Papers

Adversarial training is one of the most reliable defenses against adversarial attacks, but its high computational cost must generally be paid anew for each task. Robust foundation models offer a promising alternative: adversarially pretrain a model once and then transfer its robustness to downstream tasks through lightweight adaptation. However, a fundamental question remains open: can robustness acquired during pretraining transfer to unseen tasks without further adversarial training? In this study, we answer this question affirmatively. A single model adversarially pretrained at scale can achieve optimal robustness on new tasks without additional task-specific training. Specifically, we show that, for a family of Gaussian-mixture classification tasks, a sufficiently deep linear transformer adversarially trained across tasks can asymptotically attain the robust Bayes error on previously unseen tasks through in-context learning from clean demonstrations. By contrast, a standardly trained model cannot. We further analyze convergence under gradient flow, an accuracy–robustness trade-off, and demonstration complexity.

Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand. We study where this evidence-to-action chain breaks as agents move from deciding whether to act to executing single actions and dependent workflows. Across ten model-harness configurations, strong static action assessment can coexist with much weaker interactive execution. Failures often begin before execution: agents stop with incomplete investigation or act before required evidence is established. Once required evidence is obtained, single-action execution is usually reliable, while multi-action workflows additionally expose unresolved prerequisites and incomplete execution. For this analysis, we introduce SafeActBench, comprising 656 cases across six operational domains and five protocols that progress from static action judgment and investigated non-action to single- and multi-action workflows. A provenance-bound Evidence Ledger and deterministic trajectory evaluator track what information was established, when actions occurred, and whether downstream dependencies were satisfied. These results show that failures arise not only from missing information, but also from how agents use established evidence when deciding and executing actions.

Collaborative manipulation requires robots to perform complementary actions as interactions unfold. We study single-policy decentralized collaboration: every robot runs the same policy from its visual observations and proprioception, without task prompts, identity labels, or inter-robot messages. The challenge is to learn complementary team behaviors within shared parameters and select appropriate actions from each robot's local observations. We introduce CoRE, which learns Collaboration-Role Experts from pooled multi-task, multi-robot demonstrations. Fused appearance and geometry provide local interaction evidence. Query-conditioned cross-attention experts provide adaptable prediction paths, which a local router combines at each action-chunk position. During training, an action-expert alignment loss supervises expert selection using relative forced-route prediction errors against demonstrations under fixed inputs, without role labels. Across simulation benchmarks, CoRE achieves the highest average performance among evaluated decentralized methods. Physical experiments demonstrate effective collaboration across diverse manipulation tasks and robustness to partner delays and slowdowns. Project page: https://aus.bot/research/core/.

Real-world reasoning rarely reduces to static question answering: agents must actively gather information from tools and sensors that are often noisy and unreliable. Yet most existing active reasoning benchmarks assume that environmental feedback is trustworthy, or introduce noise without exposing an explicit, calibrated uncertainty signal, leaving open how LLMs should reason when the evidence itself is uncertain. We introduce VisualNoiseQA, a novel benchmark for active reasoning under noisy visual feedback. A text-only LLM must solve VQA problems by iteratively querying a fixed, off-the-shelf VLM treated as a stochastic visual sensor. For each query, we draw multiple samples and expose an empirical uncertainty signal via self-consistency, enabling the reasoner to probe from different angles and decide what to ask next and when to stop. Our construction is automatic and scalable: starting from diverse VQA sources and two noisy VLMs, we retain only questions where the sensor is inconsistent yet human-solvable. We evaluate multiple LLM reasoners on 1,000 instances spanning perception, chart understanding, and knowledge-intensive reasoning. VisualNoiseQA thus provides a controlled playground to study how different LLMs exploit uncertainty signals for robust reasoning.

Semi-structured tables encode rich semantic information through diverse layout elements, such as hierarchical row and column headers. Answering numerical questions over such data is challenging because it requires a joint effort of accurate structural understanding of tables, question uncertainty resolution, and complex query intents interpretation. Existing methods are rarely effective in handling the above multifaceted challenges in a holistic manner; moreover, they rely on LLMs without considering financial implications. We propose SemiBonsai, a cost-effective, LLM-powered framework featuring: (1) a Multiway Layered Table Structurer that converts a table into a multiway layered tree organized by structural elements; (2) a Context-Aware Uncertainty Resolver that grounds underspecified phrases to context-coherent structural elements; (3) a Plan-Guided Reasoner that decomposes complex query intents into a query plan for reasoning; (4) a Budget-Aware Bandit Router that learns per-instance LLM utilities and optimizes LLM selection under fixed monetary constraints. Experiments on three benchmarks show that SemiBonsai improves both QA effectiveness and cost-effectiveness. Notably, the Table Structurer improves answer accuracy by 46% over alternative tree construction methods, and Router achieves a 5% average relative gain in answer accuracy compared with existing routing methods across budgets.

Local multipath rejection can erase evidence needed for cooperative sensing. We propose staged anchored recovery: preserve strong-only confirmations, then query compatible weak reports using unused strong anchors. For any number of sensing nodes, we prove lossless residual screening and derive corroboration and bidirectional cost laws. In 1,024 five-UAV drops, recovery adds 30 matched targets and three false outputs over strict consensus, matching one-pass anchored confirmation's detection counts while reducing weak uploads by 97.4%. Equal-sized cue/report records yield 4.5% less payload than uploading all eligible reports. Independent validation recovers four additional targets with no observed false outputs.

Pose error in precision assembly can corrupt not only what a robot observes but also the coordinate frame in which it acts. On the FORGE benchmark, the official state-based policy succeeds in 97% to 99% of episodes with the true pose but only 32% to 60% at the benchmark's $\sigma=5$ mm pose-noise setting. The same estimated pose enters the observation and anchors the action frame, making the offset unidentifiable from proprioceptive state alone before contact. We supply this missing information during training in two ways. A privileged teacher observes the offset in simulation, while clean demonstrations can instead be relabelled into the displaced frame in closed form. The deployed student is trained with behaviour cloning followed by one DAgger round and receives only noisy state, a raw wrench window, and two RGB cameras at test time. On the unmodified FORGE tasks, the teacher-route student maintains 92% to 99% success across $\sigma=0$ to 5 mm, while six non-privileged baselines fall to 2% to 80%. A matched behaviour-cloning experiment isolates the source of this robustness. With the same student architecture, data budget, and training procedure, demonstrations generated without offset access yield only 31.5% success at $\sigma=5$ mm, whereas privileged and relabelled demonstrations reach 88.7% and 95.8%. Deployed zero-shot on a Franka, the student reaches 83.3% pooled success at $\sigma=5$ mm against 34.4% for the strongest state-based policy. Deployable sensing alone is insufficient. Robustness requires supervision that encodes compensation for the latent frame offset.Project page: https://drychang.github.io/displaced-frame/

Personal digital activity increasingly shapes online experiences, yet few users have been educated regarding the processes transforming raw interactions into personalized suggestions. We developed an education artifact that illustratively simulates how AI leverages users' digital activities to shape online recommendations (e.g., ads). Our artifact processes users' digital activity using a locally-hosted LLM to generate user profiles of their inferred interests and personalized recommendations. A three-layered Sankey diagram maps data sources through inferred interests to personalized recommendations. Interactive filters enable users to explore how different combinations of data sources influence personalized outcomes. This paper describes the artifact and its educational value, and reports findings of a pilot think-aloud study with six young adults. We find that the artifact effectively taught participants the conceptual relationship between digital activities and personalized recommendations. While this did lead participants to develop privacy awareness, they anticipated minimal behavior change due to the perceived unavoidability of platform participation.

Optimized kernels such as FlashAttention and FlashDecoding are crucial for accelerating today's large models. Most of them are handwritten by experts because existing ML compilers cannot match their efficiency. Producing such kernels requires fusing computations with multiple reductions, which requires both algebraic transformation of the computation graph and operator scheduling of the transformed graph. Unfortunately, searching the two jointly yields a space too large to navigate. We propose Cleave, an ML compiler built on symbolic decoupling: Cleave discovers transformations by performing superoptimization on a graph with symbolic shapes, and then schedules each resulting graph on concrete shapes. Representing shapes as symbols makes equivalence checking cheap and lets a new Split operator, with a symbolic split count, parallelize along a reduction dimension. Cleave's scheduler fuses graphs with multiple reductions through iterative tiling and horizontal fusion. Evaluation on common LLM subgraphs shows that Cleave generates kernels up to 2.8x faster than the best baseline (1.6x on average) and reduces compilation time by 5.9x on average compared to Mirage. For dynamic workloads captured from production serving traces, Cleave compiles each operator once and achieves geometric mean speedups of 1.4x and 1.7x over FlashInfer's handwritten FA2 and FA3 backends. Cleave's code is available at: https://github.com/nyu-systems/cleave

We study the online calibration of multidimensional forecasts over an arbitrary convex set $Y\subseteq\mathbb{R}^d$ relative to an arbitrary error norm $\|\cdot\|_{L}$. For forecasting $d$ binary outcomes simultaneously ($Y=[0,1]^d$), we give the first algorithm that achieves $\varepsilon$-calibration in a number of rounds that is polynomial in $d$ for every fixed accuracy. It requires $d^{O(1/\varepsilon)}$ rounds, exponentially improving the dimension dependence of previous bounds. For multi-class forecasting ($Y=\Delta_d$), we obtain the same $d^{O(1/\varepsilon)}$ rate, improving the $d^{\widetilde{O}(1/\varepsilon^2)}$ bounds of Peng and Fishelson et al. Our algorithm is simple: on each round, it outputs a harmonically weighted distribution over harmonically smoothed past outcomes. The same algorithm works for every forecast set and norm. More generally, it achieves $\varepsilon$-calibration after $\exp(O(\gamma(Y,L)/\varepsilon))$ rounds, where $\gamma(Y,L)$ is a geometric parameter defined by a matrix discrepancy problem. The harmonic weights are motivated by the fact that the discrete Hilbert transform matrix achieves the optimal discrepancy up to a universal constant, simultaneously for every $L$. This optimality result may be of independent interest.

Post-discharge risk prediction from electronic health records (EHRs) is difficult because many dependencies that link discharge-time observations to downstream complications, such as comorbidity cascades and drug-disease interactions, are absent from the record. External medical knowledge graphs (KGs) can supply these missing dependencies, but tracing them demands three properties: KG exploration must remain cost-bounded, retrieved evidence must be differentiated by source quality, and the resulting rationale must be citable for retrospective review. Large language models (LLMs) can plan and verify over structured evidence, making them natural candidates for KG reasoning, but existing LLM-based methods do not satisfy these three properties jointly. In this paper, we propose BAR, a Budget-Aware LLM Reasoning framework over medical KGs with three contributions. First, BAR refines the raw KG into disease-specific evidence graphs whose edges carry support scores and provenance records, turning the KG into a quality-annotated reasoning space rather than a static feature source. Second, an LLM then reasons over this graph through a plan-navigate-verify loop that decomposes the question into steps, retrieves evidence under a patient-specific budget, and revises when verification fails. Third, a reasoning policy is trained with a reward that compares predictions with and without acquired evidence, combined with acquisition cost and citation-integrity terms. Across 8 diseases and 3 prediction horizons on MIMIC-III and MIMIC-IV, BAR improves AUPRC by 3.4 points over the strongest baseline, raises citation precision from 59.8% to 77.9%, and consumes only 62-65% of the budget cap.

We study fair multi-armed bandits under the Nash Social Welfare (NSW) objective, which measures performance via the geometric mean of accumulated rewards. Existing work defines Nash regret as $\mathrm{NR}_T = \mu^\star - (\prod_{t=1}^T \mathbb{E}\mu_{I_t})^{1/T}$, where $\mu_{I_t}$ is the mean reward of the recommended arm $I_t$ and $T$ is the horizon. Since it applies the geometric mean to per-round marginal expectations, it ignores the joint distribution of rewards across rounds, leaving the NSW fairness motivation unaddressed at the trajectory level. We propose trajectory-wise Nash regret $\widetilde{\mathrm{NR}}_T = \mu^\star - \mathbb{E}[(\prod_{t=1}^T \mu_{I_t})^{1/T}]$, which computes the geometric mean over complete sample paths before taking expectations, capturing NSW fairness more faithfully. By Jensen's inequality, $\widetilde{\mathrm{NR}}_T \geq \mathrm{NR}_T$, making it a strictly stronger metric. We also introduce high probability Nash regret $\widehat{\mathrm{NR}}_T = \mu^\star - (\prod_t \mu_{I_t})^{1/T}$, giving the first high probability regret bounds in fair bandits. Our two-phase algorithm, Round Robin Nash Confidence Bound (RR-NCB), combines round robin exploration with a Nash confidence bound index policy. We show $\widetilde{\mathrm{NR}}_T \leq \widetilde{\mathcal{O}}(\sqrt{k\log T/T})$ and, with probability $1-\delta$, $\widehat{\mathrm{NR}}_T \leq \widetilde{\mathcal{O}}(\sqrt{k\log(kT/\delta)/T})$, matching the optimal $\widetilde{\mathcal{O}}(\sqrt{k/T})$ rate despite the stronger metrics. Optimality follows from a lower bound via AM-GM and standard $k$-armed bandit minimax arguments. Simulations validate our theory.

When several people share a 3D virtual room with an LLM agent, the agent must decide not only what to say, but whether an utterance was addressed to it and, if accessible, what profile information about the others present it may use. To study both problems, we construct LookAway, a controlled corpus of 40 sessions involving 80 distinct personas and an LLM agent (1,200 turns), including ambiguous-addressee turns in which speaker orientation agrees or conflicts with the intended addressee. Across three open-weight large language models and five conditions varying which profiles the agent sees and whether it is told where each person faces (18,000 decisions), adding speaker orientation increased addressee accuracy from 56% to 99.5% when orientation was congruent, but when the speaker faced someone other than the addressee, two of the models went by where the speaker faced on more than 85% of those turns. Warning one model that orientation could be misleading reduced this only modestly. A browser-based 3D demonstrator shows the effect live: the same sentence gets an answer when the speaker faces the agent and silence when they face the other person. Providing both personas' profiles improved responses about the person being asked about, but also increased the use of profile attributes not revealed in the shared conversation, reaching 45.3% of answers for one model. More context thus improves multi-user interaction but also leads to oversharing, so shared LLM agents need mechanisms for weighing spatial cues and controlling when user-specific information enters a response.

Dense retrieval models are typically trained with contrastive objectives that learn effective representations but do not directly optimize retrieval metrics or downstream task performance. To address this problem, we introduce RELER (REinforcement LEarning for Retrieval), a reinforcement learning framework that enables existing embedding models to learn to retrieve directly in embedding space and align to task-specific rewards. We train RELER by sampling unit-length query and document embedding actions from von Mises-Fisher (vMF) distributions centered on normalized encoder outputs, scoring the resulting retrieval or downstream outcomes as rewards, and updating the encoder with REINFORCE using a leave-one-out baseline (RLOO). As exploration in the high-dimensional embedding space is prone to sampling noise, we further propose conditional-mean projection (CMP), which projects each sampled embedding onto the low-dimensional subspace spanned by its encoder output and the candidate embeddings it is compared against, reducing noise in the policy gradient while preserving its expectation. We evaluate RELER on BRIGHT, a benchmark with reasoning-intensive queries that remain challenging for existing embedding models. RELER consistently outperforms InfoNCE and LambdaLoss in average nDCG@10 when post-training BGE-M3 and Qwen3-Embedding backbones. We further evaluate downstream utility through retrieval-augmented generation (RAG), where we adapt only the query encoder while keeping the document index and generator fixed. Across seven QA datasets, jointly optimizing retrieval and answer rewards improves both average retrieval performance and answer quality in RAG.

Typed decision models answer a declared question without generating text: a decision head returns a probability for each of the declared options in a single forward pass. A single pass is fast, intuitive System 1 thinking. We study what lies between one pass and generated reasoning: looping, in which the same layers are recursively applied several times before one typed readout. Each loop lets the model revise its hidden state before it commits to an answer, without generating a token; we call this System 1.5 thinking. We propose SanSi, which turns a pre-trained looped language model into a typed decision model. The option probabilities are read after every loop, and every loop is trained with a proper scoring rule, so that one model serves every budget from one loop to eight in a single run. On 10,027 test decisions from 59 sources, SanSi reaches 72.0% accuracy: 13.5 points above a non-looped model of the same shape trained with the same recipe, 5.3 points above a newer non-looped model of its size, and 1.8 points below one with three times the parameters. On two depth-controlled tasks, loops extend the solvable depth beyond the depths seen in training, where the larger single-pass model fails. Used as the judge for policy optimization with reinforcement learning, without gold answers, SanSi raises the share of the generator's correct answers from 37% to 52%.

Visual tokens are a major source of inference cost in vision-language models, yet simple image downsampling remains a surprisingly strong compression baseline. This raises a complementary question: under a fixed token budget, where should visual fidelity be preserved? We introduce Foveated Compression, which encodes a full-resolution image once and represents it with a mixture of native- and compressed-resolution visual tokens. A behaviorally self-distilled Foveated Merger compresses local visual tokens while preserving compatibility with their native counterparts, and a lightweight Foveated Selector chooses one of nine spatial cells to retain at native resolution using exhaustive budget-matched intervention supervision. At 11.11% visual tokens, uniform Foveated Compression shows no significant paired difference from iso-token downsampling. At 20.99%, the learned selector significantly outperforms random and fixed allocation, but remains below strong whole-image resizing, showing that localized fidelity is not universally preferable. A budget-matched region-choice oracle reaches 82.73 macro accuracy versus 69.61 for the learned selector, revealing substantial headroom within the same spatial action space. Matched probing further shows that signals predicting when compression breaks the answer are substantially more accessible after language-model computation than to the lightweight prefill-free selector. These results expose complementary bottlenecks in region selection and compressed-region fidelity.

As human–AI interactions become more conversational, full-duplex speech language models capable of natural real-time dialogue are growing in importance. Beyond generating appropriate responses, these models must coordinate turn-taking, backchanneling, and floor management in real time. Reinforcement learning (RL) provides a way to refine these behaviors through direct feedback on interaction outcomes. However, existing RL methods either apply timing feedback to a token policy or optimize semantic content, leaving the joint improvement of timing and content unresolved. We introduce HiPLEX, an RL framework that factorizes a pretrained full-duplex text policy into a control policy that decides when to emit content and a conditional content policy that decides what to emit. The first factor selects among 'pad', 'epad', and 'con'. The second selects a token only when 'con' is chosen. This hierarchy describes conditional actions within each frame and uses the model's existing text head. We route timing advantages to the token-group factor through event-causal masks derived from generated speech episodes, and route an LLM-judge semantic advantage to the conditional content factor. Across three Moshi seeds on Full-Duplex-Bench v1, HiPLEX reduces takeover rates during natural user pauses and backchannel opportunities, and shortens post-interruption response latency relative to GRPO, while maintaining comparable judged interruption-response quality. On Moshi and PersonaPlex, HiPLEX better matches pooled human turn-timing and backchannel-rate marginals than GRPO.

Videofluoroscopic Swallowing Study (VFSS) is one of the gold standard for diagnosing swallowing disorders, providing dynamic X-ray imaging of the swallowing process. Automated kinematic analysis in VFSS relies fundamentally on precise anatomical keypoint localization. However, existing studies focus on limited keypoints (e.g., cervical vertebrae or the hyoid) and overlook critical regions such as the soft palate, while annotating only active swallowing segments and ignoring abundant non-swallowing data, resulting in poor data efficiency. Moreover, leveraging this unlabeled data via standard semi-supervised learning is suboptimal, as generic methods are prone to spatial bias. In medical X-rays with fixed layouts, models tend to memorize absolute coordinates rather than understanding anatomical structures. To tackle these challenges, we introduce VFSSKep, a novel dataset that extends annotations to the soft palate and incorporates large-scale unlabeled data. We further propose S$^3$KL, a Structure-aware Semi-Supervised Keypoint Localization framework designed to overcome spatial bias. It integrates a Structure-Aware Learning strategy to extract high-resolution structural cues for structure-aware representation learning, and a Structural Representation Consistency Learning strategy with block shuffling to enforce invariant structural recognition. Experiments show our method achieves state-of-the-art semi-supervised performance, even with unlabeled and 25% labeled data surpassing fully supervised learning with 100% labeled data. Code and data will be made publicly available at: https://github.com/kaai520/S3KL.

Modern voice assistants may be shared by multiple users and should be able to answer questions about earlier conversations such as "When did I originally plan to leave?" or adapt their behavior to individual users based on past interactions. This requires more than retrieving a topically similar passage: the assistant must identify the current speaker, recover the relevant past state, and distinguish it from later revisions. We present PERSIST, a persistent memory system for multi-session, multi-speaker spoken dialogue that explicitly models Who, What, and When. PERSIST structures cross-session histories into readable event records and retrieves them with a 3W joint scoring mechanism that combines semantic content, acoustic speaker identity, and temporal state. For real-time full-duplex interaction, PERSIST further reuses intermediate representations from the dialogue backbone, avoiding query-audio re-encoding and reducing retrieval latency from 578.42 ms to 7.03 ms. We also introduce SpokenTrace, a diagnostic benchmark that factorizes evaluation along memory tasks and speaker-query types, exposing failures in recall, speaker attribution, and temporal-state tracking. On SpokenTrace, PERSIST achieves 85.08% end-to-end task accuracy and improves all-support EM@3 from 49.01% with BGE-large to 82.10%.

Safety alignment in Large Language Models (LLMs) remains vulnerable to backdoor attacks. Existing LLM backdoors are almost all input-centric: activation depends on explicit trigger patterns in the user input, so modern guardrails are built to sanitize the input space. We challenge this assumption with a novel answer-side backdoor for multi-turn dialogue. Instead of inserting the trigger into the input, the adversary uses a benign first-turn prompt to naturally induce the model to generate a specific, seemingly innocuous word. Once merged into the dialogue history, this self-generated word becomes the trigger. When a later harmful query arrives, the model detects its own trigger and bypasses its safety refusal, while the user input stays perfectly clean. Across four LLMs, our attack reaches near-perfect Attack Success Rates, approaching 100% at only a 5% poisoning rate, while preserving general utility and clean-input safety, and it evades mainstream input-centric defenses. Representation-level analysis shows that the self-generated trigger consistently suppresses the model's refusal signal, exposing a critical blind spot in current LLM defenses.

Activation steering provides a lightweight and flexible way to control large language model (LLM) behavior. However, effective steering requires more than inducing the intended behavior: it should also limit unintended changes and remain robust across inputs and training data. Existing evaluations cover these dimensions only in fragments. As a result, the trade-offs between efficacy and side effects have not been systematically characterized. We introduce SteerScope, a two-axis, multi-dimensional evaluation suite that jointly characterizes steering outcomes and method properties through 15 metrics. We score target efficacy and side effects on language quality, task capabilities, and safety and reliability, and further assess generalization and data dependence through steering-specific metrics for sample efficiency and sample sensitivity. Rather than comparing methods at a single operating point, we characterize the trade-offs between efficacy and side effects. Under matched models, tasks, and evaluation protocols, we benchmark 23 methods spanning 4 families, including prompting, LoRA, and SFT as baseline methods, and release the suite as an extensible codebase. We find that current activation steering methods do not yet surpass the Prompt Steering baseline in their overall balance between steering efficacy and side effects: across both model scales, no evaluated activation steering method achieves higher efficacy without incurring greater composite side effects. We further uncover a consistent coupling between steering efficacy and side effects. Under OOD prompts, target efficacy is often preserved, whereas side effects tend to become more pronounced, particularly through declines in instruction relevance and fluency. Methods also exhibit sharply different sample-efficiency profiles.

We study ridge regression from exactly $s$ distinct rows of a fixed design. Responses are fixed, and only the subset is random. The determinant law and selected ridge fit share one positive definite penalty. Established mean identities and exponential-family duality give the unique penalty that matches a prescribed full-data ridge fit in expectation. It exists exactly when $s$ exceeds the target's effective dimension. Our main result concerns centered covariance risk normalized by full-data penalized loss. For balanced signed coordinate replicas, a strict sector inequality gives the sharp risk and all maximizing responses at every budget from the dimension to one below the row count. This holds for any nonzero positive semidefinite query. With the target and query fixed, the maximizing response space is unchanged across these budgets. For general designs, we characterize attainment of a leave-one-out envelope. For existing real equiangular tight frames, flat row query energy characterizes when every nonzero residual response maximizes at two deletions. At three deletions, we give the sharp risk and complete maximizing space for isotropic queries, using unequal triangle weights. The balanced geometry yields a same-sample unbiased ridge–Horvitz–Thompson mixture with lower sharp risk and an exact mean-share improvement boundary. Under full recalibration after feature changes, we prove quadratic regret from searching the complete old maximizing space and a query-uniform bound on the mixture's risk gain. The strongest sector inequalities have exact computer-assisted proofs.

Multi-reference image generation requires preserving the appearance of multiple subjects while composing them into a coherent scene. However, existing diffusion transformers commonly encode references as dense visual token grids and jointly process them with global attention, making conditioning increasingly expensive as the number and resolution of references grow. We present RefRoute, a framework that addresses both reference representation cost and attention overhead through two complementary mechanisms. Compact residual conditioning combines low-resolution latent tokens with lightweight residual features extracted from full-resolution pixels, reducing reference token counts while retaining fine-grained appearance cues. Condition routing and attention routing align reference tokens with their assigned target regions and restrict cross-reference interactions, while allowing selective reference access beyond region boundaries for scene integration. We further introduce RefRoute-Data for training many-reference generation models and ManyRef100, a benchmark spanning human, object, and mixed compositions with 10-17 references. After many-reference fine-tuning, RefRoute achieves an overall Weighted-Ref-VIEScore of 36.06 on ManyRef100, compared with 8.88 for FLUX.2-Klein-9B. Separate inference-cost evaluations show substantially slower latency growth as the reference count increases: at 16 references, our 50-step and 4-step configurations achieve $18.3\times$ and $14.2\times$ speedups over their corresponding FLUX baselines, respectively. These results establish compact reference representations and spatially routed attention as an effective approach to scalable many-reference image generation.

Transformers have exhibited impressive empirical success across various domains, but their theoretical foundations remain less developed. This work constitutes a mathematical study of the measure-to-measure operators defined by transformers. We show that transformers map sub-Gaussian inputs to sub-Gaussian outputs; this ensures that taking arbitrary-length compositions of the softmax operator is well-defined. We then show that transformers are Hölder continuous with respect to the 1-Wasserstein distance on appropriate spaces of sub-Gaussian inputs. This allows us to establish estimates on the error propagation along a transformer between a sub-Gaussian input and its empirical approximation. We also study a mean-field analog of the cross-attention mechanism, which is an operator from a pair of probability measures to a single probability measure. We show that cross-attention exhibits different Hölder regularity and sample-complexity in its two input arguments. Last, we apply our results to deduce approximation guarantees for measure-to-measure transformers. Together, these results provide a firm stability and finite-sample theory for transformers on sub-Gaussian data.

We present JevCascade, an inference-time compute allocation framework for prefill-only decision models, built on readout stability and zero-label restricted readout prediction. Prefill-only decision models score every candidate in a menu in a single forward pass and never decode, which makes one call one to two orders of magnitude cheaper than a same-scale generative language model. We show that this read-out structure carries a testable property. When an intervention changes only the menu and leaves the input text fixed, the post-intervention accuracy is already determined by the cached first-pass distribution. The restriction predictor restricts the pass-1 probabilities to the menu, renormalizes and reads off the argmax, with no labels and no second pass. Across six model families, ten core datasets (sixteen in total) and two task types, menu-only interventions are predicted to within 4.2 points, exactly for one family. The property is ordinal, not probabilistic, since a probability-level variant of the same estimator errs by 21.0 points, and class-specific, since same-scale generative models err by 1.6 to 15.8 points and degrade as they grow. It makes inference-time compute decidable before deployment. Uniform extra passes buy calibration but almost no accuracy; at matched cost a confidence cascade beats every scheme that re-asks the same model, and curating the menu beats enlarging the model, a 0.8B model on a curated 5-candidate menu reaching 95.4% on CLINC150 against 80.0% for a 4B model on the full 150-label menu.

A memory system can retire an obsolete value and later restore it merely because the same old statement appears again. A re-read of an old source and a genuine revert can produce the same observed sequence of values while requiring opposite current answers. We study this ambiguity on 130 extractor-selected atomic transitions derived from software fixes. In the ordinary transition condition, identity-based temporal memory reaches 98.5% model-judged accuracy with zero observed errors under a literal stale-value proxy. Appending a verbatim re-read of the old statement reduces accuracy to 10.8% and raises the stale-value rate to 88.5%. A guard that refuses to reactivate a previously retired value restores accuracy to 97.7% and reduces that rate to 0.8% in this constructed re-read condition. The guard cannot also recognize a legitimate revert without additional change provenance. Two supporting studies examine exposing retired history to the answer model and supplying current source for changed behavior. An exploratory extraction study over 707 software fixes provides scope context, not a universal coverage estimate. The design implication is to distinguish an observation of a value from evidence that the value changed. Selected inputs, aggregate-only answer records, related-family judges and a post-failure guard evaluation limit the conclusions to the retained experiments.

Surface electromyography (sEMG) provides a wearable, noninvasive interface to neuromuscular activity for movement decoding and human-computer interaction. Population-scale decoding remains difficult because the relationship between sEMG and neuromuscular activity varies across users and sessions, while task-relevant dynamics span channels and multiple timescales. Learning waveform-to-output mappings from task labels leaves the distinction between recording variability and coordinated motor activity implicit. We introduce the Neuromotor Hierarchy Network (NHN), which learns a compact latent neuromotor state from task supervision to represent task-relevant neuromuscular coordination. NHN constructs this latent state through a hierarchy inspired by neuromotor organization. It adapts recording statistics while preserving relative intensity. Its spatiotemporal encoder uses parameter-efficient channel interactions and modulates features with multi-timescale history. The resulting features yield candidate activations of learned motor primitives, which are temporally integrated and continuously weighted to form the state. Theoretical analysis characterizes the efficiency, temporal behavior, and optimization of NHN's core mechanisms. We evaluate the architecture for both continuous hand-pose estimation on emg2pose and touch-typing recognition on emg2qwerty. On emg2pose, NHN reduces user-averaged angular error by 0.52% to 2.84% across all three generalization splits in both Regression and Tracking relative to Hadidi et al.'s best task-specific variants, using 48.42% to 48.51% fewer parameters. On emg2qwerty, NHN reduces beam-search character error rate by 19.40% zero-shot and 30.42% after fine-tuning relative to SplashNet-Upscale, using 65.86% fewer parameters. Physiology-guided inference of a latent neuromotor state supports parameter-efficient sEMG decoding.

Existing molecular and materials learning approaches often rely on a limited set of structural representations, which may capture only selected aspects of complex three-dimensional structure. Here, we introduce mathematical invariant-enabled topological neural networks (MITNNs), a framework that represents complex structures through multiple complementary mathematical views and integrates them with topological neural architectures. MITNNs combine multiscale invariants from topology, spectral theory, commutative algebra, differential geometry, and discrete curvature, capturing complementary structural information from the same system. Systematic invariant-subset, architecture-subset, and ensemble analyses show that predictive performance depends on how mathematical representations and neural architectures are paired, with selected combinations outperforming individual models and the aggregation of all available components. Across protein-ligand binding, metal-organic framework properties, mutation-induced protein solubility, and molecular toxicity prediction, MITNN consistently outperforms existing methods. These results establish MITNN as a mathematically multimodal framework for scientific machine learning.

Accurate nuclei instance segmentation is essential for quantitative renal pathology, yet general-purpose models often struggle with low contrast, dense nuclei, complex morphology, and strong background staining. In this work, we extended a human-in-the-loop framework by combining 5,901 foundation-model-generated pseudo-labels from well-segmented cases (Easy), 860 newly expert-annotated unresolved challenging cases (Medium), and 198 expert-annotated consensus failure cases (Hard). These annotations, spanning different levels of segmentation difficulty, enabled the systematic evaluation of seven single-source and mixed-source fine-tuning strategies across nine cell segmentation model configurations. Fine-tuning improved all models, with Medium data included in seven of the nine best-performing strategies. LSP-DETR achieved the highest F1 score of 0.8725 with Hard-only fine-tuning, while StarDist showed the largest improvement, increasing from 0.7380 to 0.8332 with Medium-only fine-tuning. These findings show that annotations spanning multiple difficulty levels support effective model adaptation, although the optimal annotation composition remains model dependent.

Recommender systems learn from observed user-item interactions, but explicit negative feedback is often unavailable. Since deep learning models require negative signals for training, negative sampling methods typically treat selected unobserved interactions as negatives. However, a missing interaction does not explain why a user is uninterested in an item or whether there is sufficient evidence to label it negative. This is especially important in business recommendation, where negative signals should be interpretable and aligned with business objectives. We formulate implicit negative candidate discovery to identify unobserved interactions supported by observed customer behavior. We encode these patterns as symbolic rules, score them based on support, informativeness, and product relevance, and rank the retained rules by evidence. An LLM then interprets the retained rules using business objectives and domain knowledge; the interpretations are combined with the statistical evidence in the final report. We evaluate our method in an industrial B2B setting and across five public recommendation datasets. Candidate-quality evaluations in the industrial setting and three public datasets show higher precision than the evaluated baselines, while symbolic selection improves downstream test PR-AUC by 12.5% over random selection with four negatives per positive example in the industrial task. Our results show that negative candidate validity can be evaluated separately from downstream recommendation performance. This distinction enables evidence-based, business-aligned, and explainable negative selection, improving both interpretability and model training in sparse, skewed, real-world recommendation settings.