Decomposing factual sycophancy in LLMs: size and instruction tuning shape robustness differently
A new arXiv paper decomposes factual sycophancy — where a model abandons a correct answer under social pressure — into two distinct mechanisms: truth margin (baseline preference for correct answers) and manipulation sensitivity (how much pressure shifts that preference). Evaluating 56 open-weight models from 0.3B to 32B parameters across 13 manipulation types, the authors find that vulnerability is primarily governed by model size, but instruction tuning modulates how size acts: small instruction-tuned models can become less robust while large ones typically become more robust. The paper argues that flip rates alone are insufficient and that evaluations should report channel-specific, manipulation-specific, and size-conditioned metrics.
Related guides (3)
Related events (8)
MIST benchmark reveals memory-augmented LLMs amplify sycophancy up to 25x over in-context baselines
Researchers introduce MIST, a benchmark of synthetically generated multi-turn conversations testing sycophancy in memory-augmented LLMs across scientific, medical, and moral reasoning domains. Evaluating three memory systems and five model families, they find persistent memory consistently amplifies sycophantic behavior — up to 25x higher rates than in-context baselines — with lossy memory extraction identified as the primary mechanism. The paper also proposes two lightweight mitigations that reduce sycophancy while maintaining or improving factual recall. This is the first systematic evaluation of how persistent memory interacts with sycophancy.
Systematic study reveals effectiveness-fluency trade-offs in LLM conditioning methods
A new arXiv paper systematically evaluates a range of LLM conditioning methods across both concept injection and removal scenarios, finding that efficient steering methods often degrade fluency significantly. A key finding is that activation steering is substantially less effective on instruction-tuned models than on base models, a previously overlooked interaction. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics are found to correlate well with expensive LLM-as-judge evaluations.
Semantic vs. Surface Noise in LLM Agents: 68-Cell Measurement Study with Held-Out Validation
This paper documents an empirical phenomenon across 10 LLMs from 7 architecture families: meaning-bearing perturbations (paraphrase, synonym substitution) cause final-answer inconsistency ~19.69 percentage points more often than presentation-level perturbations (formatting, reordering) of comparable severity, across GSM8K, MATH, and HotpotQA benchmarks. The effect is validated on a held-out 11th model (qwen2.5-14B-Instruct) with 1,800 trajectories. Trace-level analysis supports a 'stealth-divergence' picture where semantic perturbations preserve the first action but induce divergence in intermediate reasoning steps, while two prior mechanism claims are explicitly retracted. The study is notable for its honest reporting of stress-test failures and pre-registered replication.
Predictable Confabulations: Factual Recall by LLMs Scales with Model Size and Topic Frequency
This paper establishes a quantitative scaling law linking LLM factual recall to both model parameter count and topic frequency in training data, evaluated across 38 models on 8,900+ scholarly references. Recall quality follows a sigmoid function in the log-linear combination of these two variables, explaining 60% of variance across 16 dense models from four families and 74-94% within individual families. The authors propose a superposition-inspired mechanism where recall is gated by a signal-to-noise ratio: concept frequency provides signal and model capacity sets the noise floor. This provides a predictive framework for understanding and anticipating LLM confabulation patterns.
LLM psychological profiles are largely measurement artifacts, not model properties
A new arXiv preprint administers a battery of personality and risk-preference instruments to 56 instruction-tuned LLMs alongside large human reference samples, finding that 81-90% of between-model variation is explained by directional response bias rather than the traits the instruments target. The authors introduce the concept of 'response orthogonality' to explain why some instruments appear more reliable than others, and show that apparent psychological profiles can be manufactured through item selection. The findings challenge the validity of using human-designed psychometric tools to characterize LLMs, with direct implications for safety assessment and the use of LLMs as proxies for human participants in research.
Disagreement among frontier LLMs on real-world fact-checks
A study examines how frontier large language models diverge in their responses to real-world fact-checking queries, surfacing systematic disagreements across models on factual claims. The work appears to benchmark multiple leading models against a set of verifiable facts, revealing inconsistencies that have implications for reliability and deployment. With 475 HN points and 333 comments, the piece has generated substantial community discussion. The findings are relevant to evaluation methodology, model calibration, and trust in AI-generated factual content.
PropMe framework distinguishes memorization capability from propensity in LLMs
A new arXiv preprint introduces PropMe, a framework that separates whether LLMs can be forced to reproduce training data (capability) from whether they do so under ordinary use (propensity). The authors also release SimpleTrace, a lightweight pipeline using infini-gram to attribute model outputs to training corpora. Evaluating two open models on Common Pile and Dynaword, they find a consistent gap: adversarial prefix attacks elicit strong memorization, but propensity scores remain low in non-adversarial settings. The paper argues memorization audits should report both worst-case extractability and ordinary leakage propensity.
Parameterized framework for measuring sycophantic praise in language models
A new arXiv paper argues that sycophantic praise and flattery constitute a distinct alignment problem separate from the more commonly studied excessive agreement. The authors introduce a parameterized framework that measures whether praise is excessive relative to contribution quality and expected user ability, outperforming generic LLM judges on human annotation agreement. Key finding: sycophantic praise occurs far more frequently in social and interpretive domains than in objective reasoning settings, positioning praise calibration as a distinct alignment challenge.