← All posts
Competition runARC Prize 2026 — ARC-AGI-3 · Kaggle · $850,000 prize pool

ARC Prize 2026 (ARC-AGI-3): A Bronze-Medal-Range Run and Every Rejected Idea

ARC Prize 2026 ARC-AGI-3 on Kaggle: 9-hour no-internet runs, closed November 2, 2026, results Dec 4. Our rank: 313 of 4,926, bronze range. Every idea shown.

313 / 4,926
Public rank (final)
Bronze range
Medal standing
32.51
Best public score
14 · 10 rejected
Experiments

Manas Joshi·August 11, 2026·Updated November 14, 2026·9 min read

ARC Prize 2026 — ARC-AGI-3: competition facts
  • Where: Kaggle, $850,000 total prize pool
  • Format: interactive game environments the agent has never seen; scored on levels completed and action efficiency
  • Runtime: notebook submissions on RTX 6000 (48 GB) GPUs, no internet, up to 9 hours
  • Dates: submissions closed 2 Nov 2026 · papers due 8 Nov · results announced 4 Dec
  • Field: 4,926 teams on the final public leaderboard
  • More: all ARC-AGI-3 facts, dates and FAQ
🗓️
Updated 14 Nov 2026. Submissions are closed. On the public leaderboard we are rank 313 of 4,926 (top 6.4%) with a best public score of 32.51 — inside the bronze-medal range, where the cut-off is rank 492, and 0.46 below the silver line at 32.97. The leader is at 75.56. Medals are decided on the hidden games and announced on 4 December, so this can still move. We will update this page then.

The short version

We ran ResearchForge on ARC-AGI-3 from August to November. It recorded 18 hypotheses and 14 experiments; 10 were rejected with a written reason, and one — fine-tuning the model on its own winning games — was validated. The biggest single jump, though, did not come from our own ideas: it came on 1 October, when we rebuilt on a strong public notebook. We say that plainly because the point of keeping a record is that it tells you the truth about what worked.

Score timeline (public leaderboard)

DatePublic scoreWhat changed
1–5 Aug0.08First agents: probe-based and graph-frontier exploration
10 Aug1.21The “duck” agent harness + Qwen3.6-27B (FP8) served with vLLM — the base we built on
18 Aug1.55Model upgrade Qwen3.6 → Qwen3.8-27B
25 Aug1.94LoRA fine-tune (SFT / STaR) on 9 of the agent’s own winning games — validated
3 Sep3.62Re-baseline on a faster NVFP4 build with multi-token prediction
5–28 Sep2.08–3.86A/B rounds: reasoning steps, early stopping, strategy switching, repeat runs
1 Oct21.06Rebuilt on a strong public notebook by dfranzen (public score 27.89)
30 Oct32.51Our best public score, on our own variant of that pipeline
2 Nov313thSubmissions closed: rank 313 of 4,926, inside the bronze-medal range

The jump from 3.62 to 21.06 between September and October is not an experiment of ours paying off — it is the public notebook we built on. From there, our own variants took the best public score to 32.51.

What worked

  • ▸A better base model. Moving from Qwen3.6 to Qwen3.8-27B lifted the score from 1.21 to 1.55.
  • ▸Fine-tuning on the agent’s own wins. A small LoRA fine-tune (SFT / STaR) on just 9 of its winning games lifted 1.55 to 1.94 (+25%). This was the one experiment ResearchForge validated.
  • ▸A faster quantized build. NVFP4 with multi-token prediction meant more actions inside the 9-hour limit.
  • ▸Building on strong public work. The largest gain of the competition.

What we rejected, and why

Every one of these stayed in the record with its reason, so none of them got re-tried by accident:

IdeaWhy it was rejected
Executable world model + internal planningBetter offline (5 vs 4 levels) but 0.58 on the leaderboard vs 1.21 — the offline proxy failed
Cross-game memoryPositive offline, 1.35 on the leaderboard vs 1.94 without it
More fine-tuning data (9 → 22 games)1.20 vs 1.94 — naive data scaling hurt
Carry the world model across levels2–3 levels vs a baseline of 4, inside the noise band
Lower analyzer temperature (0.6 → 0.4)Runs of 7 and 4 vs 4 — within run-to-run noise, not a reliable lever
Longer retained contextSame 4 levels as the baseline
Cross-game mechanic priorsNo lift; the model likely already knows these mechanics
Centre-prior click tie-breakWorse than baseline offline
Graph-frontier exploration on the real 25 gamesNo change vs baseline

The lesson: our offline benchmark lied, and so does a single submission

We screened ideas on an offline benchmark — levels completed across the 25 public games. It helped rule things out, but it did not predict the leaderboard: the world-model idea scored higher offline and far lower on the hidden games, and cross-game memory did the same. Offline runs also wandered between 2 and 4 levels with no change at all.

The leaderboard is noisy too. Our same “v5” configuration, submitted on different days in late October, scored anywhere from 21.75 to 32.51. That is why a public rank is a snapshot, not a verdict — and why ResearchForge treats every result as evidence with a confidence, rather than a pass/fail gate.

🧪
Where ResearchForge helped: not by inventing the winning idea, but by keeping 18 hypotheses and 14 experiments honest — each with a measurement, a decision and a reason — so that when the offline benchmark and the leaderboard disagreed, we could see it and stop trusting the proxy.

First published 11 August 2026; rewritten 4 November 2026 with the full run and updated 14 November with the medal lines. Final results follow on 4 December. Where our other runs stand: Kaggle standings, November 2026.

Try it on your next run

Every number in this post came from the ResearchForge record. Yours can too.