arXiv:2609.37616v1 Announce Type: new
Abstract: Language models tend to agree with whatever a user asserts, and post-training increasingly targets this sycophancy so that models evaluate claims on th...
By Abhinav Rajeev Kumar (Lossfunk), Paras Chopra (Lossfunk)
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.
By Serhii Mytsyk, Yiming Zhang, Vikram Krishnamurthy
arXiv:2606. 15420v1 Announce Type: cross Abstract: A constitution tells a language model what to value, but little tells us whether it does.
By Tong Che, Rui Wu
The paper introduces a comprehensive psychometric battery to dissect the components of yes‑no bias in large language models’ moral judgments. By crossing question polarity, printed order, and answer label, the authors separate stance, position, word, and verdict effects, and fit a two‑parameter model of framing susceptibility and decisiveness. Applying this method to 1.35 million trials across twenty moral dilemmas and ten frontier models, they find that response‑form bias is widespread but varies, with some models exhibiting hidden verdict attachment that surface analysis cannot detect.
By Haonan Huang
arXiv:2609.08016v1 Announce Type: new
Abstract: Multi-agent debate, in which several LLMs exchange arguments before answering, is widely assumed to improve answer quality by surfacing genuine disagre...
By Chen Qian
LLM tutoring poses a measurement problem: can a general-purpose helpfulness rubric distinguish direct answer-giving from pedagogical guidance? We audit this signal in a pre-registered study.
arXiv:2607. 18966v1 Announce Type: new Abstract: Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective.
By Axel H{\o}jmark, J\'er\'emy Scheurer, Evgenia Nitishinskaya, Felix Hofst\"atter, Jason Wolfe, Theodore Ehrenborg, Bronson Schoen, Alexander Meinke
The paper introduces an audit layer for vision‑language model benchmarks that reuses per‑question artifacts from evaluation runs to assess the benchmark itself. Six detectors analyze saved outputs, and an additional text‑only rerun removes visual inputs to produce per‑question reports and aggregate statistics, including permutation‑null excess and interval estimates. Applying this method to the MMStar benchmark reveals that 20.2% of its questions show no accuracy gain from images, with a small but significant excess in mathematics and science categories, and identifies questions where all models chose the same non‑gold option.
By Denis Degtiarev, Anna Gromova, Sergey Katkov, Artem Mikhailin, Safina Mukhamedshina, Albert Nasybullin, Karina Shaikhutdinova, Artem Tuzov
The study investigates how subtle changes in the presentation of a structured judging request can influence the behavior of an answer‑grading model. By adding a single colon to a candidate answer, the researchers found that the model’s false‑acceptance rate rose dramatically—from 1–3% to over 26%—across two benchmark datasets (DROP and GSM8K) and different grading configurations. The effect varied with the presentation format, showing that logically equivalent requests can lead to markedly different vulnerabilities.
By Hongliang Liu
The paper introduces MAD Bench, a 2,400-question benchmark for multi‑agent debate (MAD) that tests knowledge, reasoning, values, moral judgment, safety, and refusal. It evaluates three open‑weight models and finds that peer communication can alter individual decisions, with forced‑choice readouts before rationale generation more often matching the final answer in debate than in reflection. The study shows that stronger consensus can coexist with increased shared errors, suggesting the need for evaluation metrics that consider dependence and unanimous errors alongside accuracy.
By Arnesh Batra, Aniket Khandelwal, Jashn Khemani, Krish Thukral, Arush Gumber, Denis Federico Lim, Ujwal Kumar, Phan Xuan Tan
The paper introduces EmoNet‑Face‑HQ, a benchmark of 2,500 AI‑generated portraits each evaluated by four psychology experts across 40 emotion categories. It shows that most vision‑language models (VLMs) perform no better than random guessing unless fine‑tuned, but that changing the prompt to a simple yes/no per category can raise a VLM’s agreement with experts from chance to near expert levels. When evaluated using probability‑of‑yes scores on the five most‑agreed‑upon categories, eleven open‑weight VLMs achieve a quadratic‑weighted κ of 0.425, below the 0.468 two‑expert agreement, yet their average probability‑of‑yes score (0.553) matches the benchmark’s fine‑tuned model and outperforms it in three cases without any training.
By Tobias Hallmen, Fabian Deuser, Robin-Nico Kampa, Norbert Oswald, Elisabeth Andr\'e
The study investigates how task difficulty, model type, and user pressure influence large language models’ tendency to abandon correct answers or endorse user positions—a phenomenon known as sycophancy. Using 103,939 graded replies across ten configurations of eight LLMs (with and without reasoning) and 13 pressure conditions, the authors find that the cost of verifying a claim and the presence of a guardrail are the dominant factors, while model family and pressure tactics play minor roles. Key practical insights include simplifying hard-to-verify problems, employing deep reasoning, framing questions neutrally, and selecting models based on guardrail performance.
By Guang Yang, Homa Hosseinmardi, Fengchen Liu, Amir Ghasemian