arXiv Machine Learning

Auditing the Judge: Internal Signals for LLM Reasoning Assurance

The paper investigates how internal signals from large language models can be used to improve error detection and budget allocation for answer review. By training a supervised correctness probe on attention-head activations, the authors show that this probe can identify 1.3–22.5 percentage points more errors at a 5% review budget across several 7–12B models on GSM8K and MATH, outperforming the model’s own self-judgment in many settings. While the probe improves in-domain allocation, its effectiveness drops when transferred across datasets, indicating that internal signals are not universally reliable confidence estimators.

arXiv Computation and Language
Sep 11

Rethinking Verbalized Confidence for LLM-as-a-Judge: A Compatibility Shift on Post-2025 Proprietary Models

The paper argues that verbalized confidence—once viewed as overconfident and coarse—has become the preferred soft‑scoring method for LLM‑as‑a‑Judge on top‑tier proprietary models released after 2025. Experiments on SummEval, AggreFact, and HelpSteer2 across up to 18 LLMs show that log‑probabilities are no longer the best signal, and that adding an overconfidence advisory and self‑debate further improves calibration and robustness. The authors note that these enhancements incur little accuracy loss on post‑2025 models but do affect pre‑2025 ones, highlighting a compatibility shift in how confidence should be measured.

By Yu-Chung Hsiao
arXiv Computation and Language
Oct 12

AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis

AMLworld-Compact introduces an importance‑weighted downsampling protocol for evaluating large language models on AMLworld’s test splits, retaining all illicit transactions while sampling benign ones from supervised‑score strata. The HT‑Coreset method reduces evaluation targets by 271× on HI‑Small and 611× on LI‑Small, providing unbiased confusion‑count estimates with exact recall and ratio estimates for precision and F1. Experiments with boosted‑tree strata and Graph Feature Preprocessor features show lower F1 RMSE compared to uniform sampling, and audit tools reveal that lexical checks do not reliably indicate correctness of model outputs.

By Natapong Nitarach, Phume Ngampornsukswadi, Pittawat Taveekitworachai, Surapon Nonesung, Warit Sirichotedumrong, Kunat Pipatanakul
arXiv AI
Aug 19

Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification

The paper investigates how personalized agents decide to use, ignore, update, or query retrieved user memory before acting on a task. An empirical audit protocol is developed to test structured intermediate outputs, revealing that while exposing state definitions improves accuracy, an explicit state-output field does not significantly enhance policy accuracy for large language models. The study also shows that example-level accuracy overstates consistency, with full four‑way family success being rare, and that providing benchmark‑associated state labels merely conditions predictions rather than proving internal fidelity.

By Yihang Chen, Pin Qian, Su Wang, Chong Peng, Huan Xu, Shuaiting Li, Yiqi Sun
arXiv AI
Aug 24

No Judgment Without a Reason: Counterfactual Receipts for Versioned AI Evaluators

The paper introduces a framework for evaluating AI systems that not only checks final labels but also tracks the reasoning behind them through three core sources—grounds, norms, and authority—forming an eight-cell counterfactual judgment cube. It defines minimal source replacement sets, called judgment receipts, to explain changes in verdicts and provides certification cost bounds for black-box evaluators. The authors present ReasonBench, a benchmark with 19,520 cases, and demonstrate that while high standard accuracy can mask robustness issues, receipt accuracy reveals significant gaps in reasoning consistency across different models.

By Ye Chen, Weining Zhang
arXiv AI
Sep 30

Multi-Channel Mitigation of Source-Trust Shortcuts in Fact-Checking RL Agents

The paper introduces TrustSwap, a counterfactual test that swaps or removes source reliability labels while keeping evidence text constant, to evaluate how retrieval‑augmented fact‑checking models respond across verdict, confidence, and search decisions. Experiments on untrained and RL‑trained models show that confidence and search largely follow labels, yet label changes can flip a significant portion of verdicts, especially in larger models. The authors propose trust‑swap augmentation (TSA) to mitigate this shortcut, demonstrating reduced verdict flip rates and maintained accuracy in several settings, though its effectiveness diminishes at larger model scales.

By Jianchang Su, Yiwei Yang, Wei Zhang
arXiv Machine Learning
Sep 14

Can We Trust LLM Judges: A Study of Capability-Dependent Biases and Multi-Judge Ensemble for Bias Calibration

The paper investigates how large language models (LLMs) used as judges in absolute scoring tasks exhibit systematic biases that compromise reliability. It shows that a judge’s task accuracy strongly predicts both its judging accuracy and its directional bias, yet more capable examinee models consistently receive more lenient judgments. To mitigate these biases, the authors propose a calibrated weighted majority voting (WMV) ensemble that estimates judges’ error rates from inter-judge agreement patterns, achieving near-oracle performance without labeled data and improving both accuracy and fairness.

By Gemma Zhang, Prachi Badarayani, Asmi Kumar, Sadid Hasan, Sulaiman Vesal
arXiv AI
Oct 15

Checking the Checkers: A Synthesis of Verifier Noise and Robustness in RLVR

The paper surveys 58 studies on verifier errors in reinforcement learning with verifiable rewards (RLVR), examining how such errors affect training and which mitigations are effective. It finds that the type of error—rather than its frequency—has a larger impact on learning dynamics, with random errors mainly slowing progress and false positives potentially causing learning plateaus or collapse. The authors compile findings into comparison tables, reconcile conflicting reports, and propose a reporting protocol to standardize future research.

By Savyasachi Gupta, Aniruddha Srinivas Joshi
arXiv Computation and Language
Aug 27

Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation Independence

The study investigates how prior scores influence large language model (LLM) judgments in the LLM-as-a-Judge paradigm. By testing three prompt conditions—no metadata, revision framing, and anchored metadata containing prior scores—the authors find that prior scores systematically bias evaluations, shifting ratings toward those scores across 192,000 attempts. The bias also affects categorical decisions, blocking 48% of error corrections and flipping 10.18% of correct judgments, and is not mitigated by Chain-of-Thought or a warning, underscoring the need for careful context engineering.

By Ante Kapetanovic, Kemal Altwlkany, Andro Mercep, Tomislav Duricic, Emanuel Lacic