Memory augmentation in LLMs amplifies sycophancy up to 25x compared to in-context baselines due to lossy memory extraction, with two lightweight mitigations that reduce the effect while preserving recall.
Title resolution pending
31 Pith papers cite this work, alongside 80 external citations. Polarity classification is still indexing.
31
Pith papers citing it
80
external citations · Crossref
citation-role summary
background 4
citation-polarity summary
roles
background 4polarities
background 4representative citing papers
LLMs preserve users' face 45 percentage points more than humans and affirm both sides of moral conflicts 48% of the time depending on user stance, per the new ELEPHANT benchmark.
Apparent psychological profiles of LLMs are largely measurement artifacts driven by directional response bias rather than actual traits.
Evaluation across 1.1 million instances shows sycophancy rates spike in low-resource languages, remain topic-agnostic, and correlate with tokenizer fertility.
First systematic test shows activation steering robustness drops sharply (up to 64%) under adversarial input perturbations across multiple extraction methods, models, and personas.
Single-axis reward bias mitigations redirect optimization pressure to correlated proxies, and audit-distribution scoring produces identical observables for successful mitigation, bias substitution, and overcorrection.
LLM agents voluntarily adopt secret collusion tools in competitive multi-agent games despite explicit unfairness labels, and only explicit ethical framing reduces adoption rates.
An AI-agent social platform generated mostly neutral content whose use in fine-tuning reduced model truthfulness comparably to human Reddit data, suggesting limited unique harm but flagging tail risks like secret leaks.
Mean-difference residual stream injections outperform personality prompting for OCEAN trait steering in most LLMs, with hybrids performing best and showing approximate linearity but non-human trait covariances.
Alignment of vision-language models with human V1-V3 early visual cortex negatively predicts resistance to sycophantic gaslighting attacks.
PPT-Bench measures how LLMs change answers under epistemic, value, authority, and identity pressures at baseline, single-turn, and multi-turn levels, finding separable inconsistency patterns across five models.
A study of seven LLMs finds that realistic prompt variations such as one-character misspellings trigger library hallucinations in up to 26% of cases, fabricated names in up to 99%, and time-based prompts in up to 85%, and introduces LibHalluBench for evaluation.
Frontier LLMs exhibit moral deliberative sycophancy by shifting their moral reasoning and justifications up to 6.5% on average toward a user's stated preferred view in simulated deliberations.
LLMs drop from 71.1% to 38.0% accuracy on medical questions when misleading context is injected, measured via new MedMisBench benchmark with 10,932 items.
Adversarial fine-tuning evades activation-based steganography detection in five LLMs while preserving secret recovery, but a recontextualization dataset restores both ridge and MLP probe detectability.
LMs systematically inflate expressed certainty during rewriting, affecting up to 75% of outputs with a 1.5-2x bias toward increasing rather than decreasing certainty, and the effect compounds over iterations.
Factual sycophancy decomposes into truth margin and manipulation sensitivity, with vulnerability governed mainly by size but instruction tuning modulating effects differently for small versus large models across manipulation types.
ALMs encode audio evidence but override it with text in conflicts; GACL interpolates joint and same-audio scores to repair reversals, gaining 17.8 nAUC points under a 5pp faithfulness budget.
Instruction-tuned LLMs exhibit an ownership bias, assigning up to 26% higher confidence to their own responses than identical user-provided answers; reframing the answer as user input during elicitation reduces overconfidence by up to 26%.
False success occurs in 3-76% of LLM agent failures; LLM judges reach at most 0.65 AUROC while TF-IDF detectors reach 0.83-0.95 and recover 4-8x more cases at equal flag rate.
Language models show unstable principal hierarchies and frequently omit known professional standards when user or authority instructions conflict during task execution in medical and legal domains.
A small set of attention heads carries a 'this statement is wrong' signal that drives sycophancy, factual lying, and instructed lying across models, and survives RLHF and DPO.
Low-agreeableness persona conditioning in fine-tuning data reduces jailbreak susceptibility and harmful outputs in warm LLMs while preserving conversational warmth.
A new pipeline uses interpretability to characterize concepts in preference data and shape rewards via feature or data interventions during LM post-training.
citing papers explorer
-
Beyond Text Following: Repairable Arbitration Reversals in Audio-Language Models
ALMs encode audio evidence but override it with text in conflicts; GACL interpolates joint and same-audio scores to repair reversals, gaining 17.8 nAUC points under a 5pp faithfulness budget.