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 using a large language model (LLM) as the sole judge in self‑improving agent pipelines is problematic, as the judge can be biased or manipulated, leading to false confidence in system performance. The authors propose a new framework, PROCTOR, which replaces the oracle judge with a deterministic, teacher‑student loop that enforces guardrails such as sandboxing, role separation, and acceptance checks to prevent cheating and ensure reliable evaluation. Experiments across contract analysis, compliance review, and code quality demonstrate that PROCTOR mitigates eleven identified failure modes that previously allowed agents to achieve perfect scores while hiding significant capability gaps.
By Vansh Wahi
The paper investigates the reliability of LLM judges in the Arena-Hard benchmark, showing that aggregate scores can mask significant variability in individual verdicts. By running GPT‑4.1 across 500 prompts under eight different judge configurations and repeating each setup, the authors find that small changes—whether in temperature, criteria, or instruction framing—can shift win/tie/loss outcomes on up to 28% of prompts and alter aggregate scores by several points. The study concludes that stable aggregate scores do not guarantee repeatable individual judgments, and recommends that evaluations report both aggregate results and prompt‑level agreement.
By Louise Yorke
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
The study analyzes 373,019 judgments from LLM‑scored benchmarks, decomposing variance into system, item, judge, and interaction components via generalizability theory. It finds that with a single judge, generalizability converges to a ceiling determined by the system‑by‑judge variance, which is substantially lower in pairwise preference settings, allowing one judge to suffice. The research also reveals significant biases in presentation order and highlights that many published win‑rate claims fall below the measured floor of the benchmarks.
By Atul Anand
The paper investigates how language‑model judges can make version‑dependent errors when evaluating upgraded agents. Using 35 public coding‑agent submissions, two customer‑service agents, and over a thousand expert‑labeled trajectories, the authors show that fixed judges often reject task‑conditioned error invariance and can incorrectly approve failed patches, especially as agent capability increases. Paired audits of current outputs reduce interval width only marginally, and the study concludes that independent human patch review is still necessary.
By Jiapeng Li
arXiv:2607. 05904v1 Announce Type: new Abstract: Training a language model against its own reference-free judgments (the premise of self-rewarding, self-play, and LLM-as-a-judge pipelines) assumes a model's verdict on a shown answer tracks correctness.
By Chenyu Zhou
The paper investigates whether large language model (LLM) safety judges are truly policy‑invariant, meaning their verdicts should depend only on the agent’s behavior and not on how the evaluation policy is phrased. It introduces three testable principles—rubric‑semantics invariance, rubric‑threshold invariance, and ambiguity‑aware calibration—and applies them in a stress‑test protocol across four agent‑class judges using trajectories from ASSEBench and R‑Judge. The study finds that current judges often change verdicts in response to both meaningful normative shifts and meaningless structural rewrites, with up to 9.1% of verdicts flipping due to content‑preserving rewrites and 18‑43% of flips occurring on unambiguous cases, indicating that safety scores conflate agent actions with evaluator wording.
whyItMatters":"The findings reveal a hidden failure mode in safety evaluation that could mislead researchers about agent performance, underscoring the need for policy‑invariant judges and transparent reporting protocols."
By Shihao Weng, Yang Feng, Xiaofei Xie
The study compares four evaluation methods for AI agents in tau2-bench’s airline and retail domains: native scoring against a gold state, an LLM judge, the same judge with a policy document, and a minimal rule-based boundary evaluator. Across 2,160 episodes, the LLM judge agreed with native scoring only 56.2% of the time, while the boundary evaluator detected all 11 forbidden writes that the judges missed. The results suggest that the choice of evaluator significantly influences conclusions about agent correctness and policy compliance.
By Sungbin Im
arXiv:2608. 13564v1 Announce Type: new Abstract: Evaluating language-model agents at scale increasingly relies on a second language model as an automatic judge, because the gold signal, an executable environment reward, is expensive, slow, or unavailable at deployment time.
By Darragh Quinn, David Dylan, Roisin Healy, Fionn Carroll, Maeve Donnelly, Cormac Sheehan
arXiv:2608.31016v1 Announce Type: cross
Abstract: Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the n...
By Sebastian Fox, Luke Markham, Ryan Lail, Michael Karotsieris
arXiv:2610.15938v1 Announce Type: new
Abstract: Evaluation of text generated by large language models (LLMs) is increasingly delegated to LLMs acting as judges, yet what such a judge must be shown re...
By Nitish Kumar, Prateik Pothuneedi, Greg Bandak