Science & Mathematics

Symbolic constraints (programmatic, deterministic checks for tool calls) enable deterministic verification of LLM agents' tool calls, but existing approaches primarily enforce predefined safety policies rather than tool dependencies in complex tool environments. We explore whether symbolic constraints can be mined from environments and observations beyond predefined policies, and whether they can improve end-to-end task performance. We present Symbolic Constraint Agents (SCA), a framework that mines symbolic constraints from API documentation, refines them against execution traces, and checks the agent's tool calls with an SMT solver at runtime. A violation triggers a natural-language warning or, when the valid replacement is uniquely determined, an automatic repair. On AppWorld, refinement raises error-detection F1 from 37.0 to 58.3 on GPT-5.5 traces and from 39.3 to 61.9 on Qwen3.5-27B traces. At runtime, repair raises GPT-5.4-mini's task success by 5.4 points under a 40-call budget and 9.8 points when environment errors are hidden, while removing 59--87\% of failed calls across models, and transfers to $τ^2$-bench with gains of up to 5.6 points. These findings show that automatically acquired symbolic constraints can improve tool-calling reliability and task completion, particularly when failures are costly or difficult to diagnose.

System One models, introduced with TypeSafe's Jev, answer typed questions about a state with calibrated probabilities instead of generated text: choosing among options described in natural language (Choice), placing the state on an ordered scale (Score), or judging a yes/no statement (Noul). Jev is general-purpose: one set of weights, served as an API and adapted to a domain only through the request. We study the other end of the design space. A Narrow Decision Model (NDM) offers the same interface but is trained for the decisions of one organization, is small enough to run on its own infrastructure, and has an open architecture whose properties can be stated and tested. The state can be any text or structured record about which decisions are made; our running case is the routing of court communications. The architecture serializes a request as a prefix tree (the state, then each question, then each option) and runs a bidirectional encoder under a block attention mask with logical positions, so that each answer is, by construction, invariant to the order of its options and independent of the other questions asked. Routing through an organization chart is a composition of Choice decisions with an explicit "none of these" option. We describe training with proper scoring rules through a curriculum, calibration and abstention, decisions composed by rubrics, and a protocol that accepts or rejects the mechanism before held-out data is opened. A minimal validation on public and fictitious data shows that the invariances hold to rounding in a trained model, that a small instance routes on organization charts it has never seen well above lexical retrieval, and that, at equal compute, the block architecture reaches a lower loss than a pairwise cross-encoder. A first instance trained on real legal documents agrees with the LLM classifier in production on 91% of held-out documents.

Rapid advances in image synthesis have increased the need for reliable synthetic-image detectors, yet their robustness to transferable attacks remains incomplete and understudied. Existing evaluations generally examine whether an attack constructed against one detector transfers to another, but rarely isolate how its effectiveness changes when the attacked images originate from different generators. We conduct a systematic evaluation of three representative attacks: two iterative perturbation-based methods, MIG and MuMoDIG, and the amortized generative attack Tricker. Across four detectors and five generators spanning diverse architectures, we introduce a two-axis evaluation protocol that independently varies the surrogate detector used to construct or train the attack and the generator supplying its input images. We assess attack effectiveness, cross-detector transferability, perceptual quality, runtime, and interaction with common post-processing operations. We revisit Tricker and show that its data and training requirements can be substantially reduced, enabling broader and more efficient evaluation. Our results demonstrate that transferability is not a detector-only property: the input-generator distribution strongly influences both evasion effectiveness and attacked-image quality. MIG and MuMoDIG exhibit pronounced dependence on the detector--generator pair, whereas Tricker provides more uniform white-box evasion and image quality. We additionally observe that post-processing maintains the asymmetric transferability topology. These findings expose a neglected dimension of attack evaluation and motivate transfer-aware protocols that jointly account for detector and generator variation. We will release our code, evaluation data, and trained attack models with a responsible-use statement.

Tabular foundation models (TFMs) are becoming a powerful alternative to conventional tabular learning, adapting to new datasets in context without task-specific training. Current architectures, however, face a trade-off: models that preserve cell-level representations are expensive because they repeatedly attend across rows and columns, while faster models collapse each row into a fixed embedding and lose this fine-grained state. We introduce Factorized Tabular Foundation Models (FTFM), which preserve cell representations while factorizing their computation through one contextual latent per row and one per column. Cells no longer attend to one another directly; instead, each cell is updated through the interaction of its row and column latents, reducing repeated attention while retaining feature-specific information. FTFM achieves competitive predictive performance in tabular classification on major benchmarks such as TabArena and BeyondArena.

Student feedback is a critical input for continuous improvement in higher education, yet traditional end-of-semester written surveys suffer from low response rates, superficial answers, satisficing, and limited actionability. Voice-based conversational AI agents can conduct structured interviews at scale, asking open-ended questions and probing follow-ups, and may capture richer qualitative feedback than written instruments. This research proposal asks what characteristics, depth, and strategic value feedback collected through voice-based conversational AI agents has in higher education, and whether such agents elicit detailed, honest, and actionable feedback that can support instructional improvement and institutional learning. Drawing on survey-mode-effects research, the Computers As Social Actors (CASA) paradigm, and dynamic capabilities theory, we develop three hypotheses on elaboration, social presence and disclosure, and actionability. A completed technical pilot at the Technical University of Munich established the feasibility of telephone-based AI interviews; the main field study is planned for the "Advanced Leadership Communication" course at MIT Sloan with approximately 60 students, using a convergent mixed-methods design and a four-dimensional coding framework for actionability (specificity, thematic richness, constructive orientation, contextual depth). The study extends CASA theory to evaluative educational settings, characterizes voice-based AI as a distinct feedback modality, and links micro-level feedback quality to institutional sensing capabilities.

Agents that run for a long time have a memory problem, and it is not the one most systems optimize for. Storing every trace and retrieving the nearest neighbors works early and quietly stops working later: the store fills up with failures the agent then repeats, and with facts that were true fifty turns ago. RSPM (Recursive Self-Pruning Memory) borrows the fix biology already found. At intervals the agent sleeps. It clusters its recent failures, distills each cluster into a short rule, and deletes the raw episodes it just learned from. Rules decay unless something reinforces them, and rules that recur across many contexts get promoted into general ones. We tested this on three settings with little in common: a difficulty-calibrated SWE-bench simulation, temporal consistency across 12 conversational benchmarks, and long-horizon QA on LoCoMo-10. All three improve, by 9.1, 11.9, and 7.6 points respectively. More interestingly, the two studies that instrument the store itself converge on nearly the same operating point without being tuned toward it: roughly 93% fewer retry loops and roughly 30% fewer stored entries. That hints that knowing what to forget is a general property of agent memory rather than a benchmark-specific trick.

The Internet of Things (IoT) gives large language model (LLM) agents access to physical observations, but one requested capability is usually offered by several devices, and the agent's plan does not say which one to call. Existing work retrieves relevant tools and checks plan feasibility, without learning which eligible device returns a useful observation. Unlike previous works, we study online tool grounding: binding an already specified tool call to one physical endpoint under service cost and device participation constraints. We propose a framework that learns which physical device should execute each agent-issued call. It formulates selection as a contextual bandit problem and combines adaptive context partitions with hierarchical endpoint trees, so that execution feedback is reused across related requests and devices. Selection balances task success against service cost while respecting device capability, authorization, battery reserves, and availability. Gaussian reward aggregation provides an optional mechanism for protecting individual feedback values. Theoretical results establish a finite-horizon regret bound and differential privacy for released reward statistics under fixed, reward-independent assignments. In a simulated loading-dock workload, the proposed router achieves higher step success and better cost-adjusted service, at lower simulated device energy and latency, than nearest-device, advertised-quality, UCB1, and LinUCB selection.

Bone scaffolds must reconcile a macroscale mechanical response with microscale geometric feasibility, yet conventional approaches to their design restrict attainable constitutive properties. This work presents a concurrent two-scale topology optimization framework; a conditional rectified flow generates periodic lattice microstructures, which are then refined in latent space by exact finite-element homogenization sensitivities. The resulting unit-cell geometries and their spatial assignments are jointly optimized under mechanical and feature-size constraints. In a distal rabbit femur benchmark, all eight flow-started refinements produce printable cells with maximum relative errors under 1\% across six prescribed engineering moduli. Constructing the implant from a single tiled cell reproduces intact whole-femur compliance within 0.1\% and regional strain energy within 0.35\% under four loading conditions. In human femoral reconstruction, multiscale concurrent optimization produces a cortical mid-shaft scaffold with porosity 63\% while satisfying the compliance constraint within 2\% under walking and stair-climbing loads. It achieves an 84\% porosity on a proximal cancellous scaffold with strain-energy deviations below 0.25\% under the same loads. Analysis of the learned representations reveals distinct architectural families in latent space. Further, the attention heads of the rectified-flow transformer specialize for spatial shifts -- learning the analogue of a finite-difference stencil -- and coordinate symmetries, with ablations showing their influence on generated topologies. These results establish a route to printable, mechanically constrained scaffold designs. The framework is general enough for extending to many other multiscale structural design applications including topology-microstructure co-design in metal additive manufacturing.

This study presents a large-scale quantitative analysis of 5.4 million video titles from xHamster, the world's fourth most visited pornography platform, to examine linguistic patterns used to describe different dimensions of identities (e.g., race and gender). Specifically, we investigate how nationalities, ethnic groups, and gender identities are portrayed in pornographic discourse, focusing on stereotypes and deviations from implicit norms of pornography. The results show the dominance of American pornography over other groups, which acts as the de facto norm for the adult industry. We identify persistent ethnic stereotypes reflecting colonial power dynamics, religious fetishization, and the sexualization of cultural taboos. The analysis reveals gendered representational differences across cultures, with a prevalent Madonna-Whore dichotomy in the portrayal of women and stereotypical depictions of men in physically dominant roles.

Completing complex tasks with language-model agents requires coordinating planning, execution, and review so the final result addresses the user's original objective. We present an agent execution harness that keeps the original objective separate from a revisable plan and an execution record. A planner assigns scoped tasks, workers produce outputs and artifacts, and a reviewer examines results against the original request before submitting or requesting additional work. Reusable domain-specific instructions guide these roles and are refined during construction using task inputs and execution trajectories, without official grading feedback. We evaluate complete harness configurations across terminal tasks, clinical responses, application building, and scientific questions. On Terminal-Bench 2.1, the harness passes 83 of 89 instances (93.26%); on HealthBench Professional, a Fable-configured system achieves a length-adjusted rubric score of 69.63 across 525 tasks. In a 53-task paired HealthBench comparison, the Fable-configured harness achieves an aggregate score 19.89 points higher than direct completion, despite a $-1.46$-point median task-level change; mean inference cost rises from \$0.29 to \$2.13 per task. We combine these outcomes with latency, documented instruction revisions, and reviewer decisions. Evaluating outcomes alongside execution trajectories and resource use clarifies how coordination supports task completion, where it introduces regressions, and at what cost.

Video Question Grounding (VideoQG) requires both accurate question answering and precise temporal localization. Existing supervised methods rely on costly temporal annotations, while existing zero-shot approaches are primarily designed for Video Question Answering (VQA) and are not optimized for temporal grounding. We propose DiffGround, an action-driven zero-shot framework built upon frozen Multimodal Large Language Models (MLLMs). Our framework first predicts an answer-relevant action cue that transforms the input question into an explicit semantic target for grounding. Action Cue converts a reasoningoriented question into an action-centric retrieval query. Guided by the action cue, a Vision-Language Model (VLM) efficiently selects informative keyframes for reasoning. Finally, we introduce Semantic Diff Refinement, a training-free temporal refinement method that exploits semantic differences between neighboring SigLIP2 embeddings to recover complete action segments from sparse predictions. Extensive experiments on NExT-GQA and STAR show that DiffGround consistently outperforms existing zero-shot and supervised methods, establishing a new state of the art for Video Question Grounding.

Recent work suggests that activation dynamics--how representations change during inference--might be more revealing about language model behavior than static readouts from individual layers. In this work, we investigate the layer-wise evolution of representations of factuality, using true and false statements when the model is prompted to either lie or to be honest. We report a geometric phenomenon in which the factuality directions extracted from the lying and honest conditions appear to rotate away from each other across layers. We find that such layer-wise rotation also occurs in a task that requires the model to take different visual perspectives in the 3D world, suggesting that the rotation phenomenon might be indicative of manipulating conceptual spaces in general, rather than unique to factuality. We then investigate the "real time" transformation of conceptual representations on a trial-to-trial basis. We show that it is possible to make strong predictions about the model's target task and behavior from trajectories alone. In addition, diverging points in single-trial trajectories correspond to the onset of rotation in aggregated-trial analysis. Finally, we show that both rotation and divergent trajectories occur in more naturalistic settings in which models are not overtly prompted to "lie" but rather report false statements in the course of completing a different task.

While retrieval-augmented language models increasingly serve as literature assistants in the physical sciences, existing benchmarks evaluate them primarily on qualitative text overlap or through model-based scoring rather than against verified experimental ground truth. In experimental materials disciplines such as perovskite photovoltaics, however, the most critical inquiries are inherently quantitative and aggregate measurements across dozens or hundreds of papers. Here, we present a quantitative question-answering benchmark where all 755 ground-truth answers are computed directly from 45,441 measured device records (40,889 human-curated records through 2021 and 4,552 records extracted from 2022-2026 literature). Benchmarking six retrieval architectures over 158,274 passages from 25,662 papers using an open 14B model on a single 24 GB GPU, standard passage retrieval tops out at 16-23% accuracy (compared to 1.6-6.5% zero-shot). In contrast, equipping the language model with read-only SQL execution over the structured device table elevates accuracy to 51.5% on historical and 50.8% on prospective literature -- a 33.01.8 percentage-point advantage over hybrid text retrieval. Aggregate queries spanning extrema, distributions, and filtered record counts prove virtually unsolvable via text retrieval alone, demonstrating that systematic table extraction is an operational prerequisite for literature-scale quantitative inquiry.

Identifying which visual interventions interact in a vision-language model requires comparing their joint and separate effects. The resulting four-output difference can be small enough for rounding error to obscure it. We introduce the Finite Interaction Lens (FIL), a mixed-precision method for evaluating finite intervention families. At each selected linear projection, FIL decomposes each candidate's input into a native state, two single-intervention components, and a full-dimensional mixed residual. It projects the single-intervention components at higher precision and reuses the same numerical contributions in the joint candidates, while retaining every mixed residual and nonlinear trajectory. We derive a local cancellation identity and a conditional bound on error propagation through a smooth nonlinearity. In one MLP of Qwen3-VL-4B-Instruct, FIL reduces final interaction MSE by 94.4% on 16 synthetic families and by 99.85% and 99.35% in two separate four-image COCO studies, relative to a control that has the same nominal BF16-product budget and higher-precision single interventions but projects each joint candidate independently. One image accounts for most of the absolute gain in each COCO study. Higher-accuracy multi-product BF16 arithmetic on all candidates gives lower aggregate error at similar total query times. Single-product BF16 already recovers the reference choices in both candidate-selection tasks we tested. On these development inputs, FIL improves interaction fidelity at a fixed nominal precision budget. Its practical value depends on a task that needs the extra fidelity.

We consider graph search models, in which an invisible, omniscient and fast intruder can hide anywhere in the graph, and the goal is to design a \emph{search strategy} composed of a sequence of movements of \emph{searchers} that guarantees the capture of the intruder regardless of its actions. Our point of departure is the connected search model in which the cleared subgraph is always connected during the search process. In this paper, we initiate the study of the subgraph searching problems which arise from the following question: what happens if the number of searchers available is not sufficient to clear a given graph $G$? We propose an approach in which one wants to clear in a connected way as large subgraph as possible with the given $k$ searchers. In this context we study two problems: one problem, called the \emph{subgraph search}, is to find the maximum-sized subgraph whose search number is at most $k$, and the other one, called the \emph{protected subgraph search}, is to perform a partial search on $G$ leading to as many cleared edges as possible. Informally speaking, the difference lies in the fact that in the latter a recontamination may come from an edge never to be cleared and hence the partial search strategy needs to avoid such recontaminations. Having proved that both problems are computationally hard in general, we give efficient algorithms for $n$-vertex trees: an $\bigo(\min\{kn, n \log n\})$ time algorithm for the subgraph search and an $\bigo(\min\{2^{6k}kn, n^{8.571} \log n\})$ time algorithm for the protected subgraph search.