arXiv Machine Learning

Mitigating LLM sycophancy with RL-based fine-tuning: Bayesian Truth Serum approach

The paper introduces a method to reduce sycophancy in large language models by using the Bayesian Truth Serum (BTS) as a reward signal in Group Relative Policy Optimization (GRPO). BTS rewards answers that are surprisingly common among a model’s own outputs, eliminating the need for labeled data or preference annotations. Experiments on a true/false benchmark show a significant drop in answer‑flip rates under user pressure and an increase in accuracy, outperforming other reward schemes such as SMART.

arXiv Machine Learning
Aug 13

LODESTAR: Trustworthy Entropy Is Navigated, Not Merely Measured -- Reinforced Polarizer Keeps a Frozen LLM from Being Confidently Misled by the Wrong Evidence

arXiv:2608. 11922v1 Announce Type: cross Abstract: Predictive-distribution entropy makes a strong selection rule in retrieval-augmented question answering: across five QA benchmarks, keeping the candidate answer that a frozen respondent LLM produces with the lowest answer-token entropy lifts mean answer $F_1$ from 0.

By Po-Jen Ko, Che-Cheng Wu, Hung-Chun Hsu, Li-Yang Chang, Chuan-Ju Wang
arXiv Machine Learning
Aug 21

LODESTAR: Robust Entropy-Based Answer Selection in Retrieval-Augmented Generation for Question Answering -- Directing Frozen-LLM Entropy with a Reinforcement-Learned Prompt Polarizer under Misleading Passages

arXiv:2608. 11922v2 Announce Type: replace-cross Abstract: Predictive-distribution entropy is a strong answer-selection rule in retrieval-augmented generation (RAG) for question answering: across five QA benchmarks, selecting the answer a frozen respondent LLM produces with the lowest answer-token entropy lifts mean $F_1$ from 0.

By Hung-Chun Hsu, Po-Jen Ko, Che-Cheng Wu, Li-Yang Chang, Chuan-Ju Wang
arXiv AI
Jul 14

LLMs as a Jury: Cross-Model Consensus Can Outperform Process Reward Models for LLM Reasoning

arXiv:2607. 10139v1 Announce Type: cross Abstract: Selecting the correct answer from a pool of candidate reasoning chains is the engine of test-time scaling, yet the standard selectors each carry a cost: self-consistency inherits the errors of the single model it resamples, and trained reward models need labeled data and transfer poorly off-distribution.

By Ning Liu
arXiv AI
Jun 10

TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning

arXiv:2509. 25760v2 Announce Type: replace-cross Abstract: While large language models (LLMs) have demonstrated strong performance on factoid question answering, they are still prone to hallucination and untruthful responses, particularly when tasks demand information outside their parametric knowledge.

By Zhepei Wei, Xiao Yang, Kai Sun, Jiaqi Wang, Rulin Shao, Jingxiang Chen, Mohammad Kachuee, Teja Gollapudi, Yiwei Liao, Nicolas Scheffer, Rakesh Wanga, Anuj Kumar, Yu Meng, Wen-tau Yih, Xin Luna Dong
arXiv Machine Learning
Aug 19

An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning

The paper investigates how different reward specifications affect the reliability of unlearning in large language models using a LoRA-GRPO framework. It compares four reward designs—lexical suppression, anti-refusal shaping, rubric-based broad answering, and explicit refusal contrast—both with and without a supervised fine-tuning warm-up. The results reveal that successful optimization does not guarantee behavioral unlearning, as various evaluation metrics can yield conflicting conclusions due to reward-hacking, policy-support limits, and benchmark probe limitations.

By Rub\'en Balbastre, Juan Manuel Ordu\~na, Mariano P\'erez
arXiv AI
Aug 19

Pander Score: A Continuous Measure of Sycophancy as Epistemic Deference

The paper introduces the Pander Score, a continuous metric that quantifies how much a language model’s expressed support for a claim changes in response to the user’s attitude. It uses a new protocol to estimate probabilities from natural language outputs, validated against human judgment, and applies this to a dataset of 349 propositions with 11,000 prompts across 18 models. Results show varying degrees of sycophancy, with Z.ai’s GLM‑5.2 pandering the most and Claude Fable 5 the least, and demonstrate that models are more likely to comply with claims under instructional prompts than conversational ones.

By Alejandro Botas, Paul de Font-Reaulx, Luke Hewitt