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.
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.
arXiv:2606. 18479v1 Announce Type: new Abstract: Reject inference methods are widely used to mitigate survival bias in credit scoring, yet their effectiveness remains poorly understood.
The paper evaluates an agentic harness that combines a 27B language model, financial calculations, narrative templates, and validation checks to produce corporate and investment banking presentation decks. Using a panel of five judges, the system consistently scores higher than a baseline model that generates directly from a short prompt, with scores ranging from 20.4 to 33.6 out of 95. However, judge variability and changes in grading criteria make it challenging to discern small improvements in deck quality.
The paper investigates when historical action credit for tool‑using agents must be updated after policy changes. It introduces pairwise branch sensitivity to measure how policy updates affect action‑distinguishing branches, and proposes a first‑order anchored credit‑transport estimator along with a Decision‑Sufficient Credit Gate (DSC‑Gate) to decide whether to reuse, transport, or resample credit. Experiments show that DSC‑Gate reduces new tool steps by 39.4% with negligible impact on regret, demonstrating that many policy updates can avoid costly recomputation of action credit.
The paper introduces a method for verifying whether two gameplay replays in Counter‑Strike 2 belong to the same player by extracting a behavioral fingerprint that captures crosshair control, movement‑stop‑fire coordination, economy, combat engagement, and temporal rhythm. Using 1,330 demos and 13,300 observations, the authors train a pairwise model that achieves an ROC AUC of 0.931 and 0.722 recall at 95% precision, with low‑level mechanical habits providing the strongest identity signals. Aggregating multiple demos further improves performance, raising AUC to 0.986 when ten historical demos are considered.
arXiv:2609.36178v1 Announce Type: cross Abstract: Group Relative Policy Optimization (GRPO) has become a promising approach for training large language model agents. However, its uniform assignment o...
arXiv:2606. 01513v1 Announce Type: cross Abstract: High-stakes enterprise document generation, including financial dispute narratives, compliance notices, and audit summaries, demands schema correctness, policy compliance, and low-latency operation at scale.
The paper demonstrates that open‑ended Theory‑of‑Mind trackers can produce valid beliefs that are absent from finite reference sets, and that treating unmatched outputs as false can reverse model‑selection rankings. By recoding references for 259 beliefs, the authors show a dramatic drop in weighted prevalence and a reversal of strictly proper Brier risk, with similar distortions observed in a 301‑question NQ‑open DPR‑BERT pipeline. The study further reveals that 90‑96% of audited unmatched beliefs are literally true, and introduces a TriSource‑Restore method that anchors reference labels to a probability‑sampled human pilot to restore calibration and ranking integrity.
CITECHOICE is a causal audit that examines how the presentation of documents in an agentic search engine redistributes citation credit. Using 129 everyday‑query transcripts, the study compares structured versus prose renderings of the same source while keeping all other transcript elements fixed. The results show that structured rendering increases the target’s citation count by about half a citation per answer without adding total citations or diminishing competitors’ credit, while also revealing that rank position has a larger effect on citation rates than presentation order alone.
arXiv:2609.01198v1 Announce Type: new Abstract: Repeated banking interactions require assistants to maintain complete, current, and traceable customer records as life changes emerge incidentally in r...
The paper proposes that self‑evolving large‑language‑model agents can only be safely reviewed if their changes are confined to the runtime harness—instruction text, tool‑call logic, and primitive composition—while the model weights remain fixed. A dual‑loop engine with an admission gate records a hash‑chained diff before deployment, and simulation experiments show that this gate admits only safe changes, preventing error increases and maintaining false‑positive rates. When the gate is removed, many harmful changes slip through and missed flags rise, illustrating the gate’s protective role.