arXiv AI

A Contradiction Judge Counted Reported Teammate Errors as Hallucinations: An Assertion-Aware Re-Score of One Multi-Agent LLM Harness

The paper revises earlier claims about hallucination rates in multi‑agent LLM systems, noting that a judge that flags contradictions can mistakenly label correct reports of teammate errors as hallucinations. Using a pre‑registered analysis plan, the authors re‑score outputs from a harness with Llama and Claude agents, finding that the broadcast effect on hallucination rates is negligible and that most hallucinations stem from teammates’ wrong values. The study also reports that the effect varies across judge‑prompt combinations and that the data are limited to a single harness and 13 hand‑written scenarios.

arXiv AI
Oct 26

Validating the Comparison, Not Just the Scorer: Why Overall Accuracy Can Mislead in Interactive LLM Evaluation

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
arXiv AI
Sep 2

Commit-first LLM judging inherits the judge's own errors

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
arXiv AI
Oct 2

The First Token Is Not the Verdict: Hidden Costs of Reading LLM Judges Without Generating

The paper demonstrates that reading a large language model (LLM) judge’s verdict from the logits of its first generated token—an approach used in constrained decoding and likelihood‑scoring evaluation—introduces a significant distortion in position bias. Because judges do not always start with a verdict token (12–49% of cases for Qwen3 judges and <3% for Llama‑3.1‑8B and Phi‑3.5‑mini), this readout often returns the first response rather than a true judgment, inflating position bias by up to 42 points while barely affecting judge accuracy. The authors recommend reporting the frequency with which a judge leads with a verdict token to provide a more accurate assessment of position bias.

By Gnaneswar Villuri, Hashmath Shaik, Alex Doboli
arXiv Computation and Language
Oct 23

From Clinical Judgment to Checkable Rules: Clinician-Led Guardrails for Mental-Health Chatbots

The paper presents a clinician‑led safety guardrail for mental‑health chatbots that replaces typical prompt‑engineering or Constitutional AI methods. Two clinicians design, finalize, and adjudicate a rule base of four safety checks, which is formally verified for internal consistency and applied to LLM outputs. The verification layer’s verdicts are recomputable from a named rule and sixteen facts, and it consistently avoids false crisis‑resource demands that a matched LLM judge frequently raises.

By Chenhan Lyu, Kit Wislocki, Sarah Elizabeth Stoeckl, Maitreyi Sinha, Amir Rahmani
arXiv Computation and Language
Oct 5

Silent Dissent: LLM Agents That Yield to the Majority Still Represent Their Original Premise

The paper investigates whether large language model (LLM) agents that yield to a unanimous majority in multi‑agent debate truly change their underlying premise or merely adjust their statement. Using two‑hop factual questions with an unstated intermediate entity, the authors show that agents that concede still encode the original bridge in their internal representations, as revealed by a Jacobian‑lens analysis, even when the majority answer is wrong. Experiments across several open‑weight models demonstrate that hiding or removing the agent’s earlier answer increases conformity and that the original premise can be recovered from the question alone.

By Ziang Ni, Peng Zou
arXiv AI
Oct 12

Agreement Without Alignment: Label-Free Rubric Clarification Increases LLM-Judge Consensus but Not Agreement with Humans

The study evaluates a label‑free rubric‑clarification procedure that rewrites only the most contested sub‑criteria of a rubric, aiming to reduce disagreement among large‑language‑model (LLM) judges. Across six rubric conditions and three datasets (SummEval, HealthBench, USR Topical‑Chat), the method increased inter‑judge agreement on SummEval by 0.069 Krippendorff’s α, but did not improve agreement with human expert ratings. In other datasets, neither inter‑judge nor human agreement met the preregistered effect criteria, and human agreement was actually lower after editing contested clauses compared to random edits.

By Sangdo Han, Moonsoo Kim
arXiv AI
Oct 7

Evaluate the Stack, Not the Layer: Do Deterministic and LLM Gates for Agent Actions Fail Independently?

The study evaluates how runtime gates for agent tool calls stack, assuming their errors multiply. Using 1,119 labeled agent actions from three corpora, the authors compare a deterministic rule layer with four LLM judges, measuring the effective number of multiplication-equivalent layers (n_mult). Results show that two judges compose to about 1.2–1.4 layers, while a rule layer plus one judge composes to roughly 1.86–2.09 layers, with variations depending on miss definitions and corpora.

By Chenglin Yang