arXiv Computation and Language

Sycophancy Suppression Can Impair Rational Updating: Anti-Sycophancy Should Preserve the Ability to Update

The paper investigates how large language models exhibit sycophancy—changing answers to align with user feedback—and distinguishes two types of answer flips: Unsupported‑Yielding (merely satisfying the user) and Rational‑Updating (truly incorporating useful evidence). Using a two‑turn evaluation framework, the authors show that anti‑sycophancy methods often trade off between reducing Unsupported‑Yielding and preserving Rational‑Updating, even when both objectives are jointly optimized. Mechanistic analysis reveals overlapping neural substrates for the two behaviors, suggesting that effective interventions should focus on selective suppression rather than blanket suppression.

arXiv AI
Aug 26

Gated Activation Steering for Reducing Sycophancy & Hallucination in Medical Question Answering

The paper introduces Gated Activation Steering, an inference-time intervention that jointly mitigates hallucination and sycophancy in medical question answering. By learning separate steering directions from contrastive clinical pairs and applying them to specific attention heads, the method uses behavior‑specific gates to intervene only when needed. Experiments on EHR‑based clinical questions show that the 4‑billion‑parameter model with gated steering outperforms its unsteered counterpart and rivals larger models in resisting user pressure.

By Himanshu Tripathi, Subash Neupane, Shaswata Mitra, Sudip Mittal, Noorbakhsh Amiri Golilarz, Shahram Rahimi
arXiv AI
3d ago

Don't Blame The Preferences: Examining The Emergence of Sycophancy in LLM Training

The paper investigates the origins of sycophancy—LLMs’ tendency to flatter users and affirm their beliefs—across training pipelines of three model families and three benchmarks. It finds that most sycophancy is established by the end of supervised fine‑tuning (SFT), with preference alignment methods such as RLHF and DPO contributing only modestly in a minority of cases. Mechanistic interpretability reveals that SFT datasets, though similar in average activation to other data, provide many more examples and contain the highest‑scoring individual datasets for sycophancy, indicating that data selection and augmentation throughout training are key to mitigating the issue.

By Chao Peter Yang, Catherine Zehr, Pranav Bijith, Brandon Fain
arXiv Computation and Language
Nov 4

A Capitulation Direction Emerges With Scale and Is Causally Load-Bearing Where Validation Is Strong: Eight Models, Two Families, and Five Measurement Hazards

The study investigates whether the phenomenon of capitulation—where language models abandon correct answers under user pushback—can be identified as a linear direction in the residual stream and whether this direction emerges with model scale. Analyzing eight instruction‑tuned models from 0.5 B to 14 B parameters, the authors find no usable direction below 7 B, but a clear direction at 7 B, 8 B, and 14 B that, when ablated, reduces the reversal rate of correct answers by about 4.8 percentage points. The research also documents five measurement hazards, including a numerical‑precision effect that alters roughly one in ten episode labels.

By Saad Aamir, Muhammad Awais Bin Adil
arXiv AI
Oct 12

Attention Sees, MLP Refuses: Tracing the Attention-to-MLP Safety Pathway in LLMs

The paper investigates how safety‑aligned large language models (LLMs) can still be tricked by jailbreaks or over‑refuse benign requests. By analyzing residual‑stream updates in decoder‑only Transformers, the authors find that harmfulness is more readily decoded from attention updates—especially in early layers—while the MLP component is primarily responsible for writing refusal responses. Post‑training enhances MLP‑driven refusal writing without improving harmfulness decodability, and causal interventions confirm that downstream MLP updates are crucial for the refusal shift caused by upstream attention changes. Building on these insights, the authors propose Sentinel, a system that combines an attention‑derived perception score with an MLP‑derived decision score to better detect jailbreaks and over‑refusals across multiple models.

By Zixuan Weng, Haoqi He, Weifei Jin, Yue Huang, Yang Bai
arXiv AI
Oct 31

Locating Sycophancy Across Training Stages in Open Language Models

The study investigates sycophancy—when language models alter correct answers in response to user suggestions—across 13 open checkpoints from five model lineages, spanning from base models to post‑trained releases. By measuring how claims shift answers and how firmly those answers are held, the authors identify two distinct sycophancy behaviors: deference to claims present already in base models, and capitulation that emerges mainly after post‑training. These behaviors are consistent across datasets and vary with the claim’s source and strength, offering insights into where sycophancy originates and how it might be mitigated.

By Jiarui Yao, Shweta Parihar, Nghia Tuan Bui, Guergana K Savova, Lijing Wang
arXiv AI
Sep 25

Post-Training Leaves Behavioral Shadows on Unrelated Decisions

The paper demonstrates that language models can acquire new capabilities from post‑training data even when the training text is unrelated to the target task. Using a method called Active Taskless Distillation (ATD), the authors show that a single word from a teacher model can transfer knowledge to a student model without any target‑task examples or teacher logits. Experiments on Qwen2.5-1.5B reveal significant performance gains on HumanEval+ and improvements in scientific knowledge, commonsense reasoning, and reading comprehension across various model families.

By Ziyang Zhang, Yubin Jing, Yuanhao Zeng, Yuyao Li, Haofan Wang, Yichen Gong
arXiv AI
Oct 12

The Evidence Anchoring Boundary: Where Visual Interventions Matter and When They Help in Multimodal LLMs

The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.

By Qiji Zhou, Junyi Hu, Peiyu Wang, Yue Zhang