CS Paper Radar

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

Serhii Mytsyk, Yiming Zhang, Vikram Krishnamurthy

cs.LG eess.SY ·submitted 2026-08-26 01:07 UTC · v1
benchmarktheory

Abstract

Large language models (LLMs) frequently exhibit \emph{sycophancy}: they adapt their answers to a user's stated beliefs or preferences instead of reporting what they hold to be true, which lowers factual accuracy and can amplify misinformation. This paper proposes a methodology for mitigating sycophancy that employs the Bayesian Truth Serum (BTS), a peer-prediction mechanism, as the reward in Group Relative Policy Optimization (GRPO) to fine-tune an LLM. BTS pays an answer for being \emph{surprisingly common}, that is, more frequent among respondents than those respondents themselves predicted. We treat a group of responses from a model for one question as those respondents, so the reward is a function of the model's own outputs and fine-tuning needs neither labels nor preference annotations. We prove that in the large-group limit a sycophantic response earns strictly lower expected reward than an honest one. We also prove that if the entire group agrees in advance on a symmetric answering rule, it cannot earn a higher information score than under truthful reporting. On our true/false benchmark the reference model's answer-flip rate under user pressure decreases from 23% to 4%, and its accuracy under that pressure increases from 80% to 93%. Our reward outperforms SMART and is comparable to synthetic-data fine-tuning and to pinpoint tuning, all three of which train on labels. It spends considerably more compute in exchange, which makes it suitable when labeled data is scarce. Peer Truth Serum, which also pays a premium for a rare answer but elicits no prediction report, reproduces the effect. A peer-prediction reward computed inside a single GRPO group therefore reduces sycophancy without labels, and comparing mechanisms suggests that the premium paid for a rarer answer drives the effect.

Closest papers

Nearest neighbours by meaning across the whole index.

Information Elicitation by Ordinal Agreement

cs.GT ·2026-11-05 ·Yiling Chen, Shi Feng, Grant Schoenebeck, Fang-Yi Yu
publishedtheory similarity 0.68
Abstract · 33 pages, 3 figures. Accepted to the 22nd Conference on Web and Internet Economics (WINE 2026)

Ordinal information is central to tasks ranging from conference peer review to Reinforcement Learning from Human Feedback for large language models. In many such settings, there is no objective ground truth against which reports can be directly verified, making incentive-compatible elicitation challenging. Peer-prediction mechanisms have been designed to achieve truthful elicitation despite the lack of ground-truth verification. However, existing peer-prediction methods do not provide incentives for agents to truthfully compress rich private information (e.g. reviews) into coarser ordinal reports (e.g. rankings) requested by the mechanisms. We propose a simple ordinal peer-prediction mechanism that rewards agents based on pairwise agreement, measured by Kendall tau. Under a broad class of Bayesian Stochastically Order Consistent models---which includes classical models such as Mallows and Thurstone---we prove that truthful reporting is stochastically dominant, and strongly truthful under index-equivariant strategies. This provides robust incentives even for agents with non-linear utility functions, such as risk aversion. Our mechanism further accommodates heterogeneous agents, top-K reports, and subset assignments, where each agent is scored against one randomly drawn peer. Experiments on synthetic Mallows and Thurstone data, together with a real-data stress test on the SUSHI preference dataset, show right-shifted truthful-payment distributions and the highest normalized sensitivity among tested baselines.

0.24

Expert-Guided Rubrics as Dense Reward Signals for Reinforcement Learning in Non-Verifiable Domains

cs.LG cs.AIcs.IR ·2026-10-17 ·Ashish Bansal, Hossam Amer, Keerat Kaur Guliani, Parinita Edke, Lesley Matheson +1 more
similarity 0.68
Abstract · 11 pages, 4 figures, RL4LLM-Agents workshop ICAIF ACM

Large language models (LLMs) are increasingly deployed as autonomous agents in non-verifiable domains such as financial services, where they must interpret user intent, rewrite queries, and retrieve information under domain-specific constraints. Although reinforcement learning (RL) can train such agents from the consequences of their actions, recent successes largely rely on verifiable tasks with automatically computed binary rewards. In non-verifiable retrieval tasks, assessing query quality requires scarce subject-matter expert (SME) judgment that is inherently graded rather than binary. With Group Relative Policy Optimization (GRPO), sparse rewards yield no useful learning signal when all sampled outputs within a group receive the same outcome, causing optimization to stall. We introduce a method that converts a small, fixed investment of SME judgment into a dense reward and combines it with retrieval-based metrics to form a hybrid, GRPO-compatible training signal. Experiments show improvements of up to 6\% in Recall@3 and 6.7\% in MRR@3 compared with sparse-reward baselines.

0.21

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

cs.LG cs.CL ·2026-08-18 ·Rubén Balbastre, Juan Manuel Orduña, Mariano Pérez
benchmark similarity 0.67
Abstract · 29 pages, 5 figures. Code and artifacts linked in the paper. v2: Extended the held-out evaluation to include broad-topic helpfulness, replac…

Practical LLM unlearning is usually evaluated through two objectives: suppress target-specific knowledge and preserve non-target utility. In generative QA, this leaves a third behavior underspecified: when a target-adjacent prompt admits a broader answer without target-specific leakage, the model should answer at that level rather than leak, evade, or refuse. We study this specification problem in a controlled LoRA-GRPO RWKU setting, comparing four reward designs that span lexical suppression, anti-refusal shaping, rubric-based broad answering, and an explicit refusal contrast, with and without SFT warm-up. The experiments show that optimization success is not equivalent to behavioral unlearning: RWKU forget scores, held-out completion audits, and training dynamics can point to different conclusions. We trace these disagreements to reward-hacking endpoints, policy-support limits in GRPO, benchmark probes that miss endpoint changes, and a rubric reward that selects broad-topic answering with low semantic leakage under held-out evaluation.

0.20

Mitigating Social Sycophancy via Pluralistic Preference Optimization

cs.AI ·2026-10-01 ·Stephane Hatgis-Kessell, Myra Cheng, Xiaoxuan Hou, Qian Hu, Rahul Gupta +2 more
similarity 0.66
Abstract

Personal advice, including relationship advice, now ranks among the most common uses of generative AI. But language models (LMs) exhibit sycophancy: they affirm users much more often than humans do, which can make people overconfident and less willing to repair their relationships after a conflict. Prior work on mitigating sycophancy has focused on factual settings where a response can be checked against a ground truth answer, while mitigations for social sycophancy (e.g., personal advice, where there is no ground truth) have relied on simple prompting and post-training methods with limited effectiveness. Our insight is that social sycophancy occurs in part because LMs overly center on the user and fail to consider the perspectives of other stakeholders impacted by the user's behavior. To address this problem we propose Pluralistic Preference Optimization (PlurPO): given inputs describing interpersonal conflicts, the LM identifies and simulates the relevant stakeholders, and is then trained to prefer and generate responses acceptable to all stakeholders. PlurPO uses only signals the model produces about its own outputs, without ground-truth labels. PlurPO substantially reduces social sycophancy across four datasets and four model families compared to prior methods. For example, on statements of intent to cause harm, where the users' actions should not be endorsed, PlurPO reduces the endorsement rate by 89% on average across four models. On general advice questions, where the target is to match the endorsement rate of human responses, it closes the gap by more than half, from 17.8% to 8.0% on average. The preference dataset constructed by PlurPO for an 8B model also effectively transfers to mitigating sycophancy in a larger (32B) model. Our results indicate that social sycophancy can be reduced by leveraging a model's own capabilities to simulate a plurality of relevant perspectives.

0.26

Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders

cs.CL ·2026-10-13 ·Xin Li, Hao Jiang, Jinghao Guo, Weisi Lin, Chau Yuen
positioncode/data similarity 0.65
Abstract · Project Page: https://lixin.ai/ReMaRk/

A recommendation request on Reddit draws several replies, each naming different titles. When an LLM recommender is trained on such threads with reinforcement learning (RL), which replies should a generated list be rewarded against? The state-of-the-art recipe, Rank-GRPO, picks one reply per training example, so the advantage mostly records which reply was drawn (85% of its variance) rather than how good the list is. We argue that RL should train on the request, not on each answer, and propose ReMaRk: it scores every list once against the union of all replies' titles and samples each request equally. This makes RL cheaper. On seven backbones and three datasets, ReMaRk reaches the validation recall of 2,000 per-reply RL steps in 25 to 53% of the steps and leads at that budget; trained until neither improves, it is never significantly behind Rank-GRPO and in four of seven pairs reaches Rank-GRPO's best with fewer GPU-hours. The same reward helps on ProtoQA, where about 100 people answer each question, which points to a simple rule for RL with several valid human answers: reward them jointly.

0.47

Claim Framing as a Driver of LLM Sycophancy

cs.CL cs.AI ·2026-10-13 ·Jean Seo, Pepa Atanasova, Isabelle Augenstein
hot direction similarity 0.65
Abstract

Sycophancy in Large Language Models, the tendency to align responses with user preferences over ground truth, has largely been studied using questions without a clear ground truth and with a single, plainly stated user opinion. We present a systematic study of which framings of a user's claim trigger sycophancy in factual question answering, where every answer is objectively verifiable without extra context. Evaluating 14 models on 42K questions, we express correct and incorrect user opinions through various claim framings over 8 dimensions and 21 types, and measure a broad set of opinion-driven behaviors. We find that all models are susceptible to user opinions regardless of size. Models are most influenced by how confidently an opinion is stated, whom it is attributed to, and whether it is presented as shared between the user and the model. In contrast to prior work, framings based on the user's expertise, authority over the model, or social pressure matter little in our factual setting. Larger models resist incorrect opinions better but are no better at recovering from their own errors. Our results call for sycophancy evaluations that probe how an opinion is framed, not merely whether one is present, and that measure multiple behaviors beyond a single flip rate.

0.43

Small Language Models as Judges for Rubric-Based Reinforcement Learning

cs.CL ·2026-08-30 ·Fengyu Xie, Yilun Zhao, Bingsen Chen, Arman Cohan, Chen Zhao
top venuehighly cited similarity 0.65
Abstract · 9 pages, 1 figure; EMNLP 2026 Findings

Rubric-based reinforcement learning extends RL beyond tasks with exact answers or rule-based verifiers by scoring responses against instance-specific criteria. However, this makes reward computation expensive: training requires repeated rubric judging, often with proprietary APIs or local generative LLM judges with 7B parameters or more. We study whether smaller language models can serve as efficient and reliable rubric-based judges. To make this question measurable, we construct PointRubric and RaR-Science-Static, two pointwise rubric-based evaluation datasets with instance-specific criteria and itemwise satisfaction labels. We compare three ways of extracting criterion-level judgments from small models: Generative verdicts, Yes/No Logprob margins, and Probe judges. Across both datasets, the Qwen3-1.7B Probe judge achieves the strongest criterion-level agreement among these methods, outperforming Generative and Logprob judges. Used as a GRPO reward model, it trains a policy from 0.232 to 0.643 on RaR-Science rubric score, compared with 0.594 for an 8B Generative judge baseline, while the baseline requires 10.7$\times$ more reward-judge time. Task and domain transfer experiments further suggest that Probe judges preserve criterion-level reward structure across settings.

0.66

Flat-RLVR: Sharpness-Aware Reinforcement Learning with Verifiable Rewards for Reasoning Capacity in LLMs

cs.LG cs.AI ·2026-10-21 ·Yeonbum Kim, Hyun Kyu Lee, Sung Whan Yoon
similarity 0.65
Abstract · 9 pages, 8 figures

Reinforcement learning with verifiable rewards (RLVR), particularly through group-relative policy optimization (GRPO), has emerged as a central paradigm for the post-training of large language models (LLMs). While RLVR is highly effective at enhancing the reasoning capabilities of LLMs, it commonly induces \textit{distribution sharpening}, whereby the policy distribution becomes excessively peaked and rare-yet-correct trajectories are progressively suppressed, thus diminishing sample diversity. To this end, we promote flat-reward optima in RLVR to smooth the policy distribution. We theoretically link flat reward RLVR to reduced distribution sharpening, and empirically show improved sample diversity while maintaining single-attempt accuracy. Our method, called \texttt{Flat}-RLVR, searches for flat reward regions via sharpness-aware minimization during LLM post-training, combining this with GRPO variants to achieve better trade-offs between sampling efficiency and reasoning capacity.

0.08