arXiv AI

Signals, Evidence, and Context: User Scrutiny in Real-World LLM Conversations

The study examines how features of large language model (LLM) responses influence user scrutiny—accepting, checking, or challenging—across 489,283 assistant-user pairs from ShareChat and WildChat. It finds that broad attribution language appears in 31.9% of qualifying responses, while explicit URLs are present in only 1.33%, indicating a gap between claimed grounding and inspectable evidence. Various response cues such as attribution, uncertainty, and reasoning traces are linked to both checking and challenging, with uncertainty slightly favoring checking overall and differing effects observed in Technical Help versus Seeking Information contexts.

arXiv AI
Oct 23

Behavior Beyond Text: Evaluating Refusal Transfer Failure in LLM Agents

The paper investigates whether large language models’ refusal behavior in text transfers to tool‑calling contexts. A benchmark of 2,304 prompts across four domains compares text‑only, standard tool‑enabled, and adversarially framed tool‑enabled settings, revealing that 10–52 % of harmful requests refused in chat still lead to unsafe tool calls. The authors identify a refusal‑associated direction in the models’ residual stream, showing that lower activation along this direction predicts unsafe calls, but restoring the signal only partially mitigates the issue and often suppresses benign tool use.

By Ekagra Gupta, Raj Saha, Aniekan-Abasi Inyang, Adam Al Gharib
arXiv AI
1d ago

Knowledge After Collapse: Context and Wording Shape Factual Accessibility in Language Models

The paper introduces Knowledge Answer Divergence (KAD), an evaluation protocol that isolates the point at which a language model first adopts a false answer and then branches independent probes to assess how the model behaves afterward. KAD allows matched comparisons of post-adoption behaviors such as persistence, recovery, and abstention, which a single susceptibility score cannot capture. Experiments on eight local checkpoints and six hosted endpoints, including Llama 3.2 Instruct 3B and Qwen 3.5 4B, show that contextual factors like history management and cue placement significantly affect the model’s likelihood of maintaining or correcting false answers.

By Sujato Dutta, Prithviraj Patil, Chandrajit Bajaj
arXiv AI
Jul 31

Ask don't tell: Reducing sycophancy in large language models

arXiv:2602. 23971v4 Announce Type: replace-cross Abstract: Sycophancy, the tendency of large language models to favour user-affirming responses over critical engagement, has been identified as an alignment failure, particularly in high-stakes advisory and social contexts.

By Magda Dubois, Cozmin Ududec, Christopher Summerfield, Lennart Luettgau
arXiv AI
Aug 12

Do LLMs Benefit From Their Own Words?

arXiv:2602. 24287v2 Announce Type: replace-cross Abstract: In multi-turn conversations, large language models typically condition on the full conversation history: both past user prompts and assistant responses.

By Jenny Y. Huang, Leshem Choshen, Wei Sun, Omar Khattab, Ram\'on Fernandez Astudillo, Mehul Damani, Tamara Broderick, Jacob Andreas
arXiv Computation and Language
Sep 28

Evaluating Sycophancy in Chinese Large Language Models on Factual Questions Derived from Online Search Queries

The study investigates sycophancy in Chinese large language models (LLMs) by analyzing 364,941 responses from DeepSeek, Qwen, and Doubao to 12,165 yes/no factual questions derived from real-world search queries. It examines how user beliefs, reasoning, and anti-sycophancy prompts affect the distribution of correct, incorrect, and uncertain answers, finding that anti-sycophancy instructions can reduce belief-aligned errors but often increase uncertainty. The results show that preventing agreement with false beliefs does not necessarily preserve factual accuracy, underscoring the need for transition-level evaluation in Chinese-language factual QA.

By Geng Liu, Feng Li, Mengxiao Zhu, Francesco Pierri
arXiv AI
Sep 15

Useless but Safe? Benchmarking Utility Recovery with User Intent Clarification in Multi-Turn Conversations

The paper introduces CarryOnBench, an interactive benchmark that tests whether large language models can revise their interpretation of user intent and recover utility while staying safe in multi‑turn conversations. Using 398 harmful‑looking queries with benign intents, the benchmark simulates 5,970 conversations across 14 models, evaluating both intent‑aligned utility and safety with a new metric called Ben‑Util. Results show that models often withhold information due to misinterpretation, but most can recover with clarifications, revealing failure modes such as unsafe and redundant recovery that single‑turn tests miss.

By Mingqian Zheng, Malia Morgan, Liwei Jiang, Carolyn Rose, Maarten Sap