The paper investigates how sequential knowledge editing can degrade a language model’s ability to discern reliable evidence from unreliable evidence without affecting overall accuracy. Using a conservatively tuned LoRA on Qwen2.5‑7B‑Instruct, the authors show that after 1,000 edits the model’s arbitration score for untouched facts drops by 36%, leading to higher error rates on its most confident decisions, while MMLU accuracy remains unchanged. The study also finds that in some model‑method combinations, sequential edits can reduce MMLU to chance levels even though edit success and locality remain perfect.
By Atul Anand
The paper investigates the reliability of tool‑using agents, focusing on two failure modes: selecting the wrong tool and constructing incorrect arguments. It introduces a correct‑invocation rate metric to distinguish these errors and evaluates five open‑weight models on multi‑step tasks up to depth 8, finding that by depth 6 about 70% of a model’s clean‑context capability is lost due to earlier mistakes. The study reveals that exact‑match scoring against a fixed gold trajectory forces severity and recovery parameters to extreme values, and proposes a conditional‑on‑state scoring remedy that yields more realistic severity estimates.
By Afiya Noorain, Subhranshu Mohanty, Amritesh Banerjee, Abhijit Dasgupta
The study investigates why small language model agents tend to repeat a tool call that just failed. By recording the failed call and its error message in the transcript, the authors measure a negative corrective gain—agents are more likely to repeat the failed action, with a drop of about 1.03 nats per token. The problem is traced to the harness design rather than the model’s understanding of errors, and the authors show that replacing the verbatim call with a runtime-generated description of the failure can reduce this backfiring effect by 76%.
By Esmail Gumaan
arXiv:2610.03190v1 Announce Type: cross
Abstract: Accident, defect and outage investigations end with a decision that ordinary question answering never faces: whether the evidence gathered so far is...
By Tingzhu Bi, Ping Wang, Meng Ma
The paper introduces Dependency‑Tethered Memory (Tether), a memory system for tool‑using LLM agents that records the exact versions of premises each verdict depends on. By withdrawing any verdict whose support has failed anywhere in the store, Tether reduces unsafe publications to 0.3% and cuts verification calls by 37% compared to naive fact memory, while maintaining similar recall. The study demonstrates that capturing dependencies is crucial for safe memory use, as weaker writers or delayed relation arrivals increase unsafe publication rates.
By Yiqi Wang, Jinqian Ju, Jiaqi Zhang, Zhangkai Wu, Yiqun Duan, Zequn Sun, Taotao Cai, Mingkai Zheng
The paper investigates how much of an AI model’s expressed preferences are due to the model itself versus the instrument (prompt) used to elicit those preferences. By fixing the set of outcomes and models while varying five different prompting instruments across 15 welfare-related outcomes, the study finds that the ranking of outcomes is only moderately generalizable (coefficient 0.348) and that a single instrument’s preference provides little insight into another instrument’s results. The analysis shows that even after removing any single instrument, model, or a subset of outcomes, the overall preference estimate remains robust, yet the variability across instruments remains significant.
By Jason Hung
The paper investigates whether widely used evaluation frameworks for large language models (LLMs) implement a defense called commit‑first judging, which requires a judge to solve a task itself before accepting a candidate answer. Across 24 configurations in eight popular frameworks, none use the full commit‑first method; nine use a weaker variant that is ineffective. In controlled experiments, the weaker variant allowed systems to game the judge, while the full commit‑first approach eliminated this vulnerability but sometimes worsened evaluation when the judge’s own answer was incorrect.
By Idil Gozel
The paper argues that relying solely on overall accuracy or F1 scores to validate automatic scorers in LLM evaluation can conceal prompt‑specific errors that distort comparative results. By auditing six models on 180 invented‑entity questions across six prompts, the authors show that a scorer’s overall F1 of .916 masks substantial variation—up to 33.3%—in error rates depending on the prompt. In a second audit, differences in measured prompt effects dropped dramatically when multiple raters were used, highlighting the need to validate the exact comparison rather than the scorer alone.
By Ancil Crayton
The paper treats human interventions during LLM agent failures as labels that pinpoint the internal computation that went wrong, rather than merely training data. In a controlled environment, the authors analyze one failure type—over‑asking a clarifying question when the request is already clear—using rule‑based overseer labels validated by blind human judgment. They find that the decision to ask versus act can be causally localized across several models, that over‑asking exhibits distinct mechanistic signatures, and that steering or weight consolidation can mitigate the issue, though outcomes vary by model family.
By Cesare Zavattari, Alessandro Tommasi, Giuseppe Prencipe
arXiv:2609.08279v1 Announce Type: cross
Abstract: Agent memory systems must discard stored information when their history exceeds a fixed token budget. Existing budget-accuracy frontiers quantify the...
By Chen Shen
This paper introduces a deletion interface for a pretrained language model, measuring how effectively deleted records are removed from the model’s memory. By retrofitting a support‑vector memory gate into the global attention layers of a frozen Gemma 3, the authors show that deletions can be performed without altering weights and that the resulting state is close to a reference state that never stored the record. Experiments on 4B‑parameter models demonstrate low perplexity impact and strong evidence that deleted content is hard to recover, while larger or smaller models fail to achieve the same guarantees.
By Vishwajith Ramesh
arXiv:2608. 15286v1 Announce Type: cross Abstract: We introduce AgentRelBench, an environment-agnostic reliability instrument that computes ground-truth, severity-priced damage from database state diffs across repeated runs, with no LLM in the measurement path, demonstrated on EnterpriseOps-Gym.
By Shiven Khurdi