penfever/ablation_exploration_in_rl

Reinforcement Learning Improves Agentic Software Engineering

An ablation study of reinforcement-learning (RL) fine-tuning for agentic software-engineering (SWE) models. Starting from an 8B SFT model, we fine-tune with RL across ~20 configurations — varying the objective, loss normalization, sampling, and training dataset — and evaluate each on agentic SWE benchmarks.

Result

RL reliably and substantially improves agentic SWE performance, and the improvement is general — it does not depend on the particular RL recipe or training dataset.

On SWE-bench-Verified (random-100, N = 300 = 100 tasks × 3 attempts):

	SWE-bench-Verified (pass rate)
SFT base	0.24
RL variants (20)	0.30 – 0.37, mean 0.332

• Every RL variant beats the base; the mean gain is ≈ +9 points.
• The 20 variants are statistically indistinguishable from one another — the full spread is about one standard error, so no single recipe is a clear winner.
• The gain holds across training datasets: RL on a freelancer-task dataset scores 0.337, the same range as RL on the pymethods2test dataset. The effect is RL itself, not a specific data mix.
• RL from a weaker SFT checkpoint forms a distinct lower band (0.26–0.28), consistent with the improvement being relative to the starting policy.

What the model learns

Post-RL agents close the loop: diagnose the failure, edit the real source, re-run the tests, iterate, and mark the task complete only after the tests pass. The SFT base characteristically stalls in inspection — it explores the repository but rarely commits an edit or verifies a fix. All RL variants also lengthen their reasoning and raise tool use. These changes are general, not specific to any one variant. Trace-grounded examples are in HERO_LEARNED_BEHAVIORS.md.

Documents

• COMPARISON_SWEBENCH_PINNED.md — quantitative results and statistics.
• id_eval_grid.md — per-variant scores across all benchmarks.
• HERO_LEARNED_BEHAVIORS.md — qualitative behavioral analysis, trace-grounded.
• TRACES.md — HuggingFace trace datasets (agent trajectory + verifier report) for every model, the reproducibility backbone.
• marin_issue_draft.md — the evaluation-harness bug described in the Methods note.

Methods

Models are 8B; the SFT base is GLM-4_7-swesmith. Agent evaluations use the terminus-2 harness at 32k context, three attempts per task, on SWE-bench-Verified (random-100), Terminal-Bench 2, and an internal dev set. Scores are pass rates over all attempts.

Evaluation-harness note. The terminus-2 agent periodically summarizes its transcript (an LLM call) to stay within context. When that call times out, the harness aborts the entire trial as a failure (SummarizationTimeoutError), depressing scores and doing so unevenly across models. All results reported here are from evaluations free of this failure (zero summarization timeouts); earlier affected measurements are excluded. See marin_issue_draft.md.

Next

Extend the clean evaluation to the a1/a3 model families and to the Terminal-Bench-2 and dev-set base measurements.