Eve39570/agent-plasticity-hex-go

Agent Plasticity on 6x6 Hex and 5x5 Go: Claude Sonnet 5.5, Haiku 5.5 and GLM 5.3 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, or 5×5 Go against GNU Go. 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

Every table has a game column (hex or go).

Config	Rows (Hex + Go)	One row is
runs	18 + 10	a lineage: settings, cost, plasticity fit
checkpoints	183 + 110	a checkpoint: wins per split, learning cost, reflection outcome
games	4,954 + 3,520	a game: opening, moves, per-move diagnostics, token use, full transcript with thinking where recorded
reflections	171 + 105	a reflection attempt: full transcript with thinking where recorded, validator result
inventories	814 + 578	a file the agent saved, at a checkpoint (code, notes, tracking JSON)
calls	62,236 + 48,269	a model call: tokens (input, cached, cache write, output) and charged cost
openings	366 + 152	a starting position from the legacy, contested, pooled or fresh bank (fresh: held-out openings no original run played; on Hex checkpoint says which checkpoint plays it)
thinking_replay	830 + 547	a saved reflection turn sent again with thinking summaries on: recorded and replayed calls, summary
ablation_games	144 + 240	a held-out ID game from a final checkpoint with the inventory kept, removed or forced, or on 40 fresh Go openings (arm = fresh)
trace_viewer	660	every reflection (with replayed summaries, labelled, where none was recorded) and ablation game, as a transcript in Pi's session format, for the trace viewer
traces_<run_id>	8,474 in all	every game of one run, in the same format (one config per run: the Hub's conversion timed out on one config of 3,918 files)

In trace_viewer and traces_* the harness column says pi. The traces were not run in Pi. The harness here is this project's own; its transcripts are written in Pi's session format because the Hub's trace viewer reads that format, so the Hub names the harness after the format.

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

run_id	Game	Model	Reflection effort	Checkpoints	Games per checkpoint (train / ID / OOD)	Openings
smoke	Hex	Sonnet 5.5	medium	0–1	4 / 4 / 2	legacy (10/11-ply)
lineage_a	Hex	Sonnet 5.5	medium	0 (+2 games of 1)	8 / 12 / 4	contested (stopped: ID bank harder than train)
lineage_a2, sonnet_medium_rep2	Hex	Sonnet 5.5	medium	0–10	8 / 12 / 4	pooled
sonnet_reflect_high, ..._rep2	Hex	Sonnet 5.5	high	0–10	8 / 12 / 4	pooled
haiku_r20	Hex	Haiku 5.5	medium	0–20	20 / 24 / 8	pooled
hex_haiku_medium_rep2	Hex	Haiku 5.5	medium	0–10	8 / 12 / 4	pooled
hex_haiku_reflect_high, ..._rep2	Hex	Haiku 5.5	high	0–10	8 / 12 / 4	pooled
hex_glm_reflect_high, ..._rep2	Hex	GLM 5.3	high	0–10	8 / 12 / 4	pooled
sonnet_nofeedback_high, ..._rep2	Hex	Sonnet 5.5	high, no game records	0–7	8 / 12 / 4	pooled
hex_sonnet_high_fresh, ..._rep2	Hex	Sonnet 5.5	high	0–10	8 / 12 / 4	fresh: held-out ID new at every checkpoint, train and OOD pooled
hex_sonnet_medium_fresh, ..._rep2	Hex	Sonnet 5.5	medium	0–10	8 / 12 / 4	as above
go_sonnet_medium, ..._rep2	Go	Sonnet 5.5	medium	0–10	8 / 20 / 4	pooled, training draw changes per checkpoint
go_sonnet_reflect_high, ..._rep2	Go	Sonnet 5.5	high	0–10	8 / 20 / 4	as above
go_haiku_medium, ..._rep2	Go	Haiku 5.5	medium	0–10	8 / 20 / 4	as above
go_haiku_reflect_high, ..._rep2	Go	Haiku 5.5	high	0–10	8 / 20 / 4	as above
go_glm_reflect_high, ..._rep2	Go	GLM 5.3	high	0–10	8 / 20 / 4	as above

GLM 5.3 (z-ai/glm-5.3) ran through OpenRouter's chat-completions API with 2 games at a time per lineage (Claude: 4); all four GLM lineages were stopped by the API key's spending limit and resumed, replaying the unfinished checkpoint.

Main result

Held-out ID, early → late (checkpoints 0–3 → 4–10 for Hex Sonnet, 0–2 → 3–10 otherwise):

• Hex Sonnet 5.5: at reflection effort medium, 59 % → 65 % (p = 0.32); the reflector used 5–6 of its 50 calls and wrote at most one version of a tool. At effort high, 73 % → 83 % (p = 0.04), revising its code and running it in most games. Over checkpoints 4–10 high ends at 83 % against 65 % for medium (z = 3.75, p = 0.0002) from the same start. Held-out openings are fixed and the opponents deterministic, so later checkpoints largely replay the same games; per opening, high effort is better on 10 of 12 and worse on 1 (sign test p = 0.012, report 23). Re-run with held-out openings new at every checkpoint (*_fresh, report 24), so no held-out game repeats: high effort 72 % → 88 %, medium 57 % → 56 %, and over checkpoints 4–10 high is better on 46 of 84 openings and worse on 6 (sign test p = 1e-08).
• Go Sonnet 5.5: effort makes no difference (late 61 % at medium, 60 % at high, p = 0.69). The two high-effort lineages split: one holds 85–90 %, the other falls to 15–20 % at checkpoints 6–9 after tool timeouts and a search cap it set itself.
• Haiku 5.5 stays flat on both games at both efforts (e.g. haiku_r20 near 31 % over 21 checkpoints) despite busy reflection.
• GLM 5.3 lost most of its checkpoint-0 games to the time and call limits and fixed that in one reflection. That jump alone passes the flat-curve test in three of four lineages and gives its Go lineages P_sat of 48,000–60,000, the highest here, while its held-out ID after checkpoint 1 is flat or falling (late: Hex 24 %, Go 46 %). In hex_glm_reflect_high one reflection wrote a one-line bug into its search tool, and held-out ID fell from 75 % to 0 %.
• Go, fresh openings: each Go lineage's final inventory played 40 openings no run had played (ablation_games, arm = "fresh"). Sonnet and Haiku both beat GLM opening by opening (24 better / 1 worse, 18 / 3); Sonnet against Haiku is not established (13 / 8, p = 0.38) because the two Sonnet lineages score 95 % and 50 % (report 25).

Two limitation tests on Hex Sonnet at reflection effort high (reports 17 and 18). Removing the final inventory drops held-out ID from 92 % to 50–54 % (ablation_games). Reflection without game records (runs.reflection_feedback = "none") built no tools at all; on the same held-out openings the feedback lineages won 38 games the no-feedback lineages lost, and lost 5 they won, from checkpoint 3 on (per opening: 8 better, 1 worse, p = 0.039).

Details: reports/TECHNICAL_REPORT.md and reports 12, 15, 21–25 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: the feedback reflection received, per actor move. Hex: missed immediate wins, preventable immediate losses, connection distance before and after. Go: stones captured, largest capture available, liberties after, stones in atari, area margin before and after, and the labels self_atari and pass_while_behind.
• games.opponent_level: Hex, alpha-beta nodes per move; Go, GNU Go 3.8 level (5–9 train and ID, 10 OOD).
• games.actor: red/blue (Hex) or black/white (Go).
• games.end_reason: finished, budget (32 model calls used) or timeout (a second tool call over 240 s); the last two count as losses.
• games.checkpoint_complete: false for games of a checkpoint the run never completed (two lineage_a games).
• openings.difficulty: Hex, engine-scored; Go, GNU Go level 10's mean final margin against levels 5, 7 and 9. Not comparable across games.
• inventories: one row per file per checkpoint, so artifact changes can be diffed.
• calls.cost_usd: OpenRouter's reported charge for that call.
• runs.reflection_feedback: full, or none when reflection got only a note in place of the game records (limitation test 5).
• games.messages[].thinking, reflections.messages[].thinking: the model's thinking for that turn, as the provider returned it: a short summary for Claude, the reasoning text for GLM 5.3 (see Limits). Recorded from the sonnet_nofeedback_high lineages on, so in every plan B and plan C run and in later checkpoints of the extended plan A runs; earlier turns did not save it.
• thinking_replay: recorded and replayed are the original and the new turn's text and tool calls; same_call is true when the replayed turn made exactly the recorded calls; thinking is the new turn's summary. A new sample on the same input, not the original reasoning. Every model run with reflection logs is replayed.

How it differs from the paper

• 6×6 Hex and 5×5 Go only (no chess or NetHack); three models; 11 checkpoints instead of 20 (21 for haiku_r20); 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.
• Go: 2-ply (Black) / 3-ply (White) openings, kept if always playing the first legal action loses and GNU Go level 10 wins; from the empty board GNU Go as Black wins every game. Scored by GNU Go's final_score (Chinese rules, dead stones removed), komi 9.5.
• Prompts are paraphrased from the paper's Appendix I (same rules, budgets and schemas).
• Sandbox: macOS sandbox-exec instead of a container.

Limits

• What "thinking" is depends on the provider. Closed frontier models do not return their raw reasoning. Through the API, Claude 4 and later return a summary of it, which Anthropic describes as keeping the benefits of extended thinking "while preventing misuse"; protecting the reasoning against distillation is another commonly cited reason for hiding it. Sonnet 5.5 through OpenRouter returned short summaries (median about 220 characters per turn) and Haiku 5.5 returned them only when asked. GLM 5.3 is open-weights, and OpenRouter returns the reasoning text it produced (median about 1,800 characters per reflection turn, up to 45,000). For studying how an agent learns, the GLM traces show far more of the reasoning behind each edit. The two kinds are not comparable: a Claude summary is a condensed account, the GLM text is the reasoning as generated.
• Thinking only in the newer runs. The models thought on every call, but until the no-feedback lineages the harness did not save the thinking blocks. Earlier transcripts contain actions and visible text only; thinking_replay gives summaries on the same reflection inputs (new samples, not the original reasoning).
• Small samples, replayed. 12 (Hex) or 20 (Go) held-out ID games per checkpoint, the same openings at every checkpoint except in the *_fresh runs, so later checkpoints often replay earlier games (report 23); differences under about 20 points between halves of a run are not detectable.
• Two lineages per condition, except haiku_r20 and hex_haiku_medium_rep2 (one each).
• Not every dollar is here. About $0.07 from a first start of lineage_a2 (restarted to add connection retries), $0.002 from an endpoint probe, $2.21 of Go cost probes and $0.33 of a go_sonnet_reflect_high checkpoint redone after a crash are not included, nor $2.27 from a doubled start of the *_fresh runs, which were restarted from checkpoint 0. Nor are the calls of checkpoints replayed after an interruption (ledger_discarded.jsonl in the source repository: $0.29–2.31 per affected lineage, $6.85 in all); the replay's games replace them.

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 and Claude Haiku 5.5 (Anthropic) and GLM 5.3 (Z.ai), accessed through OpenRouter; check the 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}
}