The Dataset Viewer has been disabled on this dataset.

Navigation test predictions

Predictions of every model on the FrozenLake navigation test sets (one directory per test set and model).

  • preds.jsonl.gz: one line per lake of the set, in the set's row order. External models: the final answer text, the finish reason, token usage and both readings (strict = one final \boxed{L,R,...}, the in-house rule; lenient = the last box, tolerant of format). The answer is right when its route, replayed on the true grid from the start, reaches the goal exactly on its last move (any legal route counts). In-house arms: per draft the latent codes, the reader's answer and right / wrong, and the answer without a drawing.
  • responses.jsonl.gz: every record of an external run, retries included. run_config.json: its settings.
  • info.json: the set's rows sha256, lakes answered, accuracy, and the sha256 of every file.

Accuracy in the table: external runs = right answers / all lakes of the set; in-house arms = the mean single-draft accuracy over the lakes answered (an arm run on part of a set is marked "answered lakes"). Per-size and per-group numbers can be computed from preds.jsonl.gz.

set model kind lakes answered accuracy
nav3x3x BAGEL-7B-MoT (bagel) run 144 144 strict 0.00, lenient 0.00
nav3x3x Claude Haiku 4.5 (claude-haiku-4.5) run 144 144 strict 17.36, lenient 17.36
nav3x3x Claude Haiku 5.5 (claude-haiku-5.5) run 144 144 strict 99.31, lenient 99.31
nav3x3x Claude Opus 5.5 (claude-opus-5.5) run 144 144 strict 99.31, lenient 99.31
nav3x3x Claude Sonnet 5.5 (claude-sonnet-5.5) run 144 144 strict 98.61, lenient 98.61
nav3x3x GLM-4.1V (8192; the 3 cut lakes re-asked at 16384) (glm-4.1v-9b-thinking) run 144 144 strict 0.00, lenient 15.28
nav3x3x HyLaR-7B (hybrid latent reasoning) (hylar) run 144 144 strict 0.00, lenient 2.08
nav3x3x Final model: PPO v3ppo_nav310f38r38s48_s42 step 40 from SFT mix3 4800 (adapter 349b4133), MAP r38, reader nat310 (aa7e0a92) (inhouse-final-ppo-step40) inhouse 144 144 single_draft 40.28
nav3x3x Final chain warm-up: SFT mix3 step 4800 (adapter 3767ff98), reader nat310 (aa7e0a92) (inhouse-final-sft4800) inhouse 144 144 single_draft 33.42
nav3x3x 3-6 chain: PPO main s42 step 76, reader n08b (41648ae2); not the final model (inhouse-ppo-main-s42-step76) inhouse 144 144 single_draft 50.61
nav3x3x 3-6 chain: SFT c4800, reader n08b (41648ae2); not the final model (inhouse-sft4800) inhouse 144 144 single_draft 47.57
nav3x3x LatentUM base (no drawing) (latentum) run 144 144 strict 4.17, lenient 5.56
nav3x3x LVR-7B (latent visual reasoning, 8 steps) (lvr) run 144 144 strict 0.00, lenient 0.00
nav3x3x Mirage (VSP direct, latent tokens) (mirage) run 144 144 strict 57.64, lenient 57.64
nav3x3x Monet-7B (latent visual tokens) (monet) run 144 144 strict 1.39, lenient 1.39
nav3x3x Qwen2.5-VL-7B-Instruct (qwen2.5-vl-7b-instruct) run 144 144 strict 6.94, lenient 6.94
nav3x3x Qwen3-VL-8B-Thinking (qwen3-vl-8b-thinking) run 144 144 strict 37.50, lenient 47.22
nav3x3x SkiLa-7B (latent sketches) (skila) run 144 144 strict 0.00, lenient 1.39
nav597 BAGEL-7B-MoT (bagel) run 597 597 strict 0.00, lenient 0.00
nav597 Claude Haiku 4.5 (claude-haiku-4.5) run 597 597 strict 49.75, lenient 49.75
nav597 Claude Haiku 5.5 (thinking off) (claude-haiku-5.5-nothink) run 597 597 strict 94.14, lenient 94.30
nav597 Claude Haiku 5.5 (claude-haiku-5.5) run 597 597 strict 99.33, lenient 99.33
nav597 Claude Opus 5.5 (claude-opus-5.5) run 597 597 strict 100.00, lenient 100.00
nav597 Claude Sonnet 5.5 (claude-sonnet-5.5) run 597 597 strict 99.83, lenient 100.00
nav597 GLM-4.1V (8192; the 80 cut lakes re-asked at 16384) (glm-4.1v-9b-thinking) run 597 597 strict 0.00, lenient 7.54
nav597 HyLaR-7B (hybrid latent reasoning) (hylar) run 597 597 strict 0.17, lenient 1.51
nav597 Final model: PPO v3ppo_nav310f38r38s48_s42 step 40 from SFT mix3 4800 (adapter 349b4133), MAP r38, reader nat310 (aa7e0a92) (inhouse-final-ppo-step40) inhouse 597 597 single_draft 35.45
nav597 Final chain warm-up: SFT mix3 step 4800 (adapter 3767ff98), reader nat310 (aa7e0a92) (inhouse-final-sft4800) inhouse 597 597 single_draft 25.75
nav597 3-6 chain: PPO main s42 step 76, reader n08b (41648ae2); not the final model (inhouse-ppo-main-s42-step76) inhouse 597 597 single_draft 36.39
nav597 3-6 chain: SFT c4800, reader n08b (41648ae2); not the final model (inhouse-sft4800) inhouse 597 597 single_draft 30.38
nav597 InternVL3.5-8B (internvl3.5-8b) run 597 597 strict 16.25, lenient 16.25
nav597 Kimi-VL-A3B-Instruct (kimi-vl-a3b-instruct) run 597 597 strict 0.00, lenient 0.00
nav597 Kimi-VL-A3B-Thinking-2506 (kimi-vl-a3b-thinking-2506) run 597 597 strict 5.36, lenient 6.20
nav597 LatentUM base (no drawing) (latentum) run 597 597 strict 1.84, lenient 2.01
nav597 LVR-7B (latent visual reasoning, 8 steps) (lvr) run 597 597 strict 0.00, lenient 0.00
nav597 MiMo-VL-7B-RL-2508 (mimo-vl-7b-rl-2508) run 597 597 strict 12.56, lenient 13.23
nav597 Mirage (VSP direct, latent tokens) (mirage) run 597 597 strict 23.62, lenient 23.62
nav597 Monet-7B (latent visual tokens) (monet) run 597 597 strict 0.84, lenient 0.84
nav597 Qwen2.5-VL-7B-Instruct (qwen2.5-vl-7b-instruct) run 597 597 strict 0.84, lenient 0.84
nav597 Qwen3-VL-8B-Instruct (qwen3-vl-8b-instruct) run 597 597 strict 8.04, lenient 15.41
nav597 Qwen3-VL-8B-Thinking (qwen3-vl-8b-thinking) run 597 597 strict 9.05, lenient 31.99
nav597 SkiLa-7B (latent sketches) (skila) run 597 597 strict 0.00, lenient 0.34
navbig_test BAGEL-7B-MoT (bagel) run 1200 600 strict 0.00, lenient 0.00
navbig_test Claude Haiku 4.5 (claude-haiku-4.5) run 1200 600 strict 5.08, lenient 5.08
navbig_test Claude Haiku 5.5 (claude-haiku-5.5) run 1200 600 strict 49.42, lenient 49.42
navbig_test Claude Opus 5.5 (claude-opus-5.5) run 1200 1200 strict 99.92, lenient 100.00
navbig_test Claude Sonnet 5.5 (claude-sonnet-5.5) run 1200 1200 strict 99.83, lenient 99.83
navbig_test GLM-4.1V-9B-Thinking (16384) (glm-4.1v-9b-thinking-16k) run 1200 1200 strict 0.00, lenient 0.75
navbig_test HyLaR-7B (hybrid latent reasoning) (hylar) run 1200 600 strict 0.00, lenient 0.00
navbig_test Final model: PPO v3ppo_nav310f38r38s48_s42 step 40 from SFT mix3 4800 (adapter 349b4133), MAP r38, reader nat310 (aa7e0a92); 7x7 / 8x8 lakes only (inhouse-final-ppo-step40) inhouse 1200 600 single_draft 18.06 (answered lakes)
navbig_test Final chain warm-up: SFT mix3 step 4800 (adapter 3767ff98), reader nat310 (aa7e0a92); 7x7 / 8x8 lakes only (inhouse-final-sft4800) inhouse 1200 600 single_draft 17.33 (answered lakes)
navbig_test InternVL3.5-8B (7x7 / 8x8 lakes completed by a subset run; 9x9 / 10x10 partial) (internvl3.5-8b) run 1200 751 strict 0.00, lenient 0.00
navbig_test Kimi-VL-A3B-Thinking-2506 (kimi-vl-a3b-thinking-2506) run 1200 1200 strict 0.67, lenient 0.67
navbig_test LatentUM base (no drawing) (latentum) run 1200 1200 strict 0.17, lenient 0.17
navbig_test LVR-7B (latent visual reasoning, 8 steps) (lvr) run 1200 600 strict 0.00, lenient 0.00
navbig_test MiMo-VL-7B-RL-2508 (mimo-vl-7b-rl-2508) run 1200 1200 strict 1.92, lenient 1.92
navbig_test Mirage (VSP direct, latent tokens) (mirage) run 1200 1200 strict 0.42, lenient 0.42
navbig_test Monet-7B (latent visual tokens) (monet) run 1200 600 strict 0.08, lenient 0.08
navbig_test Qwen2.5-VL-7B-Instruct (qwen2.5-vl-7b-instruct) run 1200 1200 strict 0.00, lenient 0.00
navbig_test Qwen3-VL-8B-Thinking (qwen3-vl-8b-thinking) run 1200 1200 strict 2.25, lenient 4.25
navbig_test SkiLa-7B (latent sketches) (skila) run 1200 600 strict 0.08, lenient 0.08
Downloads last month
229