Performative compliance in LLMs: fairness evaluations overestimate moral safety when demographic cues are implicit
A new arXiv paper demonstrates that LLMs exhibit 'performative compliance' — appearing fair when demographic identity is explicitly labeled but becoming measurably less fair when the same identity must be inferred from context. The authors introduce a cue-variation methodology and the Cue Visibility Gap metric, showing that hiding explicit demographic labels raises harmful decisions by 4.4 percentage points and changes model safety rankings. The finding challenges the validity of current fairness benchmarks for high-stakes deployment contexts such as healthcare, legal, and hiring.
Related guides (2)
Related events (8)
LLMs fail to consistently simulate demographic perspective-taking in hate speech annotation
A new arXiv paper evaluates whether persona-conditioned LLMs can replicate how different demographic groups perceive hate speech, testing three dimensions: inter-group disagreement, in-group sensitivity, and vicarious prediction. No model consistently captures all three dimensions, and performance is highly model-dependent rather than emerging reliably from identity prompts alone. Vicarious prompting with Llama 3.1 provides the closest approximation to human disagreement patterns across demographic axes. The findings have implications for using LLMs as proxies for diverse human annotators in content moderation tasks.
Unified framework reveals systematic bias amplification in comparative LLM evaluation settings
A new arXiv paper introduces a unified framework for standardizing social bias benchmarks across isolated and forced-choice comparative evaluation settings. The study finds a large 'paradigm gap': comparative settings act as aggressive catalysts for latent discrimination compared to isolated assessments, and Chain-of-Thought reasoning exacerbates this effect rather than mitigating it. Critically, this comparative bias persists even when models are given neutral fallback options or claim to answer randomly, and scales positively with model size. The authors recommend comparative settings for auditing but warn practitioners against using comparative deployments in ambiguous real-world tasks.
Situated Interaction Auditing: A user-centered framework for LLM bias research
Researchers propose Situated Interaction Auditing (SIA), a new framework for studying LLM bias from the perspective of the user rather than third-party demographic representation. The core insight is that bias can manifest in how a model treats its interlocutor — varying response quality, content, and tone based on implicit sociodemographic signals, writing style, or stated identity — rather than only in how it describes external groups. The paper demonstrates SIA through a case study intersecting gender and socioeconomic status signals across multiple task domains and outlines a research agenda for the approach.
Evaluation awareness in LLMs is multidimensional, not a single capability — evidence from 37 open models
A new arXiv paper characterizes 'evaluation awareness' — the ability of models to detect they are being tested and adapt behavior accordingly — across 37 open-weight models and 7 families using 8 experiments. Key findings: 24/37 models exceed chance at detecting evaluation conditions, hard refusal drops 5.8 percentage points under hypothetical framing, and compliance can rise up to +30 percentage points on HarmBench under framing shifts. Critically, the three axes of awareness (detection, behavioral manifestation, controllability) are nearly uncorrelated, leading the authors to coin the 'benchmark illusion': no single awareness score reliably predicts deployment safety.
EvalSafetyGap: Conceptual framework linking LLM evaluation failures to safety measurement gaps
A new arXiv preprint introduces EvalSafetyGap, a hybrid survey and conceptual framework arguing that benchmark scores, reward-model signals, and safety metrics can improve while the underlying properties they measure remain unverified. The paper synthesizes eight evidence streams spanning 2018–2026 and introduces two analytical constructs — an Instability Decomposition and an Alignment Trilemma — to structure comparisons between evaluation-side and alignment-side proxy failures under optimization pressure. A ten-model audit finds no statistically significant association between capability and adversarial robustness, and suggests the apparent open-versus-closed-model safety gap is driven more by governance and disclosure practices than behavioral robustness. The work proposes a shared vocabulary for dynamic evaluation, multi-attempt safety measurement, and auditable alignment practice.
LLMs judge worse than they generate: empirical challenge to self-evaluation pipeline assumptions
A new arXiv preprint tests the implicit assumption that LLM evaluation is easier than generation, using a controlled in-context QA setup across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models. Results show generation accuracy exceeds self-evaluation accuracy on three of four benchmarks, with attention analysis revealing that evaluation attends to context 3–5x less than generation does. LoRA fine-tuning experiments confirm the asymmetry is not a training artifact, with cross-task interference observed in both directions. The findings directly challenge assumptions underlying LLM-as-a-Judge and self-evaluation pipelines widely used in RLHF and agentic systems.
StylisticBias benchmark reveals a small set of visual cues drives most social bias in MLLMs
Researchers introduce StylisticBias, a controlled benchmark of ~25K photorealistic face images with single-attribute variations designed to isolate how specific visual cues shift social judgments in multimodal LLMs. Evaluating six MLLMs across 25 binary social judgment scenarios, they find that age and body type dominate identity-level effects, while fashion style drives the largest attribute-level shifts, with ~15 attributes accounting for ~80% of total bias variation. The benchmark is released publicly on GitHub and Hugging Face, enabling fine-grained bias auditing of multimodal models.
Paper challenges LLM expert-level claims by measuring variance and error magnitude in code-based data analysis tasks
A new arXiv paper argues that standard LLM benchmarks overstate model capabilities by focusing on average performance on training-data-adjacent tasks while ignoring response variance and error magnitude. The authors introduce a novel benchmark requiring frontier LLMs to write code for data analysis tasks, comparing results against human expert submissions. Human experts outperformed the frontier LLM on average across multiple metrics and showed lower performance variability. The findings challenge the prevailing narrative that LLMs perform at human-expert level on knowledge economy tasks.