aiqbench analysis
Counterfactual Rollout Replay (CRR) addresses a critical gap in training software engineering agents by leveraging forkable environments to generate step-level feedback without human supervision. This technique significantly boosts performance on complex coding benchmarks like SWE-bench Verified, offering a more efficient alternative to outcome-only reinforcement learning methods. The approach demonstrates that high-quality process rewards can be derived purely from environmental execution rather than external labeling or learned models.
Counterfactual Rollout Replay: Forkable Environments as Free Process Rewards for Software Engineering Agents
arXiv:2609.33875v1 Announce Type: cross Abstract: Outcome-only reinforcement learning gives software engineering (SWE) agents a terminal success signal but little direct guidance about intermediate decisions. We introduce Counterfactual Rollout Replay (CRR), a training-time procedure that uses forkable executable environments to obtain step-level return contrasts. CRR selects a small set of decision points, restores each state, samples an alternative action, and rolls the branch forward under the policy. It retains the realised training trajectory and replaces the advantage at selected steps with the difference between its terminal return and the sampled counterfactual return. The method needs no human process labels or learned process reward model; free refers to those supervision costs, not replay compute. With a 14B policy, CRR improves pass@1 on SWE-bench Verified, SWE-bench Live, and SWE-rebench, and combines with process-reward and trajectory-search methods. On SWE-bench Verified, an equal-wall-clock comparison on the same hardware yields 41.7% versus 36.7% for extended outcome-only GRPO, a 5.0-point gain with fork overhead included. These results apply to environments with affordable, reliable state restoration; stochastic continuations and expensive or imperfect replay remain limitations.
Abstract:Outcome-only reinforcement learning gives software engineering (SWE) agents a terminal success signal but little direct guidance about intermediate decisions. We introduce Counterfactual Rollout Replay (CRR), a training-time procedure that uses forkable executable environments to obtain step-level return contrasts. CRR selects a small set of decision points, restores each state, samples an alternative action, and rolls the branch forward under the policy. It retains the realised training trajectory and replaces the advantage at selected steps with the difference between its terminal return and the sampled counterfactual return. The method needs no human process labels or learned process reward model; free refers to those supervision costs, not replay compute. With a 14B policy, CRR improves pass@1 on SWE-bench Verified, SWE-bench Live, and SWE-rebench, and combines with process-reward and trajectory-search methods. On SWE-bench Verified, an equal-wall-clock comparison on the same hardware yields 41.7% versus 36.7% for extended outcome-only GRPO, a 5.0-point gain with fork overhead included. These results apply to environments with affordable, reliable state restoration; stochastic continuations and expensive or imperfect replay remain limitations.
Accepted at NeurIPS 2026. Includes additional experiments and analysis
Software Engineering (cs.SE); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2609.33875 [cs.SE]
(or arXiv:2609.33875v1 [cs.SE] for this version)
https://doi.org/10.48550/arXiv.2609.33875 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 281 words
Related coverage
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · Sep 17, 2026
[AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment
Latent Space · Oct 16, 2026
Echoverse: Deep, evolving environments for computer-use agents
Microsoft Research · Jul 30, 2026
[AINews] Modal’s $1B ARR, $16B Series D makes it NYC’s newest AI Infra Decacorn, 3x usage and 3x valuation in 5 months, 15x in 1 yr
Latent Space · Oct 15, 2026