 GB | 2ed8472a20fa… |
gemma4_26b_a4b/images/implicit_easy.zip |
549,610 | 10.17 GB | 01bfe3658d76… |
gemma4_26b_a4b/images/implicit_hard.zip |
613,024 | 13.76 GB | 1755b9020797… |
gemma4_26b_a4b/images/implicit_medium.zip |
579,468 | 11.59 GB | 2c679990e56f… |
gpt61_sol/trajectories.zip |
3,761 | 0.43 GB | cd3312754aa5… |
gpt61_sol/images/explicit_easy.zip |
226,054 | 4.45 GB | 7333d3c6d643… |
gpt61_sol/images/explicit_hard.zip |
431,280 | 9.72 GB | 7c9a4f30b776… |
gpt61_sol/images/explicit_medium.zip |
246,474 | 5.32 GB | 18bcfb7c5c43… |
gpt61_sol/images/implicit_easy.zip |
229,738 | 4.24 GB | 68f3a5d3118e… |
gpt61_sol/images/implicit_hard.zip |
495,924 | 11.32 GB | c752d55e9e68… |
gpt61_sol/images/implicit_medium.zip |
243,506 | 5.02 GB | ef3e2a762fa0… |
claude_opus55/trajectories.zip |
3,747 | 0.15 GB | a877cf5b8a84… |
claude_opus55/images/explicit_easy.zip |
57,136 | 1.19 GB | 7d7aac3529a7… |
claude_opus55/images/explicit_hard.zip |
184,922 | 4.12 GB | cf4ae891553f… |
claude_opus55/images/explicit_medium.zip |
72,120 | 1.58 GB | cd3e3d37014c… |
claude_opus55/images/implicit_easy.zip |
74,202 | 1.44 GB | 8c84acadee19… |
claude_opus55/images/implicit_hard.zip |
116,282 | 2.66 GB | e7289c7e2eb0… |
claude_opus55/images/implicit_medium.zip |
82,174 | 1.72 GB | 6231f3e037ae… |
benchmark.zip |
4 | 2.8 MB |
Total: 351.4 GB. Each model directory has a manifest.json with full SHA-256s.
<model>/trajectories.zip—<model>/<run>/<instance>/interaction.jsonfor every session (all turns:prompt, rawresponse,parsedaction,tool_response,execution, house and step indices,is_golden), plus each run'sresults.jsonl(one row per session: chosen vs golden house, correctness, satisfaction ratio, steps, houses visited, tool calls, parse errors, token counts) andsummary.json, the per-session logs, and the model'sintegrity_check.jsonandsweep_grid.{json,md}. A run is namedN1M19_<track>_<tier>_<config>_full.results.jsonlholds one row per session, from its final attempt: two Gemma-4-31B sessions that had already finished were re-run because they had taken a fallback step during a server outage, and their earlier rows are dropped (listed underdropped_result_rowsin that model'smanifest.json).<model>/images/<track>_<tier>.zip— the camera (turn_NNNN_ego.jpg) and minimap (turn_NNNN_map.jpg) images of every step, at<model>/<run>/<instance>/images/.benchmark.zip— the 50 instances (N1M19_explicit.json,N1M19_implicit.json: requirements, the 58-house candidate set with per-house satisfied requirements and tier windows, ground truth), their README (how they were generated and validated), and the ProcTHOR-10K train-house feature cache used to check them.
The houses come from ProcTHOR-10K (AI2). The requirements of the 38 instances added for test_1018 were written by an LLM (see the benchmark README). Model outputs are subject to each model's own license terms.
- Downloads last month
- 252