Papers

Retrieval-Augmented Generation (RAG) is widely used to improve the performance of Large Language Models (LLMs) in answering user queries. Existing privacy research on RAG has focused on preventing unauthorized users from accessing sensitive data. However, another important problem that is often overlooked in RAG privacy research is that external generators have access to the query and the retrieved documents, which may contain confidential information that could potentially be misused or accessed for unintended purposes. In this paper, we introduce the Sensitive Entity Alias Generator (SEAG), a privacy-preserving framework that empowers users to utilize powerful third-party generators without disclosing sensitive information. SEAG introduces a lightweight model that locates sensitive entities, generates corresponding aliases, and constructs an entity replacement table. The table is used to replace sensitive words in the user's query and in the retrieved documents before they are forwarded to an external generator. For this purpose, two datasets were constructed: one for fine-tuning SEAG models to generate entity replacement tables, and another for evaluating the entire SEAG framework. The experimental results demonstrate the success of the SEAG framework. As for the User metric, which measures the ability of the model to provide a correct response to the user while hiding sensitive information from the external generator, all SEAG models achieved over 80% accuracy. Additional analysis further evaluated the ability of SEAG models Qwen-3, LLaMA-3.2, and Phi-4 to hide all sensitive entities within given documents. The results show good performance with total accuracies of 77.83%, 76.73%, and 74.91%, respectively.

Maintaining price consistency and executing an Every Day Low Price strategy is critical for global retailers. However, with catalogs spanning millions of active items, manual governance of price relationships is infeasible. Inconsistent pricing across item variants distorts customer value perception and cannibalizes sales. To address this, we present a scalable, context-aware Multi-Agent Framework designed to automate the construction of "Lines and Ladders" pricing taxonomies. Our framework employs specialized LLM agents to construct these coherent pricing structures by identifying key attributes, extracting multi-modal values, and applying hierarchical grouping logic. Evaluated on real-world enterprise data and deployed in production, our 3-Agent system achieves an F1-score of 0.83 for Lines, outperforming single-agent baselines by mitigating cognitive overload. The system achieves >90% precision and >75% recall in Food & Consumables, and 80.2% assignment accuracy in the unstructured General Merchandise catalog.

Multi-layer transformers form the critical component of essentially all large language models (LLMs) in use today. Because of their ubiquity and computational capability, there is a rapidly growing body of work that aims to precisely calibrate the expressive power of transformers as language recognizers by comparing them against standard models of computation studied for decades by the theoretical computer science community. In this endeavor, circuit complexity has by and large emerged as the "correct" branch of computational complexity to analyze the expressive power of transformers; the reason is that parameterizing transformers by the various resources they use, such as attention and precision, leads to direct comparisons with different classes of circuits parameterized by resources such as type of gates, size, and depth. Here, we present an overview of selected results that delineate the expressive power of transformers using concepts and methods from circuit complexity.

IT service management (ITSM) systems accumulate large volumes of heterogeneous ticket data that are difficult for sales and executive stakeholders to convert into actionable intelligence. This paper presents a sociotechnical AI pipeline, designed and evaluated following design science research principles, that transforms raw ITSM exports into a multilevel decision-support artifact. The pipeline combines LLM-based schema normalization, HDBSCAN sub-topic clustering, and hierarchical agglomerative clustering to generate executive-facing Main-topics and granular Sub-topics. A stakeholder evaluation across six artifacts and five raters from Sales Engineering and customer success roles shows that all four decision-support metrics, interpretability, actionability, trust, and likelihood of use, on average exceed 4.0 out of 5.0, with trust as the most consistent signal. The findings position ITSM analytics as an Information Systems (IS) problem of transformation, abstraction, and human-centered design.

Every interaction between AI agents begins with a handshake: the initiating agent must locate its counterpart, learn what it can do, and establish a session before the first useful message can be sent. Surveys of agent communication protocols name handshake overhead as an efficiency criterion, yet none report measurements. This paper presents the first empirical, phase-decomposed benchmark of the handshake in three major agent protocols: MCP (in both its 2025 initialize and 2026 discover eras), A2A, and ACP. Using each protocol's official SDK, we instrument every request for round trips, per-phase latency, wire bytes, and token cost across cold-start, warm-repeat, capability-mismatch, and degraded-metadata scenarios (N=50 per cell, on loopback and under an emulated 50 ms RTT). Round-trip count dominates handshake cost. A legacy MCP handshake takes three round trips and 230 ms at 50 ms RTT; A2A needs one round trip and 61.0 ms. Round trips are not the whole bill: when an LLM reads the counterpart's metadata, handshake cost scales with the counterpart's capability inventory at 24-107 tokens per capability, so a 200-capability counterpart spends 11% of a 200k-token context window before the agent reads its own task. MCP's own 2026 revision cuts its handshake from three round trips to two and, with era pinning, to zero on warm reconnects, protocol evolution that corroborates the cost we measure. We release the full harness, raw data, and analysis at https://github.com/pvarshh/handshake-benchmark.

The rapidly increasing student population has posed serious challenges to the traditional academic advising process. This study designs and implements a multi-purpose intelligent system to support students' academic progress, based on a two-part hybrid framework: (1) an advanced model for grade prediction and (2) a rule-based recommendation engine. Using a dataset containing 416,558 educational records from the University of Birjand, students were first divided into homogeneous clusters using the Gaussian Mixture Model (GMM). Subsequently, a Stacking Ensemble model combining Random Forest, Gradient Boosting, and MLP was trained specifically for each cluster. Evaluation results demonstrated that the Stacking model outperformed base models across all clusters, achieving a final aggregated RMSE of 2.35. The second component is an expert system that provides intelligent recommendations by synergizing educational regulations with the grades predicted by the first component. This system has been implemented as a practical tool on the University of Birjand portal, offering students real-time feedback such as semester GPA prediction, probation risk warnings, and course suggestions for GPA improvement.

We present the ABAI submission to COLIEE 2026 Task 1, case law retrieval, together with a controlled study of why it underperformed. The task suppresses the cited passages themselves, which removes much of the lexical overlap a retriever would rely on. Our pipeline answers this with four independently trained stages: multi-view BM25 over citation-context windows with reciprocal rank fusion, neural reranking, graph-based features from entity communities and a graph attention network, and a LightGBM meta-learner over 34 features. Our best run reached F1=0.177 on the official test set, against a cross-validated 0.311, and we attributed that gap to a recall ceiling, temporal distribution shift, and threshold miscalibration. We then tested all three. Under leakage-free protocols threshold transfer costs 0.007 F1, decision quality is flat across chronological quartiles, and the official test queries are not measurably farther from the training manifold than training queries are from each other, in two independent embedding spaces. Decomposing the misses instead splits them exactly evenly between candidates never retrieved and candidates retrieved but ranked below the cut. Measuring the remedies for each half, BM25 length-normalisation tuning, an event-triple view, and full-content dense fusion lift top-200 recall by three to seven points, and citation-graph features add 0.014 F1 over eight seeds once own-citation leakage is removed, while per-query cutoff rules, a zero-shot reranker swap, and a date filter do not help. We also document four evaluation artifacts, each of which reversed a result once the protocol was corrected.

Large Language Models (LLMs) are progressively used for advanced engineering tasks, includes Computer-Aided Design (CAD) documentation, standards compliance verification, and knowledge retrieval. Still, they are prone to produce hallucinations, outputs that seem convincing but aren't based on context that limit their trustworthiness in high-end engineering applications where precision and compliance are crucial. The paper introduces CoVe-RAG+, a unified framework that integrates Chain-of-Verification (CoVe) with Retrieval-Augmented Generation (RAG) to mitigate hallucinations in the results generated by large language models (LLMs). CoVe-RAG+ supports LLM verification in external sources of authority, such as engineering standards, CAD information, and simulation reports, while applying an iterative self-verification process to validate important claims. CoVe-RAG+ is assessed on engineering activities such as CAD model documentation, standards compliance verification, and the reutilization of historical design data. Experimental findings indicate a 28% improvement in factual accuracy relative to baseline CoVe and RAG methodologies. Moreover, CoVe-RAG+ strengthens user confidence by providing elucidative verification reports and source traceability. The findings indicate that CoVe-RAG+ provides a scalable and reliable option for implementing LLMs in engineering design processes where factual accuracy is critical.

Encoded-prompt attacks are evaluated almost entirely on their harmful arm: a benchmark sends obfuscated harmful requests and reports how often the model complied. A high refusal rate there is reported as safety, and it is equally consistent with a model that has stopped telling the request apart from anything else in the same format. We run the benign arm through the same transformation, and the two cases are far apart. Across four 7-8B models spanning three base families and four post-training recipes, refusal of harmful homoglyph-encoded prompts spans 0.08 while the same four span 0.57 on the identical requests in plaintext. What the encoding destroys is not refusal but the harm gap: on one model the gap between harmful and benign refusal falls from +0.82 in plaintext to exactly 0.00 under the encoding, and a benchmark reading only the harmful arm scores that model and one retaining a +0.61 gap identically. Running the cell such benchmarks leave out (plaintext content wearing the attack template, with nothing obfuscated) shows that on two of the four models the loss is caused by the protocol rather than by the character transformation, and on a third by the characters. Across a full SFT -> DPO -> RLVR pipeline the harm gap rises by +0.26 with a paired interval excluding zero while the standard harmful-arm metric registers no resolved change at all. We report twelve instrument defects, each with the control that caught it, including a binary jailbreak judge that fires on 0.61-0.70 of responses to plaintext benign prompts; six of the twelve inflate apparent safety, which is the direction a broken safety evaluation fails in by default.

We show that aligned vision-language models also condition refusal on a property of a request's form: whether an image is attached, holding everything the request asks fixed. Attaching a blank canvas, an image that cannot be read, cannot relate to the request, and is byte-identical across every prompt in its condition, shifts benign refusal by tens of points. The shift is not blanket caution but a threshold shift: genuinely neutral instructions are almost unaffected (<=2 percentage points on three of four hosted models) while borderline-benign prompts move +23 to +51 points, so the cost falls on sensitivity-adjacent traffic, meaning benign questions about privacy, self-harm, violence and illegal activity. There is a benign reading of such a threshold, namely that attachment correlates with risk in real traffic, and we take it seriously; a black-box study cannot measure that correlation and we do not claim to. What it can test is whether the response to attachment is robust to variation carrying no information about the request, and on four independent measurements it is not. It varies with canvas colour and pixel count. Its sign inverts across checkpoints. It survives an explicit instruction to disregard the image. And on one model it fires on a bare assertion that an attachment exists, with nothing attached and the modality word contributing none of it. Finally we price it. On a matched harmful set the same canvas does lower attack success, so the cue buys something. But the charge is decoupled from the purchase: the checkpoint with the least harmful headroom we measure, completing only 2% of plain harmful requests, still pays the benign cost in full, and across our models the harmful-side denominator falls as alignment improves while the benign cost does not track it down. Image presence is not a conservative default that a deployer chose and priced. It is an uncontrolled variable.

Large language models can process increasingly long prompts, yet their ability to locate and use decisive evidence may degrade as irrelevant or confusable context is added. We formulate this phenomenon, which we call context poisoning, as extreme-value interference in attention: the decisive-evidence score is upper-bounded, while the maximum score among effective distractors grows with their number. Under a softmax retrieval abstraction, we derive a finite-sample upper bound showing that maintaining a fixed accuracy target above base rate requires the evidence margin to scale as $\Omega(\sqrt{\log N})$, where N denotes the effective distractor count rather than necessarily the raw context length. The analysis connects long-context degradation to score aliasing, positional aliasing, and softmax dilution. Controlled experiments show that retrieval accuracy decreases as total context grows in the presence of embedded hard negatives, that the same-format condition produces the largest observed accuracy drop among the tested distractor constructions at fixed context length, and that retrieval gating can improve evidence use while its net benefit depends on preserving evidence recall. These results motivate evidence bottlenecks, alias-resistant representations, retrieve-then-reason architectures, verifier-mediated memory, and contrastive anti-poison training.

How many dimensions does a language model's computation actually use? The question is ill-posed until one names a functional. Task-weighted charts make it well-posed: low-dimensional coordinate systems fit against a chosen functional of the representation, under the functional's own metric, turning distillation into plain least squares. Across six models from three families, spanning 70m to 7B parameters, next-token prediction needs 70–90% of the residual stream's width to stay within 5% of intact perplexity, a width consumed by the rare tail of language, and the variance profile predicts none of it: two directions carry 90% of GPT-2's activation variance and almost none of its function. Dimension is per-functional: the model's own uncertainty reads from six coordinates where the full predictive distribution needs hundreds; and it grows with depth. The dissociation is exploitable: when only a few dimensions can be kept, charts trained under the functional's metric preserve the model's predictions better than variance-based or optimal linear compression.

A large literature incorporates physical structure into learned dynamics on the premise that models respecting the underlying physics should predict better. We test this premise using exact polynomial invariants recovered from trajectories and represented canonically as reduced Gröbner bases over Q. We find that exact polynomial invariants provide little additional benefit for residual-based downstream uses. On Gymnasium's Acrobot, consistency regularisation reduces algebraic residuals without improving long-horizon rollout fidelity, while one-step validation error is more strongly associated with long-horizon accuracy; similarly, shaping potentials recovered from systems with mass errors of up to 100% achieve similar learning performance to the physically correct potential. The main benefit observed here instead appears in diagnosis, where canonical representations permit exact ideal-equality and membership decisions, and where a reference generating set supplies a physical vocabulary that generic two-sample tests do not have. We develop a two-level diagnostic comprising screening, which identifies the violated physical constraint, and attribution, which recovers the faulty invariant and identifies the responsible physical parameter. Normal-form deflation removes algebraically trivial multiples, while quotient-space recovery avoids tolerance-based nullspace selection. Across fifteen injected faults, screening localises all fifteen when no observation noise is added, with no observed false alarms. Attribution identifies the responsible parameter in all seven single-parameter faults. Paired difference tests detect all faults but provide no constraint-level localisation, showing that the advantage is localisation and not detection. In conclusion, exactness is more useful for diagnosing violations of physical constraints than for improving prediction.

Reward-free latent world models plan by scoring candidate actions with distances in a frozen latent space: an action is preferred if its predicted future embedding lands closer to the goal embedding. This silently assumes that latent closeness is action-rankable, i.e., that ordering candidates by latent distance agrees with ordering them by true cost. We audit this assumption directly. We introduce ARC-Bench, a no-leak, fixed-candidate protocol that measures whether frozen JEPA-style objectives rank candidate actions correctly, and apply it to official released JEPA-WM checkpoints across navigation and manipulation-style control. The assumption fails, severely and structurally: on the official manipulation audits the top-scored candidate is almost always suboptimal, and the same inversion appears in the maze domains. A controlled visual-backbone extension shows that the defect persists when DINOv2 is replaced by video-pretrained V-JEPA 1 and V-JEPA 2 encoders at ViT-L/ViT-G scale. Provenance, undertraining, matched-budget backbone controls, and metric-circularity controls rule out trivial explanations. We then explain why this defect has stayed invisible: closed-loop replanning masks it. When we reduce the planner's replanning frequency, success collapses in both a navigation and a manipulation domain, and the episodes rescued by frequent replanning are enriched for severe first-plan ranking failures in the PointMaze first-plan diagnostic. Closed-loop success rates therefore systematically overstate the rankability of frozen latent representations. ARC-Bench supplies the measurement, and the masking mechanism the explanation, for methods that adapt, amortize, or replan around latent-space planners without directly auditing released JEPA-WM action rankability.

Rubric-based evaluations of deep-research (DR) systems often obscure fine-grained factual failures in generated reports. We introduce CLAIMPROBE, a claim-level audit that decomposes DR reports into claims and measures hallucination, misattribution, citation hygiene, and necessary-fact recall against retrieved evidence. Using CLAIMPROBE, we find that strong DR pipelines can omit key evidence and misattribute claims even when their rubric scores remain stable. We then propose CLAIMWRITER, a hierarchical claim-based writer that extracts source facts, maps them to a query-derived outline, and drafts each section from a source-linked claim representation. Across three prior DR frameworks, replacing only the report writer with CLAIMWRITER reduces hallucination by 2.6 to 4.5 times and improves necessary-fact recall by 1.2 to 1.7 times, while largely preserving overall report quality. CLAIMWRITER also enables localized revision: when sources change, it propagates changed source facts into revised reports at the highest rate among update methods, while also being more cost-effective.

Knowledge graphs used by agentic systems are often treated as flat stores of extracted triples, with little record of who owns a fact, why it was admitted, or how it should be used downstream. We argue that reliable agentic knowledge systems require governance as an essential component of graph construction to bridge this gap. We propose MAGG, a principled multi-agent framework for constructing Governed Knowledge Graphs that introduces explicit governance decisions for reliable and trustworthy knowledge sharing. A domain classifier first induces entity and relation types directly from document content, enabling operation in open-world settings without fixed schemas. Candidate triples are assigned to domain owners, reviewed against supporting evidence, admitted through governance decisions, and stored with audit metadata. The same ownership structure is reused during question answering, where queries are routed to domain-specific graph experts rather than answered through undifferentiated retrieval. Our evaluation demonstrates MAGG's effectiveness: On SciERC, MAGG improves strict triple F1 by 47% and mapped triple F1 by 51% over flat insertion. A blinded review of 120 triples finds governed-only triples more often source-supported than flat-only ones, and revised triples supported in 100% of cases. Finally, on MuSiQue, MAGG outperforms Microsoft GraphRAG by 9.0 exact-match points and 11.2 token-F1 points.

Embodied artificial intelligence (AI) must be tested in the clinical environments where it will operate, but building realistic, robot-testable settings is costly and difficult to scale. Here we show that routine clinic images can be transformed into operational digital twins for task-based evaluation of embodied AI. Using 39 ophthalmic clinic scenes, we converted single photographs into editable, simulator-ready environments and assessed reconstruction quality, room-scale geometry, mesh grounding, multi-robot feasibility, perturbation sensitivity and closed-loop policy performance. The reconstructed scenes preserved workspace structure, while local editing enabled controlled device reconfiguration. Device meshes, collision proxies and semantic anchors converted visual reconstructions into contact-aware simulation scenes. Across three robot embodiments, shared task targets showed different patterns of reachability and contact feasibility. Small device translations and rotations produced task-specific changes in contact margins that were not captured by visual similarity alone. Digital-twin trajectories also supported local policy learning and closed-loop evaluation. These findings establish operational validity as a key principle for clinical digital twins and provide an intermediate layer between offline development and physical deployment of embodied AI in healthcare.

Large Vision-Language Models (LVLMs) have achieved remarkable performance across a wide range of tasks; however, they often inherit social biases from their training data, resulting in biased behavior when processing portraits from different social groups. Existing debiasing approaches typically compare token probabilities between the original and biased generations during decoding, but they are fundamentally limited by their reliance on a single, stereotyped viewpoint and fail to account for the diversity of social perspectives. Inspired by the social science principle that diversity fosters fairness, we propose Counterfactual Ensemble Decoding (CED), a novel framework that constructs multi-group counterfactual perspectives within the visual representation space and integrates them during decoding to promote equitable model behavior. CED first performs counterfactual steering in the visual space by identifying semantic directions associated with each social group and generating counterfactual representations along these directions, thereby offering diverse perspectives that disrupt stereotypical narratives. During decoding, CED locates the decoder layer exhibiting the greatest divergence among these perspectives and ensembles their token distributions using uncertainty-aware weights, prioritizing high-confidence tokens from different groups to yield a more balanced probability distribution that guides fairer generation. Extensive experiments on three social bias evaluation benchmarks demonstrate that \tool achieves substantial improvements over leading baselines, reducing bias by up to 47.97% across scenarios involving occupations, descriptors, and persona traits. Moreover, CED also preserves the core capabilities of the original model with minimal degradation.

On-policy distillation (OPD) has emerged as an effective paradigm for transferring knowledge between language models, where a student is trained to align its next-token distribution with the teacher's along its own trajectories. To provide dense supervision at tractable cost, many works minimize the reverse Kullback-Leibler (KL) divergence between the student and teacher's normalized distributions over the teacher's top-$k$ tokens. However, this normalized objective discards the information about tail probability: the total probability outside the teacher's top-$k$ tokens. As a result, the optimization can steadily increase the student's tail probability and entropy, empirically degrading downstream accuracy. To address this issue, we propose Tail-Aware Top-$k$ OPD (TA-OPD), a novel distillation method that restores the missing tail probability signal. In particular, TA-OPD minimizes the reverse KL divergence over the top-$k$ tokens plus a tail token that carries the tail probability. In effect, TA-OPD better aligns the student's next-token distribution with the teacher's, preventing the increase in tail probability and entropy caused by top-$k$ normalization. Extensive experiments demonstrate the superiority of TA-OPD, improving Avg@8 by up to 8.05 points on common benchmarks. Our code is available at https://github.com/HuipengHuang/TA-OPD.

The rapid advancement of intelligent transportation systems and autonomous driving relies heavily on multi-modal urban traffic datasets. However, curating high-fidelity video imagery in complex tropical urban environments—specifically Kuala Lumpur, Malaysia—presents severe challenges for Personally Identifiable Information (PII) anonymization due to high motorcycle density, dark acrylic license plates, dynamic camera tilt, and extreme tropical glare. We propose an automated anonymization framework tailored for the Kuala Lumpur Road Dataset, captured via a mobile cycling platform at 2 FPS. We document how legacy Haar cascades and YOLOv8 fail under these conditions—generating false positives on background elements while missing rotated or occluded targets. Our architecture resolves this by integrating Grounding DINO—a zero-shot open-set vision-language transformer—with a novel Spatial Vehicle Region of Interest (ROI) Containment Engine. By requiring license plate centroids to reside within validated vehicle boundaries, the pipeline suppresses environmental false positives while automatically obfuscating faces, heads, and license plates. An initial evaluation on 1,266 frames demonstrates a $\sim$95% success rate, with remaining failures restricted to small, heavily occluded, oblique, or ambiguous targets. Coupled with temporal persistence mechanisms and an automated quality-control auditor, the framework minimizes privacy-related false negatives while preserving scene context for downstream vision tasks. While formal legal compliance depends on broader governance procedures, this publicly available pipeline and demonstration notebook provide an auditable preprocessing stage for privacy-aware dataset curation.

Following the molecular discovery and synthesis revolutions, scalable automated structure elucidation from routine spectroscopic data remains an outstanding challenge. Despite decades of computational efforts, no existing system achieved reliable reasoning over unseen spectra. Here, we propose MACROS, a multi-agent system automating structure elucidation by emulating expert iterative hypothesis-testing. Trained on 100M simulated and 1.6M experimental spectra-molecule pairs, it natively supports arbitrary combinations of routine spectroscopic techniques. It achieves unprecedented zero-shot generalization to diverse real-world samples, correctly identifying synthetic compounds, natural products and metabolites above 500 Da with 1D NMR. Remarkably, MACROS spontaneously recovers textbook spectroscopic correlations from unassigned data and exhibits emergent chemical intuition such as a ring-first parsing preference, learning fundamental chemical principles rather than memorizing database patterns. MACROS augments chemists via collaboration to deliver sixfold faster, 40% more accurate elucidation. MACROS establishes a scalable foundation for fully automated structure elucidation, and catalyzes accelerated molecular discovery toward autonomous laboratories.

Multiple instance learning (MIL) is widely used for weakly supervised whole slide image (WSI) analysis. However, under long-tailed distributions, MIL-based WSI analysis faces a nested dual long-tail: an inter-slide class long tail and an intra-slide long tail of instance-level discriminative evidence. The two long tails are coupled: tail classes have few training slides, while their limited diagnostic evidence is concentrated in a few patches and obscured by abundant within-bag redundancy. This coupling biases models toward head classes and degrades rare-class recognition. To address this, we propose DeCo-MIL for long-tailed WSI analysis, which jointly alleviates the nested dual long-tail through frequency-debiased counterfactual reasoning. For the inner long tail, DeCo-MIL clusters patches into tissue-morphology anchors, replaces each anchor with its matched normal prototype to perform a counterfactual intervention, and estimates its counterfactual contribution to the ground-truth class using class-frequency-corrected predictions. These contributions guide redundancy masking to preserve scarce discriminative instances. For the outer long tail, DeCo-MIL constructs anchor-stratified pseudo-bags from redundancy-reduced bags and combines tail-aware oversampling with consistency regularization, increasing effective supervision for tail classes while preserving tissue-morphology composition. Extensive experiments on three long-tailed WSI benchmarks demonstrate that DeCo-MIL achieves state-of-the-art performance in both tail-class recognition and overall classification.

Clinical algorithms can perform differently across patient groups, a concern that grows as these algorithms increasingly take the form of artificial intelligence (AI) models. In the United States, a 2024 federal rule bars healthcare providers that receive federal funds from discriminating through their use of clinical algorithms, which exposes them to liability when such use disadvantages a protected patient group. We examine how such liability considerations reshape (i) an AI firm's model design decisions that drive group-specific accuracy and (ii) a physician's decisions to use AI in healthcare delivery. The firm designs an AI model for two patient groups, and improving accuracy for the disadvantaged group is more costly. The physician (who remains the accountable decision maker) then decides whether to consult AI, weighing the reduction in clinical uncertainty against expected liability exposure when AI errors disproportionately affect the disadvantaged group. We find that this liability can induce disparate use of AI: the physician may reduce AI use overall and, over an intermediate range of liability, rely on AI less for disadvantaged patients. The effect is non-monotone. As liability increases, the physician's use of AI for disadvantaged patients first declines, then rises as the firm reallocates investment toward reducing disparity or switches to an equal-accuracy design. Mandating equal algorithmic accuracy across patient groups can then inadvertently harm both groups, because a uniform accuracy requirement distorts the firm's investment incentives and the physician's equilibrium AI-use decisions.

Verifying whether clinical care follows evidence-based protocols is a natural neuro-symbolic problem, yet the safety-critical setting defeats either paradigm alone. We present an expert-guided pipeline that constrains a large language model strictly to semantic normalization, mapping messy drug and microbiology strings onto a fixed clinical vocabulary, while a Sugeno fuzzy inference system reasons over the normalized events. The fuzzy layer encodes eight Surviving Sepsis Campaign bundle rules and replaces binary judgments with graded scores in [0,1]. Applied to 2,438 MIMIC-IV v3.1 sepsis episodes, it surfaces antibiotic timing as the most critical breakdown (mean 0.24, 13% within one hour), Hour-1 underperformance (mean 36.7%), a 51% elevated-lactate drop-off, and descriptive differences in ICU stay across compliance groups (3.8 versus 5.1 days).

Natural-language interfaces to enterprise data must translate underspecified requests into governed, executable behavior while controlling invalid queries, policy failures, cost, and nondeterminism. SemPlan Benchmark evaluates this architectural design space with a deterministic synthetic bilingual benchmark containing 1,800 cases in English and Brazilian Portuguese; 1,200 cases form the frozen scientific evaluation subset. Four architectures are compared under the same model configuration: direct SQL generation (A1), a bounded tool-agent baseline (A2), structured semantic-request generation followed by deterministic planning and execution (A3), and a clarification/stateful semantic-plan variant (A4). Across 4,800 primary records, answer correctness was low in absolute terms: 22.25% for A1, 22.58% for A2, 25.67% for A3, and 24.25% for A4. A3 had the highest observed correctness and significantly exceeded A1, A2, and A4 in the pre-specified paired correctness analysis, while A1 retained the highest policy-correct rate and the lowest unsafe-or-invalid rate. A4 had the lowest mean API cost and lowest false-refusal rate. On a preselected 150-case stability subset, answer-correct repeatability ranged from 92.00% to 98.67%. The results support a trade-off interpretation rather than a universal ranking: additional structural constraints changed failure modes and efficiency, but did not monotonically improve correctness or solve ambiguity and multi-turn state consistency.

Probabilities of causation (PoCs) characterize individual causal responses that cannot be directly observed and therefore generally require partial identification. Tian and Pearl first derived theoretically sharp bounds for binary PoCs, including the probability of necessity (PN), the probability of sufficiency (PS), and the probability of necessity and sufficiency (PNS). Mueller et al. subsequently tightened the bounds for binary PNS by incorporating causal information encoded in covariates and mediators. More recently, Li and Pearl, as well as Shu et al., extended PoCs to multivalued settings and derived corresponding theoretical bounds. These developments naturally raise the question of whether additional causal knowledge can further tighten the bounds in multivalued settings. This paper addresses this question by deriving tighter bounds for multivalued PoCs through the incorporation of causal information encoded in covariates and mediators. We illustrate the theoretical results with toy examples, while simulation studies further demonstrate that the proposed bounds are tighter than existing nonbinary bounds.

Long-running LLM agents act through tools, and a single step can send an email, merge a pull request, or wire a payment. The steering decision is the pre-commit choice at that boundary: proceed, or hold for human or policy review. We introduce SteerBench-Work, an incident-anchored, bidirectional benchmark for that decision in workplace agents across developer operations, customer service, finance, legal, medical, HR, and security. Release v2026-05 contains 106 scenarios anchored in public incidents, paired evidence-reversed mirrors, and calibration controls, with labels split nearly evenly between proceed and hold so the two error directions get near-identical numbers of chances. A model sees the proposed action and the available evidence, returns a gate decision, and is scored on whether it crosses or holds the boundary correctly. Across 30 model conditions the failures run almost entirely in one direction: models wrongly hold authorized, evidence-cleared work on 28.1% of opportunities and wrongly allow unsafe work on 1.0%. The hardest cases are risk-resolved commits, where signed or structured evidence has already cleared a real risk trigger, and models score markedly worse on evidence-reversed mirrors of famous incidents (63.8%) than on the incidents themselves (98.5%). General capability is not the same as steering calibration: higher-capability models often over-refuse at the commit boundary, and more reasoning can repair a weak gate while leaving a calibrated one flat. The public leaderboard is at steerbench.com.

LLM judges have become central infrastructure for model evaluations, online grading, and reward modeling. Judges are typically validated by accuracy on golden data, but accuracy says little about whether they are stable under re-prompting, challenge, or sustained pushback. We introduce the Wiggle Framework, a unified stress test for epistemic stability in LLM judges. The framework decomposes judge robustness along three dimensions: Mechanical Consistency (stability under re-prompting and reframing), Single-turn Conviction (stability under a single challenge), and Multi-turn Persistence (stability under sustained or adaptive pressure). We use the framework to study 9 frontier models across 14 judging tasks spanning safety, toxicity, AI writing detection, and political-response evaluation. Every model exhibits substantial wiggle as a judge — flipping verdicts 25–71% of the time under static pushback, and 62–91% with an adversarial LLM persuader. Critically, we find that pressure that succeeds in changing a judge's verdict is almost always net-corrupting with respect to ground truth. Beyond the framework itself, we identify baseline jury majority strength as the most effective single-shot signal for anticipating which items wiggle. Taken together, this is the first apples-to-apples cross-dataset comparison of mechanical, conformity, and persuadability tests in a judging context.

Causal discovery aims to uncover the underlying causal relationships given data generated from a system. The goal, however, is not merely to predict causal edges given data, but also to be able to interpret and explain either observed or hypothesized phenomena, such as a particularly large causal effect. We consider this task of conditional causal discovery and cast it as a Bayesian inference problem, in which we target the posterior over causal graphs and parameters conditional on an event such as a causal-effect constraint. Unfortunately, this poses a computational challenge: existing approaches to Bayesian causal discovery struggle when the event has small posterior mass. To address this, we adapt rare-event estimation techniques to perform inference the joint graph-parameter space. Our method gradually drives a particle population toward the constrained region while maintaining samples that approximate the conditional posterior. Empirical evaluation on synthetic graphs validates the accuracy of our approach at small and large scales, and we show in a case study on the Sachs protein dataset how our method can be used to aid scientific exploration by providing pathway-level summaries.

While large language models can generate entire novels, there is little information about the level of formal variation in their output over many generations. Rather than asking whether individual passages can be identified as AI-generated, this study asks whether repeated AI generation can produce the same range of diversity which is found across human corpora. This paper contrasts six corpora based on generation source and target style: twenty novels generated using GPT-5.5 Thinking in a nineteenth-century British realist style, twenty novels generated using Qwen3-14B in a nineteenth-century British realist style, twenty novels generated using each of these models in a contemporary zero style, 205 nineteenth-century human-written British novels, and sixty-five contemporary human-written Zero-Style novels. At the document level, the research includes MATTR-500, Shannon entropy, average sentence length, readability, and punctuation rate measurements. The most robust and reliable result is compression of sentence structure. Repeated generations produce novels that vary far less from one another in sentence structure than human novels do. Compression is also present in the measures of readability, punctuation, and sentence length variability within novels. Lexical measures tend to be similarly compressed, with the exception of Qwen Zero-Style MATTR. Despite having distinct mean stylistic profiles, GPT and Qwen lack a stable pattern of cross-measure correlation. This article therefore distinguishes between variance overclosure, which represents a limited formal range between novels, and a more specific phenomenon of correlational overclosure. This means that an individual AI-generated novel may resemble human fiction stylistically, while a collection of AI-generated novels occupies a much narrower formal range.