Papers

Modern recommendation is shifting from platform-centric personalization toward user-governed personalization, where a personal LLM agent can act on the user's behalf across services. We formalize this emerging paradigm as Personal-Agent Mediated Recommendation: a platform recommender ranks a candidate set using platform-local information, and a personal agent uses user-authorized cross-platform history to mediate the resulting ranking and produce the final top-K slate. Such mediation is nontrivial: the platform ranking can encode strong population evidence that the personal agent cannot observe, so effective mediation must therefore balance beneficial rescues against harmful overrides. To study this trade-off, we introduce MediateRec, a benchmark that includes scalable proxy cross-platform environments and a real cross-platform test under a controlled platform-agent information boundary. To train the agent to use cross-platform history effectively, we further propose Personal Attribution Mediation Optimization (PAMO), which counterfactually masks that history to estimate personal mediation support and reallocates rank-aware advantage mass under a platform-relative value floor. We theoretically prove that PAMO preserves cutoff-level advantage mass and is locally optimal among first-order reallocations that preserve this mass without lowering average platform-relative value. Experiments on MediateRec show that personal-agent mediation enables meaningful platform corrections, yet even strong proprietary LLMs introduce non-negligible harmful overrides. PAMO consistently improves over matched outcome-only RL across seen and unseen target platforms and on the real cross-platform test, while achieving a better rescue-harm balance.

While AI weather models now rival operational forecasts, how they represent the atmosphere internally remains an open question: feature attribution reveals which input patterns matter, not what the model computes or how it combines information internally. We train sparse autoencoders (SAEs) on GraphCast to uncover its learned concepts, using atmospheric rivers as our phenomenon of focus. Both standard and Matryoshka SAEs show GraphCast computes atmospheric river intensity, measured by integrated vapor transport (IVT), as a stable internal variable, despite IVT being neither an input nor a target. In contrast to the unstructured concept retrieval of the standard SAE, the Matryoshka SAE orders concepts by importance and exposes their relations. Atmospheric river concepts persist across depth and direct interventions confirm causality. This method offers a way to find internal variables and determine which of them the model actually relies on, which is a prerequisite for asking whether those variables remain meaningful as the phenomenon changes under a warming climate.

Quantitative imaging pipelines can produce precise but systematically different measurements of the same object. We present an empirical reliability assessment of three-dimensional mitochondrial morphometry that connects representation bias, a controlled processing intervention, correction transfer, and resolution sensitivity. Using 2,720 development objects from the 3D Mitochondria Shape Library for Optical Microscopy, we find that occupancy-derived volumes exceed reference mesh volumes by 3.665% on average despite an intraclass correlation coefficient of 0.994. Boundary analysis identifies an outward label displacement of 0.00304 normalized units. In a controlled label-pipeline reimplementation, removing the depth offset reduces volume error in all 55 analyzed objects by a mean of 1.57 percentage points, approximately 45% of mean reproduced inflation; the source of the remainder is not isolated. A frozen regression using occupancy-derived features reduces median absolute percentage error from 3.481% to 0.664% in 2,728 previously unused objects from the same resource. However, its calibrated error bound covers only 92.1% overall and 49.2% in a low-occupancy subgroup, demonstrating that accuracy and uncertainty transfer must be evaluated separately. In 550 rat-cortex objects from the MitoEM resource, coarsening in-plane spacing from 8 to 24 nanometers changes median surface area by minus 10.60% and sphericity by plus 11.76%, despite a rank correlation of 0.994. These results provide quantitative checks for distinguishing processing-induced descriptor changes from candidate biological differences, without establishing biological invariance or cross-source correction transfer.

Aerospace electrical-design revisions can contain multiple genuine changes, although an engineering request may authorize only a subset. Propagating every detected difference can therefore produce overly broad impact reports. We present LOGIC, a controlled benchmark and evaluation framework in which locally deployable language models ground a request in a deterministic candidate-change inventory before selected changes are propagated through a typed electrical traceability graph. This separation permits candidate-selection errors to be distinguished from downstream propagation errors. LOGIC contains 168 scenarios, including 144 selection and 24 abstention cases. We evaluate three 7–8B models against intent-agnostic, lexical, and structured-evidence methods, with an oracle-root upper bound. On 96 explicitly anchored selection cases, gate-only structured evidence achieves candidate F1 of 1.0000, compared with 0.9677 for token-lexical matching. On 12 relational-paraphrase cases, token-lexical F1 is 0.1772 and gate-only F1 is 0.0000, compared with 0.5000–0.6400 for the large language models. Model grounding degrades as candidate inventories grow from 4 to 64 changes, while affected-element and typed-path accuracy remain comparatively stable when frozen selections are replayed over graphs of approximately 1K to 100K nodes. Strict evidence gating suppresses false positives but can remove correct semantic selections. An exploratory evidence-empty abstention policy raises strict abstention accuracy to 0.6667 for all three models and reduces unsafe-report rates to 0.1667, while decreasing answerable-case coverage by 16.0–27.1 percentage points. Four of six conflicting requests remain unsafe for each model. These findings support combining literal evidence and language-model reasoning with engineering review when intent cannot be established reliably.

In cooperative Multi-Agent Reinforcement Learning (MARL), agents are often trained under concurrent participation, while in many tasks some agents act earlier and leave task-relevant information that becomes useful to agents participating later. We study this setting as staggered participation (SP), which introduces a cross-time, cross-agent learning dependency because an early action may affect the return through the information it provides and the later policy that uses it. Learning under SP therefore requires both identifying what information is useful for future decisions and learning how later agents should use it. We propose Staggered Participation Learning (SPL), a training-time augmentation that addresses these two parts with prospective acquisition supervision for earlier agents and outcome-supervised receiver learning for later agents. We evaluate SPL across multiple policy-based MARL backbones, environments, and staggered-participation patterns. Across 60 MPE/RWARE backbone setting comparisons, SPL achieves higher observed mean task completion in every case, with an average difference of 14.1%. The gains also extend to eight-agent teams and a physics-based UAV-UGV environment in Isaac Lab, providing evidence across algorithmic, temporal, and embodied settings.

In clinical practice, agreement among independent experts is treated as evidence of reliability, and multi-round consensus has become a core mechanism of agentic medical question-answering systems. When such a system must decide whether to trust its own answer, the prevailing signal is again agreement, now among the sampled answers. But agreement is a fragile proxy for correctness. A system can be unanimously wrong, returning the same incorrect answer on every sample, and on these questions agreement-based signals carry no information. The cause is that these signals read only the final state of the consensus and discard how it was reached. Agreement that was reached by resolving disagreement with evidence looks identical, at the end, to agreement that was present from the first sample because every sample shares one misconception. ProbeGuard is a certified abstention framework that bases the abstention decision on how the consensus formed. Process features trace agreement trajectories, minority persistence, and retrieval saturation. For unanimous votes, rationale semantic entropy checks whether the reasons behind the vote cohere, and an active probe retrieves counter-evidence and measures whether the consensus survives. A stratified Learn-then-Test calibration then converts these scores into a distribution-free bound on selective risk. We evaluate ProbeGuard on three medical QA benchmarks and a hard-frontier reference, with a published multi-round agentic RAG substrate, against six abstention baselines. On MedQA, 13.4% of unanimous votes are wrong, and no agreement-based signal can flag them. Process signals raise the discrimination of correct from incorrect consensus from chance to 0.696 AUROC. The certified rule answers six in ten unanimous-layer questions at an observed selective risk of 9.0%, and nine in ten once in-domain calibration data accumulate.

Dense 3D mapping with semantic understanding is essential for robotic perception in complex environments. Recent 3D Gaussian Splatting-based mapping approaches enable high-fidelity geometry and efficient open-vocabulary perception, but typically represent semantics as unstructured feature fields that limit object-centric reasoning. In contrast, 3D scene graphs explicitly model objects and their relationships for structured reasoning, but are commonly constructed from sparse geometric representations that do not fully exploit dense semantic maps. In this work, we present OpenSplatGraph, a unified framework that constructs persistent 3D scene graphs directly from an online Gaussian-based open-vocabulary semantic map. The proposed framework augments the dense semantic map with a reliability-aware semantic field that maintains lightweight observation statistics for confidence-aware, query-conditioned object extraction. Extracted object instances are associated with persistent graph nodes, allowing object attributes and relationships to be incrementally updated across observations and queries. By tightly coupling dense semantic mapping with persistent object-centric representations, our framework supports both language-guided object grounding and structured relational reasoning while preserving the geometric fidelity of Gaussian-based mapping. Comprehensive evaluations on standard 3D scene understanding benchmarks and real-world robotic experiments demonstrate that OpenSplatGraph achieves competitive performance for online open-vocabulary perception and downstream robotic tasks. Project page: https://csiro-robotics.github.io/OpenSplatGraph.

Complementary Feature Domains (CFD) theory characterizes predictive value as a context-indexed contribution system induced jointly by representations and their realization family. We show that Shannon-information preservation does not imply preservation of this contribution system: an invertible representation transformation can leave target information unchanged while altering predictive contribution under a restricted decision family. We formalize the resulting transition through a CFD contribution defect that measures how contextual contributions change under controlled recoding. For bounded Lipschitz utility, we show that each coalition utility shift is bounded by the behavioral distance between the attainable action sets before and after recoding; consequently, every contextual contribution defect is bounded by the sum of the corresponding coalition incompatibilities. Exact behavioral closure yields invariance, while increasingly accurate compensation yields restoration. A controlled ECG experiment illustrates the mechanism: a nonlinear bijective recoding preserves the information in a frozen time-frequency representation but changes accuracy under a fixed affine learner; applying the exact inverse restores all tested coalition accuracies. The result separates information preservation from realization-dependent contribution and provides a quantitative transition law for multi-representation prediction.

Learning an ensemble of GFlowNets to sample from a discrete target distribution has become a common approach for achieving better state space exploration and convergence than that of a monolithic sampler. However, these methods often add a substantial runtime overhead to the base model, and their conceptual connection remains elusive. To address this, we first propose a general-purpose theoretical framework for describing a mixture of GFlowNets, which we specialize into continuously (CI) and discretely indexed (DI) collections. On the one hand, we show CI GFlowNets can be interpreted through the lens of a random features expansion, provably boosting the sampler's expressivity in graph-structured tasks and reducing learning instability via spectral shifting. On the other hand, we demonstrate DI GFlowNets encompass prior approaches for GFlowNet training and provide the foundation for the newly proposed Stratum-Conditioned (SC) GFlowNets. This method, which is inspired by the Doob's h-transform of Markov chains, decomposes the state space according to a prescribed modular function and restricts each component to sample from a distinct subset of it. Importantly, SC GFlowNets support centralized and component-wise embarrassingly parallel training, and we show both of them significantly speed up learning convergence and mode coverage without introducing any non-negligible extra computation.

Goal-directed molecular design has advanced rapidly, yet a substantial proportion of designed molecules remain difficult to synthesize in practice, limiting their real-world utility. Prior synthesizability-aware methods either project generated molecules back to synthesizable analogs that deviate from the intended target, or optimize directly in discrete synthesis spaces that lack a continuous landscape for efficient search. We argue that this limitation mainly comes from the search space rather than the optimizer. To address this, we propose RouteFlow, a framework that reformulates synthesizable molecular design as a search over a continuous route latent space, where each latent maps back to a complete synthesis route and synthesizability is inherently preserved. To navigate this space, we adopt reward-guided flow matching as an efficient sampler that steers toward high-property regions. Since reward optimization may push latents off the manifold of real synthesis routes, where decoding becomes unreliable, we further introduce a cycle-consistency mechanism to stabilize fine-tuning. Across 16 optimization tasks from Therapeutic Data Commons, RouteFlow achieves the best sample efficiency among synthesizability-aware baselines, with the best synthetic accessibility and the highest retrosynthesis success rate. Our results also confirm that the proposed cycle-consistency reliably keeps optimization on-manifold while improving target properties, supporting effective synthesizable molecular discovery.

Vision-Language-Action (VLA) models have become a major paradigm for Vision-and-Language Navigation (VLN). However, in safety-critical facilities, invisible risks such as radiation or temperature spikes cannot be detected by an RGB camera, and handling each risk is expensive, requiring a new encoder, new data, and model retraining. We propose Physics-Guided Visual Prompting (PG-VP), a plug-and-play multimodal perception module that instead reuses what a frozen VLA model already does well: avoiding visible obstacles. Given a proximal radiation or thermal source, PG-VP performs a physics-guided risk assessment to determine the avoidance direction and overlays a corresponding virtual obstacle that moves across consecutive frames (Dynamic Visual Prompting). The navigation policy then naturally detours around this invisible hazard. The identical virtual obstacle is used regardless of hazard type, so the visual prompting pattern remains fixed as sensors are added. When no hazard is detected, nothing is rendered, and the policy behaves exactly as it would without PG-VP. We evaluate PG-VP on OmniNav using the val-unseen splits of R2R-CE and RxR-CE, where it guides the policy toward intended low-risk actions in 84.9% and 83.2% of cases, at a cost of 6.8 and 7.9 percentage points in navigation success rate. We further test it with distinct scenarios on a real robot in the presence of actual thermal and radiation sources, all without any retraining. The real test shows that PG-VP effectively avoids these invisible hazards, improving worst-10% average trajectory safety by 63.45% and 32.59% against thermal and radiation sources, respectively.

Static-analysis checker synthesis requires agents to interpret a defect specification, inspect a repository, implement analyzer-specific logic, and refine the checker through repeated compilation and analysis feedback. Existing coding-agent benchmarks focus on tasks such as patch generation or vulnerability detection and rarely assess whether an agent can develop a working checker in a repository from start to finish. We introduce CheckerBench, an executable benchmark of 300 tasks derived from 297 CVEs across 167 repositories, 85 CWEs, and five language ecosystems. Each task includes vulnerable and fixed revisions, a pinned analysis environment, and a checker scaffold. We further introduce CheckerLab, a common evaluation framework that independently rebuilds submitted checkers and measures vulnerable-fixed diagnostic contrast, patch localization, false positives, and tool use. Across 21 model-harness configurations and three independent repeats per configuration, mean Pass@1 is 32.30%, while the best reaches 45.33%. These results show that reliable, reusable checker development remains challenging for current coding agents.

Learned orchestration can automatically construct effective language-model multi-agent systems, but existing approaches couple planning to fixed worker pools and train decomposition and collaboration from the same terminal outcome, limiting transfer and obscuring credit assignment. We introduce DeOrch, which separates worker-agnostic planning from concrete worker selection. Its two-stage planner first decomposes the task without worker information, then chooses collaboration operations using compact, worker-identity-free matchability feedback from the pool, enabling conditional credit assignment to decomposition and collaboration decisions. A lightweight matcher estimates worker suitability from behavior on a fixed probe set and adapts online with a contextual bandit, allowing new workers to be incorporated without retraining the planner or matcher. Across diverse in- and out-of-distribution tasks, DeOrch outperforms prior automatic MAS orchestration methods with fewer worker calls than competing learned orchestrators, remains effective when transferred to an entirely unseen worker pool without retraining, and shows consistent gains from both components.

LoRA fine-tuning adapts small language models (SLMs) to heterogeneous instruction data within a low-rank update subspace, making it vulnerable to three structural problems: conflicting gradients that cancel, static data selection that cannot track evolving learning dynamics, and subspace saturation that causes later updates to overwrite useful directions. We argue that effective adaptation therefore requires controlling which data-induced gradients enter the LoRA subspace and when. We propose GRADE (GRadient-Aligned Data-centric rEcipe), a data-centric framework combining two mechanisms: a state-aware selector that continually admits samples aligned with the evolving multi-task gradient field, and a self-calibrating step-level gate that rejects updates likely to cause destructive overwrite near saturation. Across three current-generation backbones and a heterogeneous seven-dataset instruction pool, GRADE outperforms strong data-selection and PEFT-stabilization baselines in accuracy and robustness. It is the only method to improve consistently over standard LoRA on every architecture, while producing more coherent gradient trajectories and less destructive overwrite. These results show that successful SLM adaptation depends not only on which data are selected, but also on which gradients are allowed to enter and persist in the constrained update subspace.

Random access (RA) is one of the most foundational medium access control (MAC) layer scheduling schemes for handling unpredictable data traffic from multiple terminals. While multi-agent reinforcement learning (MARL) has been explored to optimize RA-based wireless networks, its reliance on experience-driven, distributed policy learning incurs significant training overhead for each optimization task, limiting its feasibility in real-world applications. In this work, we propose to leverage a foundation model (FM) to improve MARL efficiency across diverse RA network optimization tasks. Specifically, we design an FM-aided actor-critic algorithm within a consensus-based decentralized MARL architecture and provide its convergence analysis under local reward exchanges and nonlinear value function approximations to show that our algorithm achieves the same convergence order as the conventional MARL with critic model exchanges and linear approximations. Our numerical results show that our FM-based approach significantly enhances MARL speed for RA network optimization.

Large language models (LLMs) are increasingly used to judge how well an advertisement matches a query, but the fairness of these judgments has received limited attention. We conduct a systematic study of fairness in relevance judgments made by LLMs for queries and advertisements. Our counterfactual framework examines the effects of advertiser identity and possible popularity, input language, and demographic wording. We study GPT-4o as a categorical relevance judge and a Qwen-7B model trained specifically for relevance prediction. The advertiser and language experiments use query and advertisement pairs sampled from real advertising logs. Controlled synthetic queries are used to study demographic associations in employment, housing, and credit. For both models, changing the advertiser identity or input language can alter the relevance assessment. Selected demographic comparisons also show patterns consistent with common stereotypes, particularly those involving gender and occupation. We further study mitigation during model inference and training. The results indicate that its effectiveness depends on whether advertiser information is relevant to the query and how advertiser labels are distributed in the training data. These findings can help advertising practitioners identify fairness risks and develop suitable mitigation methods for LLM relevance systems.

Multi-agent simulations with large language models (LLMs) often operate networks of agents with a single base model. This overlooks the inter-model effects which may dominate engagement dynamics in real-world deployments. To show this, we simulate a heterogeneous social network powered by several different base models and show that the amount of engagement an agent receives depends more on its base model than on its assigned persona. The attraction or repulsion effects of a base model strengthen dramatically when more models are added in the mix, suggesting that networks dynamics may converge to base model effects at scale. To help explain this effect, we conduct a series of content-mediating analyses, showing the predictability of base models across contexts as well as the relationship between a model's lexical patterns and an engagement-maximizing style. In light of recent developments in mass multi-agent interaction, this work underscores the relevance of heterogeneous compositions in driving the outcomes of those networks

LLMs can generate Lean specifications from task descriptions, but those descriptions are often underspecified or ambiguous and may omit verifier-relevant edge cases, yielding imperfect formal specifications. We study whether an elicitation agent (EA) can improve verifier-checked Lean specifications by seeking clarification from a reference-grounded simulated developer before formalization. The agent decomposes the task, identifies unresolved decisions, asks bounded clarification questions, and uses the answers to write a refined natural-language specification that an LLM formalizer then translates into Lean. On 189 Verina Lean 4 specification-generation tasks, we compare single-shot generation with elicitation under question budgets of k=3 and k=10, allowing at most three or ten EA questions per task, across same-model and cross-model configurations. The same-model Claude Sonnet 4.6 pipeline has negative strict pass@1 estimates at both budgets, showing elicitation is not a uniform accuracy booster. In contrast, the Sonnet 4.6-to-GPT-5.6 and Sonnet 4.6-to-Gemini 3.8 Flash pipelines have positive strict estimates. The largest strict result reaches 65.6%, a +6.3-point change (+12 tasks), in the Sonnet 4.6-to-GPT-5.6 pipeline. Six of eight conditions have positive compile-conditional estimates, with a maximum cross-model difference of +4.9 points among compiling outputs. Treatment estimates also differ across the basic and advanced splits; the largest advanced-task change occurs in the Sonnet 4.6-to-GPT-5.6 configuration at k=10, from 42.0% to 50.6%. Failure analysis identifies postconditions as the largest source of degradation. These findings suggest future autoformalization systems should treat elicited natural-language specifications as intermediate artifacts and optimize jointly for semantic correctness among compiling outputs and reliable translation of clarified information into Lean.

Low-rank compression reduces the cost of pretrained language models by replacing linear transformations with low-rank factorizations. However, conventional methods use a fixed rank allocation during inference, assigning the same amount of compute regardless of the input token. We introduce Low-Rank Conditional Computation (LRCC), which adds token-dependent computation to pretrained models by training one lightweight router per Transformer block to select among a small set of nested low-rank paths. During training, the low-rank factors remain frozen, and only the routers are optimized. We evaluate LRCC on Llama and Qwen models for language modeling and zero-shot downstream tasks. Within the same average active-parameter budget, LRCC improves the predictive performance over static low-rank compression, including a 7.6 percentage-point gain in average downstream accuracy on Llama-2-7B over static methods. At matched batch-size-1 decoding latency, LRCC improves both perplexity and downstream accuracy on Llama-3.2-1B and remains competitive on Llama-2-7B, without specialized kernels. Finally, we assess the usefulness of assigning a token-wise path by analyzing the routers' path choices.

LLMs have advanced rapidly, raising growing concerns about their safety. Recent work has proposed approaches to detect and defend against attacks including defenses at decoding stage that leverage models' hidden states. However, existing decoding-stage defenses suffer from two limitations. First, they introduce a trade-off between safety and over-refusal, where strengthening safety degrades the model's helpfulness on benign queries. Second, many of these methods rely on internal hidden states and are thus restricted to specific architectures, incurring substantial overhead and limited generalization across models. To address these limitations, we introduce LADE (Latent Safety Signals for Defense), which leverages latent safety signals extracted by contrasting harmful and benign queries from dark knowledge (i.e., information carried by the output probability distribution beyond its argmax) in the first-token output probability distribution. Our key insight is that, beyond surface-level refusal tokens, the dark knowledge in the first-token distribution contains latent safety signals, defined as tokens whose probabilities differ sharply between harmful and benign queries. We show that these signals consistently align across LLMs, forming a model-agnostic direction that emerges from safety alignment. LADE consists of three components: (1) Extracting Latent Safety Signals from Dark Knowledge, which selects top-k safety-discriminative tokens from the first-token probability distribution; (2) Tokenizer Mapping, which maps these tokens across different tokenizers to enable model-agnostic application; and (3) kNN-based Discrimination, which classifies queries via a k-Nearest Neighbors search over the mapped tokens. Across diverse LLMs and benchmarks, LADE is robust against a wide range of jailbreak attacks and lowers attack success rates while maintaining a competitive safety-utility trade-off.

Driving models increasingly ground reasoning in causal relations, spatial structure, perceptual evidence, and predicted futures. These advances make reasoning more faithful to the driving scene, but leave a fundamental question unresolved: what should groundedness mean when the model ultimately outputs an action? Correctly grounded reasoning does not, by itself, ensure desirable driving outcomes. We introduce GroundAct, which starts from a simple premise: driving unfolds through physical entities and their interactions. Entities therefore become the unit of grounding; a lightweight reference token keeps each selected entity's continuous state addressable through symbolic reasoning; and only the referenced entities' interactions with the evolving proposal correct the plan. The result is an explicit path from what reasoning grounds to what the plan does, which we call grounded planning. To assess its practical value, we evaluate GroundAct in both open- and closed-loop settings. GroundAct shows strong open-loop planning across normal, out-of-distribution, and safety-critical scenarios, with closed-loop results extending this evidence to driving in simulation.

Safety auditing of post-trained large language models typically relies on model behavior, requiring model execution and depending on the coverage of available evaluations. This work asks a different question: Do the target behaviors optimized during supervised fine-tuning (SFT) leave readable evidence directly in checkpoint updates? We find that harmful-compliance SFT induces a continuous, objective-dependent ordering in checkpoint-update space. Using a reference geometry defined by pure harmful-compliance, safety-targeted, and benign-utility SFT, we find that a checkpoint-level coordinate s_H tracks controlled harmful-objective composition with Spearman correlations of 0.986-0.992 across four 7-8B backbones, with the same ordering persisting at larger model scales. Matched compliance-versus-refusal controls show that this checkpoint trace reflects the SFT objective rather than harmful-input exposure, while additional controls rule out simple explanations based on harmful-example count or generic training intensity. Building on this structure, we introduce TRACE, a weights-only auditing method that localizes an unknown checkpoint update relative to frozen harmful and non-harmful reference prototypes and converts this geometry into a continuous harmful-objective score. TRACE requires neither model queries nor access to the unknown SFT data, and can be evaluated directly from checkpoint updates. Across distribution shifts, unseen data, different SFT configurations, partial checkpoint access, and LoRA/full-parameter fine-tuning, the trace remains stable and is positively associated with independently measured attack success rates. TRACE remains informative even at low harmful-objective proportions, providing a complementary auditing signal when behavioral evaluation is unavailable or incomplete. Code is available at https://anonymous.4open.science/r/Code4TRACE-54D3.

A vision-language model may need to combine an image with a prompt to recognize a safety risk that neither reveals alone. Where does this joint safety judgment become accessible inside the model? We introduce SSU-Bench, a dataset of matched safe and unsafe image-text combinations constructed using single-item prompt edits or image edits with annotated intended regions. Using three vision-language models, we transfer internal states between paired inputs and measure the resulting change in the safety verdict. Across models and both types of counterfactual, interventions at the changed input positions are effective in earlier decoder layers, while interventions at the final input token become effective later. Directions estimated from other examples produce similar late-layer effects. A linear readout of the final-token state also predicts the model's own verdict, including incorrect judgments, and cross-model comparisons reveal similarities in the patterns of counterfactual change. These findings identify a recurring transition in where interventions can influence a joint safety verdict and distinguish a readable model decision from a correct safety judgment.

Information elicitation is an open-ended information-seeking problem in which an interaction can unfold in many potentially valuable directions, requiring an elicitor to continually determine which information to pursue as new information emerges. In agentic elicitation, these decisions may be delegated to a foundation model, yet how model choice shapes the resulting information-seeking behavior remains understudied. We study how judgments about information value vary across LLMs and how these differences shape sequential information seeking. We first examine these judgments across 11 LLMs spanning multiple model families and parameter scales, using a shared set of information and elicitation objectives. We then develop a controlled elicitation simulation in which different models encounter the same information space and use the same selection rule, isolating these judgments from question generation and respondent behavior. Using this setting, we characterize the breadth-depth behavior that emerges from model-specific information-seeking preferences over the course of elicitation. We further examine how interaction history changes the evaluation and subsequent selection of prospective information. We test the robustness and boundaries of these findings through sensitivity analyses and ablations over the opportunities available to the elicitor, the response labels used to operationalize information-seeking preferences, the presence of interaction history, and whether redundancy is explicitly relevant to the assessment. The project code, data, and trajectory files are available at https://github.com/infosenselab/open-elicitation.

Speech agents are reactive and dyadic: they speak when spoken to, and to one person at a time. We ask what it takes for a speech agent to instead take part in a conversation among several people and speak up only when it can help. We introduce Jarvis, a real-time proactive speech agent that audibly participates in multi-party human conversations. Grounded in a document shared beforehand, Jarvis follows the discussion and intervenes when the group misses or misstates a fact and does not correct itself within a few turns. We make three contributions: a problem setting based on epistemic breakdowns that makes proactive intervention measurable, realized as CHI-180-proactive, a synthetic multi-party dataset seeded with known gaps, errors, and self-corrections; a proactive backbone that harnesses a small, open-weight model with deterministic checks and grounds every claim in a source sentence; and interaction techniques for taking the floor in live speech and showing the cited evidence on screen. On CHI-180-proactive, Jarvis is correct on most events it addresses and stays silent 97% of the time when the group resolves an issue itself. A live study with 23 participants confirms these trends with real-time interventions.

Long-history recommenders often compress each user's history into a compact, candidate-independent memory that is cached and reused to score large candidate pools. We show that real user histories exhibit multi-scale semantic structure, with short-lived intent, medium-term interests, and long-term preferences coexisting in one sequence, and that monolithic cached memories preserve these scales unevenly: linear probes recover recent and mid-range content far worse than long-range content. We call this failure mode temporal aliasing. We propose MARS, a multi-resolution user memory that writes the full history into recurrent state tracks anchored to different half-lives, and a sparse routing reader that materializes compact seed memories by selecting the relevant temporal resolutions for each seed, preserving fixed-size candidate scoring. MARS outperforms strong baselines on three public datasets, with gains that grow with history length. Component-matched ablations with paired tests show that temporal diversity and selective routing each contribute beyond what hard-window memories or added capacity provide. The advantage of MARS over its interface-matched baseline also widens after within-user behavioral shifts, at about $1.02\times$ that baseline's warm-cache serving latency for $1{,}000$ candidates per user.

Muon combines current and past gradients into matrix momentum. For $M=U\Sigma V^\top$, the idealized polar update $Q=UV^\top$ gives every singular direction the same weight. We refer to this as the flat profile. Several recent optimizers replace this flat profile with fine-grained spectral maps that give each direction its own gain. We ask how much of this spectral detail a Muon update needs. Our spectral diagnostics show that approximately $94$–$97\%$ of measured singular modes lie below an estimated noise edge, yet collectively align positively with a reference gradient. We introduce BulkBoost, a two-band spectral reweighting framework with fixed-rank and noise-calibrated variants. The latter uses split-minibatch gradient differences to calibrate a Marchenko–Pastur reference edge for Muon's Nesterov input, separating the bulk below the edge from the spikes above it. Both variants increase the bulk's relative weight through one shared gain while preserving the Frobenius norm of each matrix's unreweighted direction. For a fixed partition, our theory gives the first-order condition under which moving weight toward the bulk lowers the loss. It also quantifies the fraction of the maximal first-order improvement rate, over all per-mode reallocations, that two bands can capture. Across 30 continued-pretraining settings spanning Pythia-14M to 410M and six corpora, two-band reweighting is competitive with the fine-grained power-law profile of Freon and outperforms Spectra. Measured against Muon's flat profile, Freon reduces final loss by $0.022\%$ of the pre-adaptation loss on average, whereas the two-band variants achieve reductions of $0.073$–$0.147\%$. These observations suggest that useful departures from the flat profile are surprisingly low-dimensional: a single bulk-to-spike gain captures at least as much benefit as the fine-grained spectral profiles.

When multiple agents share a cost budget, a common Lagrange multiplier can enforce the aggregate constraint but does not determine how its penalty should be allocated across agents. Uniform penalties ignore heterogeneity in the rewards agents sacrifice, while agent-specific multipliers may still rely on the same aggregate cost signal. We introduce Lagrangian Responsibility Allocation (LiRA), which learns each agent's share of a common multiplier by optimizing social welfare over a finite training horizon. The multiplier enforces the aggregate budget, while responsibility shares redistribute its influence without modifying the original rewards or constraints. For convex games under standard regularity conditions, varying these shares induces a smooth family of normalized generalized Nash equilibria in which active constraints remain at their budgets while welfare varies. To optimize responsibility before convergence, we derive a welfare gradient that accounts for both learning updates and the induced change in data distribution. Across CityLearn, MABIM, Harvest, and MetaDrive, spanning 3 to 400 agents, LiRA improves average social welfare by up to 29% over uniform and agent-specific multiplier baselines. Grid and driving costs remain within budget, inventory violations decrease, and Harvest makes more effective use of available budget.

Software-based legal assistance systems have leveraged many different forms of knowledge representation and reasoning. This article explores how document automation services rooted in expert system style and other symbolic approaches can usefully enhance and be enhanced by current generative AI approaches. We discuss the possible benefits and challenges, and report on preliminary experiments in using large language models to identify and fix issues in texts written by laypeople.

Frontier AI treaties or agreements on limiting computation require external verification; an external auditor must be able to confirm how much computation actually ran and that parties are adhering to the agreement. Analogue, off-chip measurements such as power draw provide an information channel for verification. It is unknown how well these analogue channels can constrain computation against an adversary who actively tries to subvert the audit. We derive a closed form for $\beta$, the largest hidden computation a power trace cannot exclude, as a fraction of the declared machine capacity. Measurements on NVIDIA A100 GPUs constrain $\beta = 1.16$ in the worst case, while adversarial matched-energy strategies are shown to hide at least $\beta = 0.41$ of compute. Analogue power measurements alone therefore constrain compute weakly. Additional restrictions granted by the threat model, such as the ability of the verifier to re-execute the declared work at an observed operating point, let the verifier push $\beta$ down to $0.059$ in the maximally restricted case. This gives a quantitative estimate of what analogue measurements can contribute to compute verification.