Papers

Optical satellite image correlation measures wide-area deformation in regimes where coherent interferometric synthetic aperture radar fails because displacement gradients are too large. However, standard pairwise workflows lack a pre-acquisition error budget and rely on extensive stable terrain. We formulate repeat-pass optical correlation as a differential estimation problem without surveyed ground control. Nominal georeferencing defines the coordinate frame, stable-area constraints and displacement priors resolve the datum, and surface displacement is estimated jointly with inter-epoch revisit-bias coefficients. The model yields a predictive accuracy envelope and calibrated per-point posterior uncertainty, bounds along-track uncertainty through a displacement prior, and represents pushbroom jitter using per-line revisit offsets. Simulations and Sentinel-2 and WorldView-2 experiments on the 2019 Ridgecrest earthquake, the 2023 Kahramanmaraş earthquake, and the Baltoro glacier validate the predicted noise floor, control-free accuracy margin, and leakage caused by view-angle and digital elevation model errors. The measured noise floor reaches approximately $0.05$ pixel at $10$,m ground sampling distance. With only five stable tiles, conventional destriping changes the estimated Baltoro trunk velocity from $106$ to $1251$myr$^{-1}$, whereas the prior-constrained estimate remains $87$myr$^{-1}$. Closure analysis attributes approximately $88\%$ of pair-error variance to individual scenes, consistent with $25{,}354$ ITS_LIVE glacier-velocity triplets. Three matching methods lead to the same conclusions. The framework therefore turns pairwise correlation into a robust measurement with a predictive error budget, reduced dependence on stable terrain, and conclusions independent of the matching method.

Post-training pipelines must select one language-model policy from many checkpoints, prompts, and decoding rules. Mean evaluator scores can conceal rare failures, whereas simultaneous candidate-wise confidence bounds can be unnecessarily conservative. We introduce StepCOPS, which uses an independent proposal split to nominate one lower-tail floor per candidate, exact binomial tests on a fresh certification split, and Holm's step-down procedure to certify a set of floors. With probability at least $1-\delta$, every certified floor, including the largest floor used for policy selection, is below its candidate's population lower $\alpha$-quantile. This guarantee assumes i.i.d. evaluation units while allowing arbitrary within-unit dependence across candidates. Across 24 predeclared configurations and 11 benchmarks, StepCOPS obtains 96.4% selected-policy coverage over 500 paired trials, raises the certified floor by 1.5 points over both proposal-Bonferroni and exact COPS, remains 0.6 points below the large-reference jury oracle, and abstains in 2.4% of trials. Shadow-judge, benchmark-native, artifact, and leave-one-judge-out audits characterize the proxy boundary: the guarantee applies to the fixed jury score, not directly to human safety.

Language-model advice can accelerate reinforcement learning, but every call has a price, and a returned suggestion may be unparsable, unsafe to execute, or no better than the agent's own action. We introduce certified predictive advice gating, a cost-aware controller that forecasts possible advisor responses, values the action actually executed after parsing, execution checks, and cache reuse, and requests advice only when a calibrated lower bound on that value covers the price. Our analysis separates the gap of executed advice, cumulative regret under an intervention-stable fallback, and myopic acquisition loss. On BabyAI GoToObj with Qwen2.5-1.5B and 7B advisors, the gate improves return over no advice by paired $0.029\pm0.016$ and $0.030\pm0.015$ while using more than $97\%$ fewer calls than always asking. Matched-budget comparisons, however, do not resolve an advantage over an early query schedule, so the gain reflects economical advice volume rather than demonstrated state selection. Groupwise calibration raises coverage on positive-value opportunities from $0.47$ to $0.85$, still below the $0.90$ target. A separate headroom-aware gate reports paired gains over an early schedule of $0.032\pm0.012$ and $0.029\pm0.011$ on GoToObj and $0.027\pm0.011$ and $0.028\pm0.012$ on GoToObjDoor, and adds BabyAI and ALFWorld success, decision-stratum coverage, and component timing. These results separate return gains, query placement, and execution screening; finite-sample deployment guarantees remain unestablished.

Large language models (LLMs) are increasingly deployed as multi-turn agents that must sustain goals, use tools, and adapt to other agents over extended interactions. However, existing research lacks auditable, multi-turn, multi-factorial experiments that quantify LLM behavior under explicit constraints, with time-resolved statistics that reveal how behavior unfolds over long horizons. To address this gap, we develop a multi-agent micro-benchmark inspired by the Stanford marshmallow experiment: ReAct agents operate minute-by-minute with a "raise a question" tool under a per-step budget, while we factorially manipulate social context (broadcast vs. isolated), personas (age, hedonic drive), and metacognitive policy (mandatory vs. optional tool use). We analyze outcomes with Kaplan-Meier (KM) survival curves and discrete-time hazard models over a long risk horizon across 19,200 agent trajectories in 64 cells. Behavior shows a sharp early "eat" impulse, and only 75.9% of agents persist to the end. In a discrete-time hazard model, isolation reduces per-minute risk relative to broadcast, whereas a must-use self-questioning policy increases risk. On average, agents ask $\approx 7.12$ questions and hit the per-step budget in $\approx 6\%$ of minutes. Questioning declines faster under broadcast than isolation. Ablation experiments demonstrated that removing hedonic drive and/or persona age increases survival and completion, narrows the broadcast/isolated gap, but leaves the must vs. may ordering intact. The combined ablation (no hedonic + no persona age) yields the highest completion (approaching $1.0$). These results establish delay-of-gratification as a compact, multi-turn interaction benchmark that captures social contagion and tool-use dynamics in LLM agents, providing a reproducible testbed and statistics for analyzing long-horizon, multi-agent behavior.

Large language model (LLM) agents may perform well on isolated tasks yet drift into inconsistency over extended interaction. We evaluate temporal consistency in a controlled 20-step multi-agent setting inspired by delayed-gratification studies. At each step, an agent chooses between continuing to delay a reward or claiming it immediately (terminating the episode). Across a full-factorial manipulation of social visibility (private vs public), persona stressors, and deliberation policy, we run 84,540 trajectories spanning 8 model families. Treating the first reward-claim as a time-to-event outcome, we estimate Kaplan-Meier survival curves and fit discrete-time hazard regression to quantify how experimental factors shift failure risk over time. Then, to analyze rationales and language patterns associated with failure, we build a seven-category taxonomy from 13,780 deliberation traces from agents who choose to terminate the episode, using an LLM-assisted labeling paired with human audit ($\kappa=0.83$). Rationale profiles change systematically with time and context: early failures are more impulse-driven, later failures more fatigue- and cost-benefit-framed, while public settings increase norm-oriented justifications. We also find a deliberation-inconsistency association: among failures, longer deliberation correlates with higher rates of intra-rationale contradiction (simultaneous pro-delay and pro-claim statements), challenging the assumption that more reasoning text implies greater consistency. Together, the survival and rationale analyses reveal distinct temporal reliability regimes and model-specific "failure fingerprints", offering an evaluation lens for diagnosing inconsistency in multi-turn agent behavior.

Natural language explanation generation serves as a key mechanism for exposing and evaluating vision-language reasoning. Prior work on explanation-driven vision-language models predominantly follows a post-hoc (answer-first) paradigm, implicitly suggesting that supervised rationales can reflect underlying reasoning processes. In contrast, modern large vision-language models increasingly exhibit a rationale-first generation tendency, which more closely aligns with structured, stepwise reasoning. In this work, we systematically evaluate whether explanations are causally tied to model predictions within a single generation step under a controlled experimental setup, explicitly eliminating unnecessary chain-of-thought or other intermediate reasoning processes across knowledge-intensive QA, visual entailment, and compositional grounding benchmarks. We find that larger models emerge as a prerequisite for reliably supporting rationale-first reasoning at scale. However, answer-first generation is less prone to format-related errors in structured output. Overall, explanation ordering, model scale and pre-training knowledge, task-specific fine-tuning, and task structure jointly influence both prediction accuracy and reasoning faithfulness.

Knowledge graph completion models optimize ranking, although many downstream applications require calibrated probabilities. We present BLADE, a variational model that separates latent truth from graph recording and distills offline language-model judgments into a frozen teacher regularizer. The LLM is absent during inference. Posterior samples provide predictive probabilities and epistemic uncertainty, while the compact teacher remains available only as an optional triage factor. Across five benchmarks, BLADE remains competitive under a common ranking protocol and reduces adaptive ECE by a macro-average of 60.1% relative to deep ensembles and 78.1% relative to temperature-scaled RotatE. On identical FB15k-237 candidate sets, BLADE also improves ECE, Brier score, and NLL over validation-selected histogram binning and a matched generative ComplEx2 model, with these improvements persisting on a prespecified near-miss pool. Under controlled injected missingness, the full triage score achieves a mean AUC-PR of 0.863, compared with 0.805 for its strongest non-teacher variant. Leakage stress tests show that aligned semantics matter, but they cannot exclude knowledge acquired during LLM pretraining. We therefore claim calibration only for the declared candidate distributions, not for all unobserved triples.

We study hidden-target seeking in an anchor-free regime where neither the vehicle, the relay, nor the target has an accessible global pose. The vehicle never senses the target directly; it receives only range-bearing observations through a single relay of unknown position and orientation, with motion known only via integrated body-frame odometry. Absolute localization is fundamentally impossible: the joint configuration retains an exact three-dimensional SE(2) gauge no estimator can resolve. Yet the quantities needed for control remain fully recoverable: motion collapses calibration to a task-relevant quotient, the relay-to-odometry yaw and target displacement, identifiable in closed form from two distinct vehicle views. We introduce an O(K) multi-view self-calibration estimator with an exact first-order yaw-uncertainty certificate that propagates the cross-view correlations integrated odometry induces: the full certificate attains 95.0% pooled coverage at the nominal 95% level, versus 82.1% when correlated poses are treated as independent. The certificate drives a hybrid policy that excites until calibration is trustworthy, refuses uncertified estimates, seeks using continuously re-measured geometry, and detects relay-frame changes via persistent certified inconsistency, avoiding unbounded dead-reckoning drift. Across 200 randomized closed-loop trials, the method attains 0.064 m median station error versus 0.065 m for an oracle given the true relay yaw, despite 27 m median dead-reckoning drift over long horizons; 90 physics-based ROS 2/Gazebo trials retain 30/30 success under nominal operation, communication degradation, and relay-frame disturbances. These results show global localization is unnecessary for reliable hidden-target seeking under this single-relay model, even when both the sensing infrastructure and the vehicle's own reference frame are uncalibrated.

Large language models (LLMs) are increasingly deployed in multi-agent systems where a principal agent decomposes tasks and delegates them to subordinate agents that may invoke external tools. Safety alignment, however, is still evaluated almost exclusively under a single-agent threat model, treating safety as a property of the individual LLM. We show that this assumption breaks down: individual safety alignment fails to transfer to multi-agent settings. Two failure mechanisms emerge under delegation: responsibility diffusion on the principal side and role-bias compliance on the subordinate side, jointly converting language-level refusal into actionable harm. We refer to this phenomenon as delegated misalignment and study it through a three-condition protocol across 6 frontier LLMs on 49 hazardous tasks. Delegation amplifies end-to-end harm substantially: DeepSeek-V3.2's full-execution rate rises from 30.6% to 77.6% once delegation is introduced, and the same model behaves very differently across roles (GPT-5: 22.5% as a single agent vs.\ 61.2% as a subordinate). Ablations further show that standard single-layer defenses each fail on their own and can even backfire. We call on the community to move beyond per-model alignment and toward composite safety mechanisms before multi-agent LLM systems are deployed at scale.

Comparative studies of predictive models often end by ranking candidate models, yet these rankings depend on evaluation protocols whose influence is rarely treated as a source of uncertainty. We formalize this problem as protocol-induced ranking uncertainty and introduce a framework that compares ranking displacement caused by switching protocols with displacement produced by conventional choices within a fixed protocol. We quantify these effects using the Protocol Sensitivity Score (PSS) and a full-refit intraprotocol reference. We validate the framework in a large-scale sequential prediction study of daily PM10. Static-split and rolling-origin evaluation are compared across 425 European background stations and 365 US EPA monitors. Switching protocols produces mean PSS values of 0.801 and 0.772 and changes the selected model at 35.3% and 31.5% of stations, respectively. In Europe, intraprotocol perturbations with identical scored targets produce PSS values of 0.072 and 0.230, with winner-swap rates of 0.8% and 4.9%. Between-protocol displacement is therefore substantially larger than the selected within-protocol references. Expanding the candidate set from three to nine models increases the between-protocol winner-swap rate to 60.2% in Europe. The pattern also persists under a frozen protocol applied to held-out background and non-background stations. These results show that model-selection conclusions can depend materially on legitimate evaluation choices. We recommend reporting ranking stability under a small set of defensible intraprotocol perturbations alongside claims of model superiority.

We present CLAIM, a production pipeline for K-12 assessment-item generation coupling a two-stage generate-then-attack protocol (the model drafts as a "curriculum architect", then re-enters the same conversation as a hostile adversarial reviewer), bi-directional few-shot conditioning on accepted and rejected items, the latter carrying the evaluator's diagnosis, and a knowledge dictionary of 44,844 error-correction rules mined from that feedback and retrieved per standard and item type. Across 43,227 scored items over 755 Common Core ELA standards, three item types, and ten LLMs, the pipeline reaches a 97.8% expert-evaluator pass rate on a 9,074-item production run. We then ask what that rate certifies. Re-scoring a stratified sample with three judges from other vendors, blind to the deployed verdict, reproduces the format ordering under every judge and recovers a larger open-set deficit than the deployed evaluator does; but agreement on the accept/reject binary is weak at production prevalence (kappa about 0.13), and the judges agree with each other no better. The level is therefore judge-relative, and with no student-response data our quality evidence is evaluator-judged throughout. The corpus also exposes a robust asymmetry. Multiple-choice and multiple-select generation saturate at 98% or above for both frontier models under a dozen static rules, whereas fill-in-the-blank generation is capability-tiered (82.8-96.7% across five models under a matched rule set, standards, and judge) and plateaus under prompt-only optimization, with error mass shifting between answer-key over-inclusion and omission as rules accumulate. We analyze this as open-set boundary determination, a task autoregressive decoders are structurally ill-equipped to solve, and show the asymmetry recurring when the evaluator itself is distilled: fail-recall rises from 8% to 63% while F1 saturates at 0.25.

Entropy over chain-of-thought tokens decides which tokens receive the policy gradient, which get pruned, and whether a run has collapsed, yet each such statistic reads a next-token distribution mixing three choices: whether to emit connective scaffolding, which connective, and what the substantive continuation should be. Designating a scaffold vocabulary subset separates the three, exactly, for entropy, Kullback–Leibler divergence, and the first-order entropy velocity of a softmax policy. We prove the raw and content conventions disagree about which position is the larger fork on an explicit open region, and bound answer diversity by the content channel plus a leakage term a measured witness certifies. Across twenty-three configurations the scaffold side carries up to 41% of the raw high-entropy set; on a matched-tokenizer ladder, coupling changes only at the math-corpus step while the scaffold's entropy share keeps growing through distillation; a closed-form forecast from one channel correlation tracks selection retention over a 54-point range to five points, unfitted. On compression, the content convention beats raw surprisal in every cell; an answer-leakage audit then corrects our own headline control: re-fed chains earn a quarter to a half of their accuracy from restated answers, and once stripped, no token scorer beats a random contiguous block.

Precision healthcare, particularly for conditions like hypertension and cardiovascular disease, necessitates monitoring of dietary sodium intake. However, tracking this is hindered by the prevalence of hidden salt in cooking, such as sodium in soy sauce and ketchup. While recipes offer a valuable data source for dietary analysis, sodium-rich seasonings are frequently omitted or described ambiguously in instructions. To solve this issue, we propose SALT, a Sodium Assessing & Level Tracking framework adopting an RAG framework to assess sodium content in recipes. Our framework first introduces a Joint-Factor Embedding Retrieval module to locate similar recipes with specified sodium content for addressing the lack of contextual references. These retrieved samples provide contexts for subsequent inference. Then we design a structured 4-hop Chain-of-Thought inference module to refine the vague estimation from language models through a multi-step sodium estimation. To facilitate our study, we further construct a recipe dataset SALT54k with $54,151$ entries labeled with sodium quantities across $11$ common seasonings. Results on SALT54k demonstrate that our method achieves state-of-the-art performance in sodium estimation. Additional real-world validations confirm the effectiveness of our method, demonstrating its potential as a practical solution for AI-assisted precision healthcare.

Language models are frequently employed in settings where they are asked to make value judgments and choices. These observed choices often exhibit intransitivity: A model may prefer item $A$ to $B$ and $B$ to $C$, while also preferring $C$ to $A$. Existing work that models LLM preferences treats such inconsistencies as sampling noise around a single latent ordering. We instead propose that intransitivity reflects the aggregation of multiple latent, internally consistent orderings. We first show that observed inconsistencies cannot be explained by a single ordering under any monotone link function. We then introduce a noise-augmented mixture Bradley-Terry (MBT) model that infers latent preference components from repeated pairwise comparisons. Across seven models and four tasks, a mixture of orderings often explains structural inconsistencies better than single-utility models. We find that aggregate preferences often hide underlying preference heterogeneity. A case study on Moral Machine dilemmas shows that models which disagree on aggregate orderings can still share latent components. Together, these results suggest that LLMs reflect plural preferences. Alignment and evaluation pipelines that treat LLM preferences as a single function, therefore, risk averaging over coherent orderings that different users may endorse differently.

Employers are increasingly using large language models (LLMs) to automate their hiring process. This paper investigates the risk of monocultural biases, in which the widespread deployment of large language models homogenizes biases across the labor market, leading to greater systemic exclusion for certain demographic groups. For ten LLMs, we measure hiring biases across their base and post-trained versions to identify which stage, pre-training or post-training, lead to monocultural biases. We find that, compared to their base models, post-trained models are 3.6% less likely to callback older applicants. This negative shift occurs in eight of the ten models that we evaluate. Post-trained models have much more correlated decisions than base models which is likely driven by human capital traits like skills or college major. However, greater consensus among models increases global systemic exclusion rates from 5.6% to 17.3% and exacerbates demographic inequalities, with intersectional systemic exclusion rates ranging from 12.2% to 21.7% for post-trained models. We find that this inequality is primarily driven by age-based discrimination that is exacerbated in post-training. These results indicate that while post-training techniques may improve models' abilities to select the best applicants, they may raise systemic inequality risks for those at the margin by uniformly introducing new biases.

The aged population is exposed to risk situations, like falls, sudden changes in vital signs, and faints. These situations become common at this life stage due to a decrease in body capacity and illnesses increase. This scenario imposes attention on families and the national health system. Some studies propose to overcome this problem through elderly health monitoring systems based on smartphones, smartwatches, and other equipment that usually cover low-efficient communication systems and low battery life. This paper proposes an efficient health monitoring system based on Sigfox IoT service and a watch-type wearable device for sensing elderly motion, location, and vital signs to predict risk situations designed to achieve at least 30-days of battery life. Our experimental results show that this architecture is a low-cost, low-power, efficient, and feasible elderly health monitoring solution.

Predicting reaction yield from molecular structure and reaction context can cut experimental trial-and-error and speed up condition screening in synthetic chemistry. Recent methods for this task use learned representations such as graph neural networks or Transformer encoders over reaction SMILES (Simplified Molecular Input Line Entry System), but these approaches carry heavy preprocessing overhead and can break when input formatting is inconsistent. We propose MFP, a reaction yield prediction method built on role-aware Morgan fingerprints where count-based circular fingerprints are computed for each reaction component, aggregated by chemical role (reactant, reagent, product), and combined with transformation-sensitive difference features into a fixed-length reaction descriptor fed to a feed-forward neural regressor. We test MFP against state of the art methods such as YieldBERT (with and without data augmentation) and GNAN (graph neural network) on the Suzuki-Miyaura and Buchwald-Hartwig benchmarks using a shared preprocessing and evaluation protocol. MFP reaches R2 = 0.878 on Suzuki-Miyaura and R2 = 0.969 on Buchwald-Hartwig while training an order of magnitude faster than graph- or Transformer-based alternatives. A formal complexity analysis confirms that MFP folds all representation cost into a one-time preprocessing step, removing the per-epoch message-passing overhead that graph methods carry. An ablation over fingerprint radius and folded vector length shows that radius-2 representations at nBits =2048 give the best balance of accuracy, speed, and cross-split stability on both datasets. These results establish MFP as an effective, reproducible, and efficient baseline for reaction yield prediction.

The rapid advancement of Internet of Things (IoT) technology has led to the widespread deployment of smart, interconnected devices across a range of domains. However, this expansion has also resulted in a substantial increase in network traffic, creating more opportunities for malicious actors to launch cyberattacks and compromise sensitive information, thereby increasing the need for effective anomaly detection. The state-of-the-art in anomaly detection has predominantly focused on point anomalies. In contrast, the detection of collective anomalies remains relatively under-explored in the literature. In this paper, we introduce Unsupervised Graph Collective Anomaly Detection (UGCAD), a novel frame- work designed to identify collective anomalies in IoT network traffic. Unlike many existing methods, UGCAD operates on graph-structured data without any prior knowledge of group labels or membership. It leverages a variational graph autoencoder (VGAE) to learn the graph representation, which is subsequently used to enhance a clustering algorithm for effective grouping of nodes. To detect collective anomalies, clusters identified as normal are first aggregated and refined, after which anomaly scores are applied to detect collective anomalies. Extensive experiments conducted on the CICIoT2023 and ToN-IoT network datasets demonstrate the effectiveness of UGCAD in both clustering and collective anomaly detection (CAD). Furthermore, comparative evaluations against several traditional and state-of-the-art clustering-based CAD approaches confirm the superiority of UGCAD in accurately detecting collective anomalies.

Automatic synthesis of analog and Radio Frequency (RF) circuits is an emerging area that requires an efficient circuit modeling method. In recent years, Machine Learning (ML) solutions have played a promising role in this regard. However, many existing ML approaches require separate training data for each circuit topology, even when a single circuit component is added or removed. In addition, they overlook circuit topology information, which limits their ability to capture complex component interactions. Furthermore, they rely on fully connected neural networks with flat feature representations, which require substantial amounts of training data. In this work, we propose an open-source topology-aware RF circuit modeling method, TARGet. Our model considers the circuit at two levels: sub-circuits and the overall circuit topology. At the sub-circuit level, TARGet leverages S-parameter representations to capture sub-circuit behavior rather than relying on individual circuit components, providing a reusable behavioral abstraction for RF building blocks. Moreover, TARGet explicitly incorporates circuit topology information into the model, enabling it to learn across multiple topologies. TARGet introduces a novel fusion-based architecture that integrates Graph Neural Networks (GNNs) and sub-circuit connectivity-aware neural networks to improve data efficiency. Experimental evaluation across multiple RF circuit topologies demonstrates that TARGet achieves sub-1% prediction error while reducing the required training data by up to 35.5x compared to state-of-the-art (SOTA) approaches. Furthermore, TARGet achieves 9.7x higher prediction accuracy under a strict 1% error threshold relative to SOTA models. A held-out matching-network evaluation further demonstrates zero-shot transfer to an unseen sub-circuit topology, where TARGet reduces NMAE by up to 45%.

Structured EHR is abundant but sparse, coded, and difficult to use directly for note-centric clinical modeling. We present MedNotes, a multi-agent synthetic data generation pipeline that converts longitudinal structured EHR into source-grounded clinical note representations under explicit quality control. MedNotes treats structured-data-to-text synthesis as a closed-loop agentic process: a generator proposes a note, evaluator agents diagnose factual, coverage, structural, and hallucination-related failures, and a router accepts, revises, or rejects the draft. On 1,485 EHRSHOT encounters, MedNotes achieves a 91.4% pass rate, with mean factual accuracy of 0.980, completeness of 99.1%, structural fidelity of 0.761, and 0.028 critical hallucinations per encounter. Iterative refinement improves acceptance from 69.4% to 91.4%. The resulting synthetic corpus improves downstream CPT prediction and paragraph-level section prediction when combined with limited real data.

Interpretability research on large language models has produced accounts of factual recall in feed-forward layers and of token relationships in self-attention, but little work offers a unified way to compare the causal contribution of different architecture components to a model's output. We introduce MechaTerp (the Mechanistic Interpretability suite) -TRACE (subset for Teacher-forced Registry of Ablated Component Effects), an architecture and study that measures how much each registered component of a language model supports the production of a named entity. TRACE ablates one component at a time and measures the resulting change in the output distribution at a fixed answer token, so component types from whole transformer blocks down to individual neurons and output logits can be compared on a common scale. We apply it to thirteen instruction-tuned dense decoder models spanning five families and one to thirty billion parameters, ablating 49,656 components across 48 medical and 42 general-knowledge prompts. We find that the components carrying the most effect are the same few, positionally fixed components in every model, regardless of which entity a prompt asks about, and that once these are removed, the remaining support is close to evenly spread in eleven of the thirteen models. Apparent localisation of entity knowledge is therefore largely attributable to generic generation machinery, which has direct consequences for methods that assume entity knowledge sits in a findable place, including targeted knowledge editing.

Retrieval-augmented generation (RAG) improves the factuality of large language models (LLMs) and vision-language models (VLMs) by grounding generation in external knowledge. However, most existing RAG frameworks assume a centralized retrieval corpus, which is often impractical in sensitive domains such as healthcare, where data are inherently distributed and raw content cannot be directly shared across institutions. Recent efforts on decentralized RAG primarily follow prompt-based paradigms that exchange raw, human-readable retrieved content, leading to substantial inference-time computational overhead and direct exposure of retrieved information. To address these limitations, we propose Representation-based Federated RAG (FedRepRAG), a decentralized RAG framework that keeps raw documents at their owning clients and exchanges only compact latent representations during cross-client retrieval. To integrate retrieved knowledge, we introduce a collaboratively trained projector that converts retrieval embeddings into generator-compatible representation tokens for a frozen LLM/VLM backbone. Experiments across decentralized visual question answering (VQA) and question answering (QA) benchmarks show that FedRepRAG consistently outperforms direct inference and local retrieval baselines while substantially reducing retrieval-context length and inference-time computational overhead compared with raw-context transfer. Further analyses confirm the importance of query-relevant retrieved representations and characterize the residual representation-level leakage associated with representation exchange. Overall, FedRepRAG provides an effective and efficient framework for federated RAG without transferring raw retrieved content.

Medical large language models are commonly trained on mixtures of didactic data (e.g., textbooks) and clinical data (e.g., patient records), yet how these data types differentially shape model capabilities remains unclear. We address this issue with token-matched experiments that vary the didactic-to-clinical ratio and analyze how data composition affects performance, capability profiles, and error patterns across knowledge-intensive and clinic-oriented tasks. We uncover an asymmetric transfer across task types: clinical data improves clinic-oriented tasks while remaining competitive on knowledge-intensive ones, whereas didactic data mainly improves knowledge-intensive tasks. Error analysis suggests a knowing-doing gap, where improvements in knowledge recall do not reliably generalize to clinical reasoning. We further observe that modest amounts of clinical data yield most of the gains on EHR-grounded tasks, while the optimal mixture ratio varies with the knowledge and clinical reasoning demands of downstream tasks. These findings suggest that medical LLM data curation should be application-driven, with higher proportions of clinical data preferred for reasoning-intensive use cases.

Remaining Useful Life (RUL) estimates guide maintenance in semiconductor manufacturing, where a late intervention can disrupt production but an early one discards usable equipment life. A single prediction does not express the uncertainty needed to choose between these outcomes. We combine Deep Learning (DL) sequence models with Simoultaneous Quantile Regression (SQR) to estimate RUL at selectable quantile levels, and evaluate maintenance decisions using Unexpected Breaks (UB), Unexploited Lifetime (UL), and a cost-weighted objective. On the $F_1$ fault mode of the 2018 PHM Data Challenge (PHM18) ion-milling dataset, we compare recurrent, Transformer, and State Space Models (SSM) architectures. Diagonal State Spaced (S4D) gives the strongest overall RUL accuracy across the evaluated quantiles. Varying the quantile and maintenance window exposes the trade-off between failure risk and early intervention; under the examined cost ratios, the resulting decisions have lower cost than the Preventive Maintenance (PvM) baselines.

Key-value (KV) caching is essential for efficient autoregressive large language model (LLM) inference, but the cache grows linearly with context length, increasing storage and decoding costs. KV cache compression mitigates this cost by retaining only a subset of cached tokens. This challenge is particularly important for multi-step LLM agents, where a query expands into trajectories of reasoning, tool interactions, and retrieved observations. Existing pruning methods typically treat the cache as a flat token stream and rank tokens by recency or attention saliency. This creates a mismatch between the unit of compression and the unit of reasoning: token-level pruning removes individual entries, whereas useful information in multi-step agents is often organized into reasoning steps with uneven and delayed importance. Consequently, an early observation or intermediate decision may receive little recent attention yet remain essential for later evidence synthesis. We term this failure mode Reasoning Continuity Disruption.These observations motivate KV cache compression that jointly considers token- and reasoning-step-level information. StepKV addresses this goal by treating reasoning steps as first-class retention units. It associates cache entries with their generating steps, estimates step utility from trajectory-derived signals, and combines this utility with token-level saliency. The resulting scores globally rank prunable tokens, from which StepKV retains the top-scoring entries under a target budget. StepKV thus provides a step-centric perspective for agent KV cache compression. Across multi-hop QA and long-horizon web reasoning tasks, StepKV sustains accuracy under low KV budgets where token-level baselines degrade sharply, offering a more robust efficiency-accuracy trade-off for multi-step agent inference.

KV-cache eviction can do more than compress. In long-context LLMs, keeping only some cached tokens sometimes matches or exceeds full-cache accuracy, because many redundant prefill tokens otherwise dilute attention away from the tokens that carry the answer. This benefit is not uniform, and evicting the wrong tokens can drop accuracy to zero on tasks that require precise retrieval, so the useful question is not only which tokens to keep but also whether to evict this input at all. We show that one label-free number computed from the prefill attention, the drop between early and late layers in how much attention heads agree on which tokens to read, predicts per input, before any decoding, which of the two cases an input falls under. We build this into PAGE (Partition-Aware Gated Eviction), a wrapper that runs any SnapKV-style evictor when the drop is large and keeps the full cache when it is small, with no training, labels, or fine-tuning. PAGE is a safety mechanism rather than a compressor, so we measure it by the failures it prevents. It cuts the harm rate on capacity-bound inputs from 0.75 to 0.026, and on multi-key retrieval with Mistral-7B plain SnapKV falls from 99% to 0% as the budget shrinks, while PAGE holds it at 89%. Elsewhere, it passes the base evictor through unchanged, which is the intended behaviour and is what we observe in 8 of 16 cells. Code is available at https://anonymous.4open.science/r/PAGE-018239.

Speculative decoding in Mixture-of-Experts (MoE) models faces the problem of unstable verification cost caused by input-dependent expert loading. To study the physics of this process, we formulate speculation-budget selection as an offline Stochastic Shortest Path (SSP) problem over reference sequences and build a diagnostic Oracle that uses counterfactual simulation to account for MoE verification cost. A detailed analysis of the Oracle's decisions on the Qwen3-Coder and EAGLE-3 pairing, in the space of marginal deltas (Delta Space), shows that rejected candidates form a strict linear boundary. This result demonstrates that a complex global optimization is locally governed by a necessary condition balancing marginal cost against expected progress ($\frac{\Delta \mathbb{E}[Cost]}{\Delta \mathbb{E}[a]}$), providing a rigorous mathematical reference point for designing future adaptive online heuristics.

Clinical decision support tools are most useful when accurate predictions are accompanied by understandable explanations. Rule-based models provide transparency, but rules derived directly from raw clinical measurements may miss patterns arising from interactions between multiple variables. We present a four-step pipeline that mines decision rules in the latent space of an FT-Transformer and translates them back into measurable clinical features. Embedding dimensions that consistently separate patient groups are treated as latent biomarkers, rules are mined using small decision trees, and selected rules are translated using gradient-input saliency and CLS attention attribution. We evaluate the framework on six public clinical and population health datasets at four embedding dimensions. Translated rules outperformed raw-feature rules in five of six datasets, with mean AUROC gains ranging from 0.04 to 0.23. On the heart disease dataset, embedding-space rules reached 0.98 AUROC, but translation reduced this to 0.72, showing that high-performing latent rules cannot always be represented by simple raw-feature conditions. These results show that latent-space rule discovery can uncover predictive patterns while translating them into clinically measurable features that can be evaluated by clinicians.

Tabular data is the most common format in clinical practice, encompassing laboratory results, medication records, diagnostic codes, and patient demographics. As foundation models for tabular data have grown in number and variety, a practical question has become harder to answer: which model should a clinician or data scientist actually choose for a given task, and why? Existing surveys catalogue what these models can do, but they stop short of providing a structured way to compare them against the specific demands of a real application. We introduce \system{}, a comparative evaluation framework that scores and ranks tabular foundation models (TFMs) across six clinically meaningful dimensions: how well a model generalizes to new datasets, how effectively it protects patient privacy, how much data it needs to perform well, how it scales with growing datasets and feature spaces, how interpretable its predictions are to clinicians, and how fairly it performs across patient subgroups. Each dimension is broken down into measurable sub-components, and groups of sub-components can optionally be combined into supplementary compound scores, called super-metrics, that provide a diagnostic view of how a model performs across several dimensions simultaneously. To show how the framework works in practice, we apply it to two healthcare use cases, screening for iron deficiency and predicting heart failure, demonstrating how the same set of metrics leads to different model rankings depending on what matters most in each clinical context. We also provide a taxonomy of 45 TFMs organized by their underlying architecture, which serves as a reference for researchers and practitioners looking to navigate this rapidly expanding field.

Methods for addressing safety drift in fine-tuned Large Language Models (LLMs) are scattered across incompatible implementations, lifecycle stages, and evaluation protocols, making them difficult to adopt and compare. We introduce SafeTune, a source-available library that unifies four intervention paradigms: post-hoc weight recovery, safety-constrained fine-tuning, gradient-based unlearning, and inference-time steering, alongside shared interpretability, evaluation, and deployment utilities. SafeTune provides a consistent configuration-driven workflow while preserving the distinct inputs and intervention points each paradigm requires. Its modular registry supports new methods, benchmarks, judges, models, and fine-tuning domains without redesigning the surrounding pipeline. We demonstrate SafeTune through controlled comparisons and finance and medical deployment case studies, showing how it characterizes safety drift, evaluates feasible interventions on common refusal-behavior and capability evaluations, and supports calibrated or layered mitigation.