Eve39570/agent-plasticity-hex

Agent Plasticity on 6x6 Hex: Claude Sonnet 5.5 traces

Agent traces from a scaled-down, independent reproduction of Agent Plasticity: Measuring Self-Improvement Through Experience (Singh et al., 2026). The paper's authors did not release code or data; this harness was rebuilt from the paper's equations and appendix.

A frozen model plays 6×6 Hex against an alpha-beta engine. Between checkpoints the same model reflects on its training games and edits a persistent inventory (Python tools, notes, tracking files) that the next checkpoint inherits. Held-out games are never shown to reflection.

What is in it

Config	Rows	One row is
runs	7	a lineage: settings, cost, plasticity fit
checkpoints	56	a checkpoint: wins per split, learning cost, reflection outcome
games	1,906	a game: opening, moves, per-move diagnostics, token use, full transcript
reflections	53	a reflection attempt: full transcript, validator result
inventories	227	a file the agent saved, at a checkpoint (code, notes, tracking JSON)
calls	23,919	a model call: tokens (input, cached, cache write, output) and charged cost
openings	234	a starting position from the legacy, contested or pooled bank
trace_viewer	1,959	a game or reflection transcript in the Hub Session Trace format, for the trace viewer

Runs (via OpenRouter; games always at effort medium):

run_id	Model	Reflection effort	Checkpoints	Games per checkpoint (train / ID / OOD)	Openings
smoke	Sonnet 5.5	medium	0–1	4 / 4 / 2	legacy (10/11-ply)
lineage_a	Sonnet 5.5	medium	0 (+2 games of 1)	8 / 12 / 4	contested (stopped: ID bank harder than train)
lineage_a2	Sonnet 5.5	medium	0–7	8 / 12 / 4	pooled
sonnet_medium_rep2	Sonnet 5.5	medium	0–7	8 / 12 / 4	pooled
sonnet_reflect_high	Sonnet 5.5	high	0–7	8 / 12 / 4	pooled
sonnet_reflect_high_rep2	Sonnet 5.5	high	0–7	8 / 12 / 4	pooled
haiku_r20	Haiku 5.5	medium	0–20	20 / 24 / 8	pooled

Main result

With reflection at effort medium, Sonnet 5.5 shows no plasticity in two lineages (held-out ID 59 % → 62 % between checkpoints 0–3 and 4–7): the reflector used 5–6 of its 50 calls, rarely read its game logs, and wrote at most one version of a tool. With reflection at effort high, two lineages rise from 73 % to 83 %, revising their code and running it in most games. From the same start (15/24 each at checkpoint 0), high ends at 83 % against 62 % for medium over checkpoints 4–7 (z = 3.25, p = 0.001). Haiku 5.5 stays near 31 % over 21 checkpoints despite busy reflection. Details: reports/12_comparison.md in the source repository.

Fields worth knowing

• games.messages: the actor's conversation. Assistant turns hold visible text and tool calls (bash, python, submit_action); tool turns hold outputs.
• games.diagnostics: per actor move, missed immediate wins, preventable immediate losses, and connection distance before and after (the feedback reflection received).
• games.checkpoint_complete: false for the two lineage_a games played after the checkpoint it stopped at.
• inventories: one row per file per checkpoint, so artifact changes can be diffed.
• calls.cost_usd: OpenRouter's reported charge for that call.

How it differs from the paper

• 6×6 Hex only; one model; 8 checkpoints instead of 20; far fewer games per checkpoint.
• Opponent: a pure-Python alpha-beta engine at 3k–100k nodes per move (train and ID) and 1M (OOD), scaled down from the paper's 10k–1M and 10M.
• Openings: 4/5-ply instead of 10/11-ply (except smoke), filtered so a weak player loses and a 30k-node engine wins, and for lineage_a2 dealt to the splits by engine-scored difficulty.
• Prompts are paraphrased from the paper's Appendix I (same rules, budgets and schemas).
• Sandbox: macOS sandbox-exec instead of a container.

Limits

• No hidden reasoning. The API returns thinking blocks empty by default, so transcripts contain actions and visible text only.
• Small samples. 12 held-out ID games per checkpoint; differences under about 20 points between halves of the run are not detectable.
• Two lineages per reflection-effort condition for Sonnet, one for Haiku.
• Not every dollar is here. About $0.07 from a first start of lineage_a2 (restarted to add connection retries) and $0.002 from an endpoint probe are not included.

Privacy and provenance

Local file paths are replaced by <repo> and /Users/<user>. The export script stops if any string looks like an API key, bearer token or email address. No personal data is in the games; the model only saw game states and its own files.

Model outputs are from Claude Sonnet 5.5 (Anthropic) accessed through OpenRouter; check both providers' terms before reuse.

Citation

Please cite the original paper:

@article{singh2026agentplasticity,
  title  = {Agent Plasticity: Measuring Self-Improvement Through Experience},
  author = {Singh, Harman and Bakhtin, Anton and Shao, Rulin and Synnaeve, Gabriel and
            Kulikov, Ilia and Fergus, Rob and Arora, Sanjeev and Keutzer, Kurt and
            Weston, Jason and Mahajan, Anuj and Goyal, Anirudh},
  journal = {arXiv preprint arXiv:2610.08902},
  year   = {2026}
}