arXiv AI August 21, 2026 By Haiyue Zhang

Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay

arXiv:2608. 19760v1 Announce Type: cross Abstract: Audited against causal ground truth from executed replay in a single-agent tool environment (ALFWorld), none of the step-level credit signals used to train LLM agents -- LLM-judge scores, outcome-conditioned logprob ratios, or the policy's own confidence -- identifies which steps causally matter better than chance.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

arXiv Machine Learning

Jun 16

Re-feeding Is Not Replaying: Measuring Replay Noise in Counterfactual Token-Credit Estimation

arXiv:2606. 15621v1 Announce Type: new Abstract: Per-token counterfactual credit estimation asks which token in a language-model rollout caused the final answer to be right or wrong: cut the transcript at a pivot, substitute an alternative token, replay continuations, and compare outcomes.

By Nils Matteson

efficiency

arXiv AI

Jun 9

Causal Agent Replay: Counterfactual Attribution for LLM-Agent Failures

arXiv:2606. 08275v1 Announce Type: cross Abstract: When an LLM agent fails -- issues a refund it should not have, calls the wrong tool, leaks data -- existing tooling answers what happened (observability) or whether it passed (evaluation), but not which step caused the failure.

By Jaineet Shah

llms agents benchmarks

arXiv AI

Aug 11

Privileged Likelihood Is Not Automatically Value: Three Checks for Token Credit in On-Policy Self-Distillation

arXiv:2608. 09263v1 Announce Type: new Abstract: Outcome verifiers score completed reasoning traces but do not assign credit to intermediate tokens.

By Xuan-Phi Nguyen, Shrey Pandit, Yiran Zhao, Anurag Koul, Zeyu Liu, Shafiq Joty

efficiency

arXiv AI

Jul 15

Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs

arXiv:2607. 12985v1 Announce Type: new Abstract: Aligned language models routinely misreport under non-evidential incentive pressure: they agree with a confident user or overstate certainty even when their internal belief is unchanged.

By Sen Yang, Yuen-Hei Yeung

llms fine-tuning benchmarks

arXiv AI

Jul 14

Auditing the Risk Claims of Distributional Reinforcement Learning

arXiv:2607. 11607v1 Announce Type: new Abstract: Distributional reinforcement learning agents learn full return distributions that are increasingly read at face value: for interpretability, risk-sensitive control, and safety monitoring.

By Hari Prasad

agents reinforcement-learning benchmarks safety

arXiv AI

Aug 20

SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents

SkillGate is a method that trains agents to select the correct skill from a large slate during an episode by separating credit signals for skill selection and execution. It addresses the problem of selector credit starvation, where traditional outcome-rewarded RL fails to give sufficient credit to the skill-naming tokens, especially in long-horizon tasks. Experiments on five benchmarks show that SkillGate improves a 9B policy’s success rate from 40.8% to 53.2%, reduces exposure to misleading candidates, and requires fewer skill reads.

By Qingyao Li, Wenxiang Jiao, Shuai Shao, Kangning Zhang, Yuan Lu, Yi Guo, Weiwen Liu, Weinan Zhang, Yong Yu

agents reinforcement-learning benchmarks

One email a morning, machine-written

Related stories

Re-feeding Is Not Replaying: Measuring Replay Noise in Counterfactual Token-Credit Estimation

Causal Agent Replay: Counterfactual Attribution for LLM-Agent Failures

Privileged Likelihood Is Not Automatically Value: Three Checks for Token Credit in On-Policy Self-Distillation

Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs

Auditing the Risk Claims of Distributional Reinforcement Learning

SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents