Searcharxiv⌕ Search

arXiv subjects

Sanmi Koyejo

Publications and source records attributed to Sanmi Koyejo.

At least 37 records · Page 2Linked to original sources
arXiv2026

What AI Benchmarks Actually Measure: Adapting Convergent and Discriminant Validity to Interrogate Fifty-Six AI Benchmarks

Benchmarks play a central role in the development and governance of models, yet it is often unclear whether they actually measure the concepts they purport to measure (e.g., reasoning, refusal). We adapt convergent and discriminant validity from the social sciences into an approach for interrogating AI benchmarks, applying it to 56 capability and safety benchmarks across 53 models. We label benchmarks with substantively similar purported concepts to a shared assigned concept, and ask whether model rankings on benchmarks with the same assigned concept correlate more strongly than rankings on benchmarks with different assigned concepts. We ask analogous questions at the item level using item response theory (IRT) models. We find that correlations between model rankings on benchmarks with the same assigned safety concepts are often weak, suggesting these concepts may be conceptualized inconsistently across benchmarks. For assigned capability concepts (e.g., reasoning, knowledge), model rankings are often as strongly correlated among benchmarks with the same assigned concept as between benchmarks with different assigned concepts, suggesting these capability concepts may not discriminate well from one another. In some cases, benchmarks that share design elements (e.g., score format) correlate more strongly than benchmarks with the same assigned concept. Finally, some individual benchmarks correlate more strongly with benchmarks assigned a different concept than with benchmarks sharing their own assigned concept, suggesting they may measure a different concept than they purport to. For example, BBQ-accuracy correlates more strongly with benchmarks labeled reasoning than with benchmarks that share its assigned concept, bias. To support future empirical work on benchmark validity, we release our extensive dataset of model outputs and scores at the item- and benchmark-level.

cs.CY↗
arXiv2026

CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment

Improving the safety of large language models (LLMs) often comes at the expense of utility, as globally applied safety tuning may affect model responses to both harmful and benign inputs. We propose \textbf{C}ontinuous \textbf{L}at\textbf{E}nt \textbf{A}dapter \textbf{R}outing (CLEAR), a conditional safety adaptation framework that uses a lightweight hidden-state gate to continuously control the activation strength of a safety low-rank adapter. CLEAR aims to reduce harmful completions while avoiding unnecessary changes to the frozen backbone that could degrade performance on benign prompts. Experiments on widely used safety and utility benchmarks show that CLEAR improves robustness on HarmBench while reducing the utility degradation observed with globally applied safety tuning such as SFT or standard low-rank adaptation (LoRA). On Llama-3-8B-Instruct, CLEAR reduces HarmBench ASR from 32.3\% to 0.5\%, while retaining most of the base model's utility and achieving up to 7.1 percentage points higher GSM8K accuracy than globally applied SFT or LoRA. These results suggest that CLEAR is a promising mechanism for improving the safety--utility trade-off in LLM alignment.

cs.AI↗
arXiv2026

Bringing analytic rigor to agentic AI for science: The Brain Researcher platform for neuroimaging data analysis

AI agents can execute scientific analyses, but an analytic output becomes a defensible claim only after alternatives are weighed and the claim is limited to what the evidence supports. Agents may reproduce failures including selective analysis, premature declarations of success and optimization of imperfect criteria. We present Brain Researcher, an agentic research harness operating in a neuroimaging researcher's computational environment under rules for admissible analyses, required checks and claim scope. In benchmarks, Brain Researcher increased first-choice tool-selection accuracy across seven models by 70.2 percentage points (23.3% without it versus 93.6% with it) and verifiable grounding from 4.6% to 22.0%. In collaborator-led and self-evolving studies, multiverse analyses exposed analytic-choice sensitivity, and scientific review classified claims as accepted, qualified, revised, blocked, rejected or deferred. By linking decisions to evidence and provenance, Brain Researcher embeds methodological judgment within the workflow, not after it.

cs.AI↗
arXiv2026

What to Forget in Unlearning? Forget Set Curation for Language Models

Machine unlearning aims to remove targeted data or behaviors from a trained model without retraining from scratch. Yet most evaluations assume that the examples to forget are already known. In realistic language-model deployments, a requester may ask a model to stop reproducing a song or book without knowing which spans, documents, quotations, or near-duplicates in a trillion-token corpus support that behavior. We study this missing upstream problem, forget set curation: mapping a suppression request to the data passed to an unlearning algorithm. We introduce CleanSlate, a benchmark for verbatim output suppression over songs and books, with model-specific extraction profiles, content-grounded QA, and capability-retention evaluations. CleanSlate exposes two failure modes. Natural lexical and exact-substring curators often yield forget sets that lead to weak suppression. An evaluation-aware curator suppresses requested continuations almost completely, but causes collateral regression on non-requested content and model-dependent capability loss. These results show that practical unlearning is not only an optimization problem once a forget set is given: the data chosen for forgetting determines both what can be unlearnt and what else is damaged.

cs.CL↗
arXiv2026

Why Do Safety Guardrails Degrade Across Languages?

Large language models exhibit safety degradation in non-English languages. Standard evaluation relies on Jailbreak Success Rate (JSR), which confounds several safety-driving factors into one, obscuring the specific cause(s) of safety failure. We introduce a latent variable model, a Multi-Group Item Response Theory (IRT) framework, that decouples language-agnostic safety robustness ($θ$), intrinsic prompt hardness ($β$), global language processing difficulty ($γ$), and a prompt-specific cross-lingual safety gap ($τ$). Using the MultiJail dataset, we evaluate the safety robustness of 61 model configurations across 5 closed-model families and 10 languages of varying resource, aggregating a dataset of 1.9 million responses. Exploratory Factor Analysis shows safety is primarily unidimensional: models refuse different harm types mainly through a shared mechanism. Contrary to the expected trend that safety degrades largely in low-resource languages, 22 model configurations are more vulnerable in English than in low-resource languages. Low-resource languages produce more uncertain responses (high entropy) than high-resource languages. Also, high-$τ$ prompts cluster in physical harm categories like Theft and Weapons and lower-resource languages, trends validated through cross-dataset generalization. While global translation quality shows low correlation with $τ$, severe mistranslations drive high-bias outliers, as validated by native speakers. Cultural and conceptual grounding mismatches may also contribute to $τ$. In predictive validation, the IRT framework achieves $\mathrm{AUC} = 0.940$, and unlike rate baselines stays predictive when a whole language is held out ($0.875$). Our framework reveals concept-language vulnerabilities that aggregate metrics obscure, enabling fairer cross-lingual safety evaluation and targeted improvements in dataset construction.

cs.CL↗
arXiv2026

In-Situ Behavioral Evaluation for LLM Fairness, Not Standardized-Test Scores

LLM fairness should be evaluated through in-situ behavioral pattern rather than standardized-test Q&A benchmarks. We show that the standardized-test paradigm can be structurally unreliable: surface-level prompt construction choices, although entirely orthogonal to the fairness question being tested, account for the majority of score variance, shift fairness conclusions in both the direction and the magnitude, and result in severe discordance in model rankings. We develop MAC-Fairness, a framework that embeds controlled variation factors into in-situ behavioral evaluation, examining how models' disparate-treatment behaviors shift when identity is varied as part of natural multi-agent conversation. Repurposing standardized-test questions as conversation seeds rather than as the evaluation instrument, we evaluate within-model differences in position persistence (how they hold positions, from the self-perspective) and peer receptiveness (how receptive they are to peers, from the other-perspective) across 8 million conversation transcripts spanning multiple models and identity presence configurations. In-situ behavioral evaluation reveals stable, model-specific, disparate-treatment behavioral signatures that could generalize across different fairness benchmarks, a form of evidence the standardized-test paradigm does not offer.

cs.CL↗
arXiv2026

When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation

Artificial intelligence benchmarks are an important mechanism to measure model progress and guide deployment decisions. However, benchmarks quickly "saturate", making it difficult to differentiate models and diminishing their long-term value. In this study, we define benchmark saturation and analyze it across 60 language model benchmarks using 14 properties that relate to saturation. We find that nearly half of our benchmarks exhibit saturation, with rates increasing with age. Further, we find that resilience to saturation is impacted by expert-curation, not by public test data. Our results suggest that design choices can extend benchmark longevity and inform more durable evaluation approaches.

cs.AI↗
arXiv2026

Positive-Unlabeled Preference Optimization For Chest X-ray Report Generation

Vision-Language Models (VLMs) for radiology report generation are typically trained on retrospective clinical reports, which suffer from omission noise: clinically present findings are left unreported due to the omission of subtle findings. For example, prior studies show that cardiomegaly may be omitted from ICU chest X-ray reports when the imaging request is focused on monitoring support device placement. As a result, models trained with standard approaches inherit these omissions, learning to under-report findings themselves. We propose PU-DPO, a preference optimization framework to prevent omission noise from corrupting the preference signal. We reformulate the objective under a positive-unlabeled (PU) learning framework, treating absent mentions as unlabeled rather than truly negative. Our framework provides preference supervision using constructed contrastive pairs, generated using edits to model responses, producing variants that explicitly mention or omit a specific finding. Generated responses that mention the finding are naturally preferred in the context of visual evidence. Across semi-synthetic experiments and analyses on real-world chest radiograph benchmarks where adjudicated labels are available, PU-DPO yields consistent gains in detection rates and recovery of hidden positives across multiple pathologies, and is more robust to omission noise than prior approaches.

cs.CV↗
arXiv2026

When Behavioral Safety Evaluation Fails: A Representation-Level Perspective

Safety evaluation of large language models (LLMs) is largely behavioral: a model is certified safe when it refuses harmful requests and answers benign ones. But refusing on the prompts an auditor happens to try does not show that the model is far from harmful behavior. Behavioral tests observe outputs; they do not measure how easily an intervention on the model turns a refusal into compliance. We call the gap between what static audits certify and what an intervention can reach the audit gap, and we show it is realizable: one can build a model that matches its safety-aligned base on every static audit yet gives way to a small, known perturbation of its internal state. We construct such dissociated models from three safety-aligned bases (Gemma 2 2B, Llama 3.2 3B, Qwen 2.5 3B) and audit the base, dissociated, and openly harmful models with the same soft interventions in parameter and latent space; the latent attacks are summarized by the Latent Vulnerability Score (LVS), the safety degradation produced per unit of bounded latent perturbation. Every static audit we run gives the dissociated model the same verdict as its base, since its refusals match the base, jailbreaks show no consistent signature, and a strong fixed probe on clean activations cannot tell it from the base. The same interventions an auditor could run reverse the verdict. At the targeted mid layer the dissociated models score 2.5 to 3.1 times higher LVS than their bases. A bounded latent attack elicits harmful compliance on 54 to 86% of prompts, against 3 to 48% for the bases, while matched random perturbations stay at or below 12%. Harmful fine-tuning reaches high compliance within five gradient steps, where the bases need 10 to 25. Behavioral testing, even with static latent probing, cannot certify representation-level robustness: a safety audit must intervene on the model, not only observe it.

cs.LG↗
arXiv2026

Underwriting the Agent Economy: The Blueprint for an AI Insurance Stack

From maritime trade to commercial nuclear power, insurance has been the enabler of major economic and technological developments by pricing risk, limiting downside, and spreading best practices. The emerging AI agent economy, projected to handle trillions of dollars in transactions by 2030, looks to be the next such development. Yet insurers' exposure to AI agent risk currently sits largely unpriced across existing insurance lines; between this silent coverage and growing exclusions, coverage is not fit for purpose. Furthermore, insurability is trending the wrong way: AI agent capabilities appear to be outpacing reliability, leading to rising incident severity; concentration among a few foundation model providers threatens correlated losses; and traditional actuarial modeling will struggle to keep pace with a technology evolving as rapidly as frontier AI. This report argues that affirmative AI coverage with limits in the billions is achievable by 2030, but only with industry-wide coordination. Drawing on successful historical precedents such as Underwriters Laboratories, the Closed Claims Project, and others, we lay out an eight-component AI insurance stack spanning incident data collection, catastrophe modeling, standards, contract design, risk selection, pricing, monitoring, and claims management. Building out this infrastructure is what will enable insurers to cover and manage AI agent risk sustainably and at scale. Finally, we discuss coverage for catastrophic risk from frontier AI ("AI CAT"), including CBRN, critical infrastructure collapse, and loss of control scenarios. Addressing these tail risks will require purpose-built instruments, potentially including a frontier model developer mutual, catastrophe bonds, bespoke liability regimes, and government backstops.

cs.CY↗
arXiv2026

The Illusion of Robustness: Aggregate Accuracy Hides Prediction Flips under Task-Irrelevant Context

As large language models (LLMs) grow more capable, they are increasingly deployed in context-rich settings where task inputs are often accompanied by long, partially irrelevant context. In a controlled setting, we find that state-of-the-art models often appear robust to task-irrelevant context at the aggregate level: prepending it to benchmark questions causes little change in overall accuracy. This aggregate stability, however, masks significant per-example instability. Even semantically meaningless pseudo-words, formed by randomly combining characters, can markedly shift model predictions on a small fraction of examples, degrading performance on some while improving it on others. This two-sided effect holds consistently across a wide range of models and datasets, yet the affected examples are largely model-specific. We further show that this instability is modulated by context type, context length, test-time compute, and model development stage. Together, our findings reveal context-induced tail risks concealed by aggregate accuracy, motivating per-example reliability evaluation of language models.

cs.CL↗
arXiv2026

Auditing of Unlearning Algorithms

Evaluating whether unlearning algorithms truly remove training data influence remains an open challenge. We propose a practical auditor that computes data-dependent lower bounds on the unlearning parameter $\varepsilon$ using membership inference attacks. Evaluating multiple unlearning algorithms, we find a sharp separation: algorithms with rigorous guarantees, such as model clipping and rewind-to-delete, achieve very small $\varepsilon$ bounds that do not falsify their unlearning guarantees, whereas empirical methods such as Hessian-based unlearning, interleaved ascent-descent, ascent on the forget set, and fine-tuning on the retain set exhibit large bounds, indicating poor unlearning. Our auditor provides a practical tool for empirically falsifying unlearning claims through a hypothesis-testing framework, and we validate it on CIFAR-100 and Shakespeare text.

cs.LG↗
arXiv2026

Stop Automating Peer Review Without Rigorous Evaluation

Large language models offer a tempting solution to address the peer review crisis. This position paper argues that today's AI systems should not be used to produce paper reviews. We ground this position in an empirical comparison of human- versus AI-generated ICLR 2026 reviews and an evaluation of the effect of automated paper rewriting on different AI reviewers. We identify two critical issues: 1) AI reviewers exhibit a hivemind effect of excessive agreement within and across papers that reduces perspective diversity. 2) AI review scores are trivially gameable through paper laundering: prompting an LLM to rewrite a paper could significantly increase the scores from AI reviewers, demonstrating that LLM reviewers are easy to game through stylistic changes rather than scientific results. However, non-gameability and review diversity are necessary but not sufficient conditions for automation. We argue that addressing the peer review crisis requires a science of peer review automation -- not general-purpose LLMs deployed without rigorous evaluation.

cs.AI↗
arXiv2026

Interactive Multi-Objective Probabilistic Preference Learning with Soft and Hard Bounds

High-stakes decision-making involves navigating multiple competing objectives with expensive evaluations. For instance, in brachytherapy, clinicians must balance maximizing tumor coverage (e.g., an aspirational target or soft bound of >95% coverage) against strict organ dose limits (e.g., a non-negotiable hard bound of <601cGy to the bladder). Selecting Pareto-optimal solutions that match implicit preferences is challenging, as exhaustive Pareto frontier exploration is computationally and cognitively prohibitive, necessitating interactive frameworks to guide users. While decision-makers (DMs) often possess domain knowledge to narrow the search via such soft-hard bounds, current methods often lack systematic approaches to iteratively refine these multi-faceted preference structures. Furthermore, DMs often require confidence that they have not overlooked superior alternatives, a paramount necessity in high-stakes scenarios. We present Active-MoSH, an interactive local-global framework designed for this process. Its local component integrates probabilistic preference learning with an active sampling strategy to adaptively refine Pareto subsets while minimizing cognitive burden. To bolster decision confidence, Active-MoSH's global component, C-MoSH, leverages multi-objective sensitivity analysis to identify potentially overlooked, high-value points beyond immediate feedback. We demonstrate Active-MoSH's performance benefits through diverse synthetic and real-world applications. A high-stakes case study with real cervical cancer brachytherapy treatment plans and an image selection user study further validate our hypotheses regarding the framework's ability to improve convergence, enhance DM confidence, and provide expressive preference articulation.

cs.AI↗
arXiv2026

Gaming Consensus: Coordinated Manipulation in Crowdsourced Fact-Checking

Crowdsourced fact-checking systems have been adopted by major social media companies such as X, Meta, TikTok and Google with the aim of combating misleading information at scale without relying on centralized editorial control. These systems have been developed around a common underlying concept: a bridging mechanism that identifies notes flagging misleading information when they receive support from people with different perspectives rather than simple majority support. To our knowledge the only publicly disclosed bridging algorithms deployed for fact-checking are based on matrix factorization, as deployed by both X and Meta, augmented with additional components addressing abuse, targeted manipulation, and contributor brigades. This work examines the core matrix factorization portion of these systems, presenting theoretical and empirical evaluations of the degree to which coordinated users could vote strategically by leveraging the latent representations to fabricate the appearance of synthetic consensus within the bridging mechanism. Using historic production data, we find that up to 10.7% of lower quality notes could be manipulated above consensus thresholds using less than 10 ratings. We complement these findings with a theoretical analysis, revealing counterintuitively that rating a note as "Not Helpful" can increase its helpfulness score, as well as a cost model quantifying manipulation effort. We have developed and deployed mitigations within X's Community Notes algorithm to address synthetic consensus.

cs.LG↗
arXiv2026

AI Writes Faster Than Humans Can Review: A Longitudinal Study of an Enterprise 2x Mandate

Enterprises increasingly mandate AI coding tools and report large productivity gains, yet longitudinal evidence on how such a mandate unfolds is scarce. In this paper, we present a quantitative case study of a documented enterprise "2x" mandate at a mid-sized, AI-forward company that has been committed to doubling merged pull requests per engineer since mid-2025. In a panel of 802 developers and 196,212 pull requests (January 2024-April 2026), per-capita throughput eventually doubled, reaching 2.09x the pre-mandate baseline in April 2026, among the largest gains reported from a field deployment of AI coding tools to our knowledge. A staggered difference-in-differences design links the within-developer share of this gain to AI adoption and to a further gain that grows with accumulated use, with the mandate acting as a catalyst rather than a direct driver. Because adoption and usage intensity were not randomly assigned, we read this evidence as strongly implicating an adoption-and-use channel rather than as exact causal attribution. The gain is broadly shared across seniority yet concentrated in newer code and not separable across model generations. Adoption also restructured code review around automation: per-reviewer load roughly doubled and automated review overtook human review, while merge and revert rates held steady.

cs.SE↗
arXiv2026

Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models

Standard benchmarks for multimodal large language models (MLLMs) score each item on one canonical ordering and miss whether order-irrelevant shuffling changes the answer, a baseline reliability property called for by emerging AI evaluation guidelines. We introduce Facet-Probe, a five-facet audit (option, evidence-chunk, document-rank, image-set, and mixed-modality ordering) of 18 frontier and open-weight MLLMs. A Bayesian item-response model separates ordering noise from per-facet bias, and a same-ordering control estimates the decoder-stochastic floor for observed flips. We find that none of the 18 MLLMs we audit are order-invariant: screened per-facet panel-mean flip rates span 24-50%. A Gemini same-ordering control at temperature 0 estimates a substantial ordering excess over a same-input decoder-noise floor in verified cells. Capability predicts but does not eliminate flips; the best model still flips on 13.4% of trials. In our Gemini mitigation tests, training-free prompt changes are modality-conditional and do not transfer from text to visual reasoning. These results suggest that prompt-level mitigation alone is unlikely to provide general order robustness, motivating future work on training-time and architectural approaches. We propose cross-ordering flip rate as a standard reporting axis for MLLMs.

cs.CL↗
arXiv2026

Internal Data Repetition Destroys Language Models

Language models are running out of high-quality training data, and even aggressively deduplicated corpora retain some amount of repetition. Earlier controlled studies predated Chinchilla-style scaling laws and could only measure the cost of repetition indirectly. We revisit repetition in the Chinchilla era, using a fitted no-repetition scaling law to report Compute-Equivalent Gain and Compute-Equivalent Loss. We show that under this modernized paradigm, repetition damage is systematic in three ways. First, holding compute allocated to repeated data constant, eval loss peaks at an intermediate repeat count $\Rep$; repeating a moderately sized subset a moderate number of times damages performance more than repeating a large subset a few times or a small subset many times. Second, the location of this peak is well-fit by a power law in model size; this scaling law reveals that the most damaging number of repeated data grows more quickly than compute. Finally, when repeated documents consume 10\% of the FLOPs budget in a controlled exact-document repetition setting, the compute-equivalent loss can be large: on FineWeb-Edu-Dedup, the most damaging repeat count for a Qwen3-style 344M-parameter model at $\OT=1$ matches the loss of a no-repetition run using 67% of the FLOPs. We demonstrate that these phenomena are not language-model-specific, and can be analytically understood in a simple statistical model: a misspecified linear regression with verbatim duplicates reproduces the same qualitative loss peak, quantifying how such peaks can arise from a statistical tradeoff between memorization and generalization. Our findings add precision to the study of duplication in language models, allowing practitioners to quantify the wasted compute incurred by the presence and repeat structure of duplicates in pretraining corpora.

cs.LG↗