milestone2/SYNTHESIS-github.md

SYNTHESIS — GitHub research on ARC-AGI-3 / Milestone 2 (Task 2-b)

Date: 2026-10-01 · Method: GitHub REST API (repo search ×13 queries, code search, org/user listings, README+tree fetches). Raw artifacts in this dir: searchrepo_*.json, searchcode_milestone2.json, all_repos.json (115 unique repos), tree_*.json, gh_*_README.md, orgs_arcprize.json, commits_arcagi3agents.json.

Headline

All four Milestone-2 Kaggle leaders are the same lineage: Tufa Labs Duck/TAAF harness + Qwen3.8-Flash-Next on one RTX PRO 6000, and the 20→28 gap is dominated by inference engineering of memory (FP8 KV cache → ~1.0–1.4M-token KV pools → much longer rolling per-game history), not by new agent ideas. Confirmed by three independent sources: rellik13's write-up (14.49→22.53 with zero prompt changes), the 1zuki/arg-agi public crawl doc (gh_1zuki_milestone2_crawl.md — a full config comparison of the 4 winners), and da-fr's write-up (93.43% prefix-token reuse; "with the increased context, completely turning off the original text-based world model mechanism worked best").

Repo landscape (115 unique repos found; top entries)

#	Repo	★	Pushed	LB/score	Relevance
1	da-fr/arc-agi-3-solution	9	2026-09-30	27.89 (LB #1)	⭐⭐⭐⭐⭐ Daniel Franzen's full Milestone-2 solution: patched harness + SGLang builder + patches + WRITEUP + CONFIGURATION (43KB feature reference incl. disabled experiments)
2	LohitSiriki/arc-agi-3-milestone2-solution	1	2026-09-30	22.53	⭐⭐⭐⭐⭐ rellik13's duck-18-1gc-submit: every notebook cell as .py, full run artifacts (prompts, sglang log, benchmark.json), WRITEUP with score history + memory ablation
3	Tufalabs/duck-harness	90	2026-07-01	M1 winner (12.58)	⭐⭐⭐⭐⭐ The Duck + TAAF + example run + viewer; base of every top LB entry
4	— (Kaggle only) lordhansolo arc-agi-3-milestone-2	—	—	23.84	⭐⭐⭐⭐⭐ vLLM-nightly 0.29.1rc1 + NVFP4/FP8-mixed + FP8 KV 1.417M tokens, 14 streams, 147k ctx. No GitHub repo found. This is the closest config to OUR stack
5	— (Kaggle only) richardcsaky arc-agi-3-milestone-2-submission	—	—	20.00	⭐⭐⭐ vLLM, expert-overlay checkpoint, 262k ctx, 35 seqs, int4 per-token-head KV, Marlin, expert pruning. No GitHub repo
6	Alexyskoutnev/TWIN-ARC-AGI-3	9	2026-09-17	93.3/100 (API agent)	⭐⭐⭐⭐⭐ Test-time executable digital twin (Python world model, replay-validated, plan-in-model). Best-documented offline world-model method
7	alexisfox7/PRO-LONG	458	2026-08-21	97.4% best@2 (Fable 5)	⭐⭐⭐⭐ Programmatic memory: one append-only log + agent greps it; +18pts over coding-agent baselines; arXiv 2607.20064; reproduction harness in research/arc-agi-3/
8	pbshgthm/arc-skill	89	2026-08-19	100.00 RHAE, 25/25 (Opus 5)	⭐⭐⭐⭐ Prediction-gated actions (harness refuses unplanned presses), claim grammar, one-page NOTES.md, escalation ladder
9	DriesSmit/ARC3-solution	96	2026-01-27	preview winner (12.58 StochasticGoose)	⭐⭐⭐ CNN action-change predictor for exploration; Tufa Labs origin
10	arcprize/arc-agi-3-benchmarking	52	2026-09-30	official baselines	⭐⭐⭐ Official M2-era benchmarking harness (gpt-5.4/5.6-sol-max, opus-5, gemini-3.8, grok-4.7 provider adapters); ARC-AGI-3-Agents is stale since Aug 3
11	U4AR/qwen38-arc3-rl	0	2026-09-24	baselines ~1.6	⭐⭐⭐⭐ LoRA RL of Qwen3.8-27B on Duck (2×H100): reward = level-solve + token-efficiency bonus, Dr.GRPO, verbatim history for prefix-cache/packing
12	agno-agi/arc-agi-arcade	14	2026-08-26	100.00 warm GPT-5.6	⭐⭐⭐ Cold/warm/seeded modes; "manuals" as transferable knowledge; contamination flagging
13	theredbluepill/arc-interactive	57	2026-10-01	—	⭐⭐⭐ 249 community game environments (practice distribution!)
14	samrishtt/arc-agi-3-kaggle-competition	3	2026-08-31	0.33	⭐⭐ Go-Explore graph explorer (deterministic, no ML): volatility-masked state hash, win-path caching, tiered click proposals
15	sonpham-org/arc-3	3	2026-09-28	—	⭐⭐ instrumented Tufa fork, GCP spot kit, reproduction matrix
16	dp-web4/ARC-SAGE	4	2026-09-09	94.85 (frozen, engine-reading)	⭐⭐ honest historical snapshot; world-model cartridges
17	1zuki/arg-agi	—	2026-10-01	their LB 3.40	⭐⭐⭐⭐ docs/arc-agi-3/milestone-2-public-crawl-2026-10-01.md = best third-party M2 analysis (saved: gh_1zuki_milestone2_crawl.md)
18	Shih-Yu-Yeh/arc3-dev-agi-journal	0	2026-09-26	LB 0.87→2.56	⭐⭐ 33-day dev journal, same resource bracket as us
19	54xkeee/duck-harness-glm53-flash-experiments	0	2026-09-06	—	⭐⭐ GLM-5.3-Flash on Duck, evidence-prediction loop
20	pinion05/arc-agi3-duck-harness-vllm	0	2026-08-08	—	⭐⭐ Duck vLLM repro on A6000: conv1d CUDAGraph patch

Others of note: farmountain/aera-arc3-paper (epistemic agent), DenisSud/su-JEPA (latent world model, JAX), pomagrenate/dsg-ewm (executable world model; the EWM 63.8 baseline in TWIN), kanishkpaul/arc-agi3-world-models, PrimeIntellect-ai/arc-agi-3-prime-agent, aws-samples/sample-strands-agents-bedrock-arc-agi-3, LiDeChi/arc3-world-model-agent, Alexyskoutnev TWIN site. Negative lookups: Jeroen Cottaar (jcottaar) has NO ARC repo (only packing/ariel2/seismic); NVARC3 GitHub account exists but 0 repos; "taaf arc" query → 0; Tufalabs org has duck-harness + infra forks (vllm/sglang/transformers/flashinfer forks updated 2026-09-09) + miles RL framework + ladder.

---

Deep dives — the three public Milestone-2 winners

1. da-fr (Daniel Franzen) — 27.89 — github.com/da-fr/arc-agi-3-solution

Same Duck core (LM + Python tool). Changes, by impact:

Serving/memory (the big lever). Intel W4A16 AutoRound Flash-Next + Albucino INT4 MTP draft (3 spec steps), BF16 PLE kept in host RAM; SGLang "Pennyroyal" v2.5.3 + custom patches (Low-M BF16 GEMM, spec-state memory-budget correction, Marlin scale-dtype fix, Mamba-checkpoint retention prefix-cache patch, shard prefetch patch); FP8 E4M3 KV; harness context 128 Ki (vs Duck's ~37k), server 136 Ki, drain to 58 Ki on trim (hysteresis: trim one large block, then let history regrow → 93.43% of prompt tokens served from cache); image-aware token estimation (base64 no longer counted as text); admission limited to 10 streams = server limit so returning games find their prefix resident; Mamba cache 60 slots; end-of-prefill checkpoint kept + LRU-refreshed.

Scheduler. All 25 games started at once; priority scheduler admits highest P=(A+B)·C where A = immediate score gain of finishing current level (with action-efficiency penalty, normalized by N levels), B = continuation-value lookup (3+ levels left: 8, 2: 7, 1: 5, 0: 0) faded to 0 in final 20% of runtime, C = Gaussian completion-probability falloff calibrated on 725 observed level attempts: C = 0.25·2^-(a/115)² + 0.75·2^-(t/62000)² (a=actions, t=generated tokens on level). Slot handover happens only at context eviction (protects prefix cache).

Perception. Intermediate animation frames + text timelines exposed as REPL variables (strongest perception win; hint only when "transient pixels" exist; inspired by jakobbrggen's TAAF-Anim notebook); frame_diff() object-level structured diff (moved/rotated/appeared/disappeared/recolor) on 4-connected components; images upscaled 10× (was 4×); difference images after actions; game-over fatal frame supplied.

Acting/agent. UNDO (ACTION7) exposed as a real action — "helped substantially, suspected very important on the competition set"; persistent model-defined Python functions retained for the whole game (functions only, not variables); guards: batch no-op guard, stale-state guard (blocks next action call in same snippet after ineffective one; from level 2), terminal-state guard; gameplay_changed vs board_changed (4-cell border excluded to ignore budget bar); tool output budget ~3,072 tokens with middle truncation; faithful per-action results + explicit game-over diagnosis prompts (budget-bar rule: only fully-depleted bar means budget death) and level-transfer guidance prompts.

Explicitly negative (their §6): structured world-model updates (turned OFF entirely — raw long history beats it), summarization (no clear gain), animation composite images, auto-inserted timelines, stronger verification hints, death ledgers, more elaborate priority rules.

Sampling: temp 0.7, top-p 0.95, top-k 20.

2. rellik13 / sirikilohit (LohitSiriki repo) — 22.53 — cleanest causal evidence

Score history: 0.99 → 3.81 (model: Qwen3.8-Flash-Next) → 6.42 (SGLang + more games) → 7.14 (49k ctx) → 9.96 (solved-level rule memory pinned in prompt, +2.8, the only text addition that helped) → 13.40 (Wang's harness port: stricter sandbox, object segmentation, test-ideas-with-code prompts) → 14.49 (harness fixes) → 22.53 (FP8 KV + longer history, no prompt change).

Final config: W4A16 + RadixArk FP8 PLE swapped in + RadixArk NVFP4 MTP (3 steps/4 draft tokens), SGLang Pennyroyal, FP8 KV pool 1,004,288 tokens, 16 concurrent, ctx 69,632, history trim 57,344→45,056 (was 36,864→26,624), one-pool UCB scheduler with 2,400 s slices, hierarchical host tier (48GB) for KV/Mamba, thinking mode, temp 0.6.

Why FP8 KV won: server logs showed KV pool 91–97% full with requests queued for cache space — memory, not throughput, was the bottleneck. Earlier "FP8 failed" test was confounded by MXFP8 re-quant (~25% slower decode) and flags lost to comments. Key lesson recorded in cells/21_patch_cell_v14_43_reasoning_ledger... etc: an explicit "ledger for proven/refuted assumptions" was never used by the agent; adding advice/summaries consistently hurt (each repeated line pushed out the agent's own turns at fixed ctx); restarting stuck levels with fresh context didn't help; temp 1.0 worse; expert pruning worse; higher-res images worse; fine-tune "Swift 1.5" no better per token.

3. lordhansolo — 23.84 — the vLLM route (our closest template)

No GitHub repo (notebook only), config per 1zuki crawl + Kaggle discussion 744792: vLLM nightly 0.29.1rc1 (commit e975732), mixed NVFP4/FP8 target with BF16 PLE, built-in NVFP4 MTP, FP8 E4M3 KV, 1,417,100-token pool, 147k server / 127k harness context, 14 live streams, prefix caching + FP8 indexer KV + BF16 Mamba cache, 3,918-second per-game cap, 8-wave admission. Proves NVFP4 on vLLM is fully competitive when the KV pool and history are sized right — exactly our stack.

richardcsaky (20.00, vLLM): expert overlay checkpoint, --max-model-len 262144, --max-num-seqs 35, int4 per-token-head KV, Marlin backend, async scheduling, expert pruning.

---

Adjacent breakthrough methods (API-model agents — techniques to port)

• TWIN (93.3, 23/25 games, 0.61× human actions): coding agent writes step(grid,action)->grid + goal_reached(grid) Python twin; harness refuses scored actions while twin replay fails; plans inside the twin (depth 8/20k nodes; goal search 14/30k); mismatch = counterexample that repairs code; probes 7.1% of actions; fits EWM baseline (63.8) with 7.7× fewer tokens. Frontier = goal inference (twins fit 99.4% on-log but only 68.6% off-log).
• PRO-LONG (97.4% best@2, $1,750, Fable 5): single append-only log.txt of all observations/actions/outcomes; agent retrieves with grep/Python; ~30-line system prompt; NO subagents, NO summarization, NO vector DB. Ablations: last-25-sections window degrades a lot; no-log collapses. Over matched baselines +18 pts at 4.2–5.8× fewer tokens.
• arc-skill (100.00 RHAE, 7,645 actions vs human 17,135): hard harness gate: every press needs a falsifiable prediction (cell X,Y=V; move X,Y DX,DY; region; vanish; level+1; win; change; noop) graded against the returned frame (7,627 predictions, 443 misses); plan = sequence of predicted steps, halts at first miss; one-page NOTES.md per game survives 115 compactions; escalation ladder sentences→Python→simulator→full world model (full model used once, never fitted, game won on rung below); planned sequences missed 2.9% vs single test presses 37.1%.
• agno-agi arcade: "manuals" (structured game knowledge) transfer across models — seeding GPT-5.6 manuals lifts Gemini-3.7/Grok-4.6/G by 20–40 RHAE. Suggests writing Duck prompts/knowledge from solved games into reusable per-archetype manuals.
• U4AR RL: LoRA RL on Qwen3.8-27B through the Duck; reward = level-solve + fewer-generated-tokens bonus; GRPO over (game,level) segments; verbatim replayed history → prefix-cache hits + 4.7× fewer forward passes when packing. Baselines only ~1.6 ARC score — RL from self-play on 5 games does NOT beat serving/memory engineering yet.
• samrishtt Go-Explore (0.33 LB): deterministic state-graph explorer; volatility-masked status-bar hashing (mask rows/cols with change-freq > 0.5); win-path caching + re-entry replay on regression (the "regression memory bug" fix); tiered click proposals (component centroids, small/rare objects first). Good seed for an offline/free exploration controller.

Official infra status

• arcprize/ARC-AGI-3-Agents: last real commit 2026-08-03 (URL fix). Superseded for benchmarking by arcprize/arc-agi-3-benchmarking (pushed 2026-09-30): Standard vs Provider-Adapter harnesses, manual_rolling vs continuous_conversation, native compaction (Anthropic 175k threshold, xAI encrypted reasoning replay) — i.e., ARC Prize itself is standardizing on long native context + compaction, mirroring the LB leaders.
• arcprize/ARC-AGI-Community-Leaderboard (pushed 2026-10-01): new showcase; requires scorecard URL for AGI-3.
• Tufalabs org: duck-harness (90★) + forks of vllm/sglang/transformer-engine/flash-attention updated 2026-09-09 (inference-infra experimentation) + miles RL framework (fork) + tufazip.

---

What WE can learn (our stack: Duck fork, Qwen3.8-Flash-Next-NVFP4, LB 4.54 → target 27+)

Our 4.54 sits almost exactly where rellik13 was on Sep 18 (6.42 w/ SGLang) and da-fr's early runs; the crawl doc's diagnosis of a similar 3.40 entry applies to us nearly verbatim: same solver lineage, radically different serving envelope. Ranked by expected LB impact:

1. Memory envelope first, prompts last. Enable FP8 E4M3 KV cache on our vLLM nightly (lordhansolo proves 1.42M-token pool on NVFP4 + vLLM) and raise history retention to ~45–57k with large-block hysteresis trimming (trim 57k→45k once, not 2k/turn; each trim invalidates the prefix). Add cache-occupancy telemetry (their smoking gun: pool 91–97% full + queued requests = raise pool, don't add prompts). If vLLM nightly limits us to ~8 resident seqs, cut admission to that (cache-resident useful state ≫ nominal concurrency).
2. Prefix/Mamba-state retention. Keep prefix caching ON, retain end-of-prefill Mamba checkpoints (LRU refresh), image-aware token counting (base64-as-text bug wrecks trim decisions), and re-tokenization-safe history (U4AR's verbatim-history trick improves hits + enables packing later).
3. Harness fixes bundle (all no-context-cost, rellik13's 14.49 step): label ACTION7 as UNDO; replace "The game is over." with "Level restarted." + diagnosis (budget-bar rule); remove the "ignore the border bar" advice (it is the action budget); expose mid-action animation frames (+ hint only on transient pixels); keep per-action results faithful. da-fr's UNDO + game-over prompts are copied in gh_da-fr_WRITEUP.md §4 verbatim — reuse them.
4. Solved-level rule memory (pinned rules from cleared levels; +2.8 for rellik13) — the ONLY text addition that paid across all write-ups. Everything else that adds recurring text (advice, per-action summaries, ledgers, archetype playbooks) is measured NEGATIVE. Our governor/schema-lite addenda should be re-audited against this evidence: keep them additive-only-at-reset, never recurring.
5. Persistent model-defined Python functions across calls (da-fr) — free token savings, no risk; pairs with our sandbox.
6. Priority/UCB scheduling over one pool (da-fr's P=(A+B)·C formulas or rellik13's UCB+2400s slices; also gives us priority_pace_reference.json + priority_scheduler.py as copyable code from da-fr's tree).
7. If we get memory right and stall: port TWIN-style twin-gating or arc-skill-style prediction-graded actions into the Duck's python tool (they're harness changes, model-agnostic); use theredbluepill's 249 community games for offline practice distribution; write "manuals" from our solved runs (agno's seeding evidence) but only as pinned solved-level memory, not recurring advice.
8. Variance protocol (everyone's top lesson): single runs resolve nothing (rellik13: same notebook 5.02/5.50/6.42; 37 practice → 9.98 LB; discussion 743723: identical builds 2.60 vs 1.45). Compare at equal generated tokens per game (their metric), read transcripts, keep our full-length practice runs, gate on public score only (1zuki's rules: offline means diagnose, never promote).

Copyable code right now: da-fr/arc-agi-3-solution → ARC3-Inference/inference/agent/{priority_scheduler.py,noop_repeat_guard.py,runtime_state.py,vision_context.py}, inference/utils/{animation.py,frame_diff.py,retained_functions.py,segmentation.py}, serving/patches/*, ARC3-Inference/CONFIGURATION.md (43KB, saved). LohitSiriki/... → cells/44_patch_cell_v16_7_m72_one_pool_ucb_scheduler.py, cells/42_...solved_level_memory.py, cells/33_...context_guard.py, cells/29_...mid_action_frames.py, cells/19_...action7_fix.py + analysis/{sawtooth,tps_by_running}.py for cache/trim telemetry.

Files saved in research/milestone2/

Readmes/writeups: gh_da-fr_arc-agi-3-solution_README.md, gh_da-fr_WRITEUP.md, gh_da-fr_CONFIGURATION.md, gh_LohitSiriki_arc-agi-3-milestone2-solution_README.md, gh_LohitSiriki_WRITEUP.md, gh_LohitSiriki_analysis_README.md, gh_LohitSiriki_tps_by_running.py, gh_Tufalabs_duck-harness_README.md, gh_Tufalabs_ARC3-Inference_README.md, gh_arcprize_arc-agi-3-benchmarking_README.md, gh_arcprize_Community-Leaderboard_README.md, gh_alexisfox7_PRO-LONG_README.md, gh_alexisfox7_PROLONG_arcagi3_README.md, gh_pbshgthm_arc-skill_README.md, gh_DriesSmit_ARC3-solution_README.md, gh_Alexyskoutnev_TWIN-ARC-AGI-3_README.md, gh_U4AR_qwen38-arc3-rl_README.md, gh_U4AR_RESULTS_REPORT.md, gh_samrishtt_arc-agi-3-kaggle-competition_README.md, gh_agno-agi_arc-agi-arcade_README.md, gh_theredbluepill_arc-interactive_README.md, gh_dp-web4_ARC-SAGE_README.md, gh_1zuki_milestone2_crawl.md, gh_tsitsino_MILESTONE_2.md. Structured: all_repos.json (115 repos), searchrepo_*.json (13 queries), searchcode_milestone2.json, orgs_arcprize.json, commits_arcagi3agents.json, tree_*.json (12 trees incl. da-fr + LohitSiriki + Tufalabs 3,607 files), user_*.json (jcottaar/dfranzen/Tufalabs/NVARC3 etc).