The Flow has not summarised this story yet — read it at arXiv Machine Learning.
arXiv:2608. 16003v1 Announce Type: new Abstract: Automated checking pipelines increasingly place one language model as the checker and another (or the same one) as the fixer.
The study investigates whether adding specialist language‑model contexts can enhance software assurance when deterministic tool evidence is already available. Five experiments were conducted using Qwen2.5‑Coder‑7B‑Instruct across Java warning controls, OWASP Benchmark, Defects4J test generation, SV‑COMP assertion verdicts, and misleading evidence packages. Results show that increasing context count does not improve unsafe clearance, that verifier feedback can raise correct verdicts but also incorrect ones, and that models reject most counterexamples while scope checks classify packages accurately, supporting deterministic evidence admission but highlighting limits in verdict aggregation and feedback interpretation.
The paper investigates how large language model (LLM) agents that triage security alerts can be misled when the detection tools they rely on provide conflicting or incorrect evidence. By corrupting up to three of four heterogeneous detectors on the CIC-IDS2017 dataset, the authors show that naive LLM investigators become highly fragile, with accuracy dropping to as low as 13% and abstaining on up to 63% of flows. They propose Conflict‑Aware Evidence Verification (CAV), which combines an equal‑weight vote, an evidence‑conflict gate, and a nearest‑neighbour trust check; CAV consistently outperforms naive investigators across all corruption scenarios, improving accuracy to 84–88% when three detectors are corrupted while only querying the LLM on 0.4% of clean traffic.
The paper introduces ClaimReceipt, a specification and verifier that checks whether a claim in an agent evaluation can be recomputed from retained evidence (sufficiency) and whether the evidence covers the entire experiment set (coverage). Using the CR‑2 verifier on 1,392 historical records, the authors demonstrate accurate reproduction of audit verdicts, non‑redundant field groups, and zero false positives on semantic faults. In a prospective CR‑3 run, the system correctly flags missing receipts and preserves coverage when private evidence is withheld, while adding minimal overhead to inference time and transaction size.
ClaimScope is a 72‑task benchmark that audits mathematical claims by requiring a verdict, scope, counterexample, and repair, covering areas such as optimizer‑state transport and exact $2 imes2$ linear algebra. The study evaluates Qwen3‑8B using a witness checker that only examines the model’s submitted state, showing that verifier feedback improves counterexample acceptance and safe completions compared to generic feedback. The results demonstrate that separating claim classification from executable evidence can yield gains in witness‑structure correction without necessarily improving broader reasoning.
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.