Papers

We study how to detect and mitigate visual attention sinks in vision-language models (VLMs). Visual attention sinks are image tokens that absorb a large share of attention over the image regardless of the query or task, and they mislead methods that treat attention maps as a record of where the model “looked.” Prior work has detected sinks through massive activation dimensions and proposed redistributing their attention, but none have addressed what makes a token become a sink. Comparing six candidate detectors on ten VLMs from four families, we find that only the centered cosine similarity between an image token's key and the key of the language model's text-sink detects visual sinks. Thus, a visual sink is an image token whose key aligns with the text-sink's key. Moving a sink's key component along this direction to the median of the image keys sends its attention back to the text-sink (Are Borrowed) while leaving model answers unchanged~(Not Seen). Building on this, we propose a training-free, entropy-triggered re-look that first cleans the attention map, then crops the most-attended region of the cleaned map and inserts it into the reasoning trace. Cleaning adds negligible computation to the forward pass, and the re-look improves accuracy with a mean gain of $+4.3$ points on fine-detail VQA benchmarks across model families.

Grounded spatial reasoning about indoor scenes rests on a geometric reconstruction that the multimodal system does not itself produce. Feed-forward multi-view reconstruction models operate under a fixed view budget, making frame selection a critical upstream decision. We show that under this setting the conventional preference for sharp frames can be actively harmful. Across nine handheld room scans, selecting the sharpest 16 of 200 candidate frames is the worst strategy we test, reaching 48% room dimension error where uniform temporal sampling reaches 3%. We trace this failure to a coupling between image quality and camera motion: sharp frames tend to occur during slow motion and therefore concentrate the selection in a narrow angular region, leaving $192^\circ$ of the room unobserved in the case we illustrate and over $300^\circ$ in the worst cells. A $2\times2$ factorial experiment that independently manipulates viewpoint spread and per-frame sharpness shows that spread has roughly three times the main effect of quality, and that blurry but well spread selections outperform sharp but clustered selections in 8 of 9 captures. The advantage also diminishes as the frame budget increases and angular coverage saturates, consistent with this mechanism. Finally, we ask whether the failure is detectable at inference: the reconstruction model's multi-view consistency score increases with reconstruction error, while a vision-language model given the same frames reports confidence unrelated to its own spatial estimation error. Under small view budgets, viewpoint coverage is a more important constraint than per-frame image quality, and neither uncertainty signal available at inference reveals this upstream failure.

Sparse autoencoder (SAE) features are typically discovered from how their activations co-vary with behaviour, yet they are often used through interventions such as steering or ablation. We take intervention attribution, the first-order effect of a specified intervention on the model output, as our estimand. For a broad class of response functions, we show that observational feature-discovery evidence generally does not identify this attribution, and derive an exact gradient-based decomposition of the gap between linear-probe predictions and intervention attribution. Across four model-SAE pairs, median probe-to-attribution magnitude ratios range from tens to hundreds, while attributions closely track small-step steering. For single-latent ablations with nonzero effects, gradient predictions match the realized sign in 91-97% of cases, compared with 17-35% for probe-implied changes. Attribution-selected sets also produce larger absolute mean ablation effects on held-out prompts than sets selected by correlation, rank tests or probe coefficients. Together, these results establish intervention attribution as a principled, gradient-estimable target for selecting SAE features for intervention.

Information-theoretic generalization bounds usually control dependence on the task sample, whereas selective privacy may constrain only a different, private coordinate. We study i.i.d. pairs $(X,Y)$, where a learner observes $(X^n,Y^n)$, leaks at most $\epsilon_n$ nats about $X^n$, and incurs loss on $Y$. A chain-rule transfer converts this mismatched constraint into the closed-form universal certificate $I(W;Y^n)\leq\min\{nH(Y),nH(Y\mid X)+\epsilon_n\}$. Combining the certificate with mutual-information generalization yields a bound governed by residual task uncertainty and normalized privacy leakage. We derive complementary task-compression and individual-sample refinements, identify a sharp obstruction showing why selective privacy alone need not give a vanishing certificate, and invert the bound to obtain the exact leakage allowance for a prescribed generalization tolerance. We prove exact information-level achievability for deterministic-task and independent-residual source families and show that the resulting $\sqrt{\epsilon_n/n}$ generalization rate is tight up to a universal constant. We also give a fixed-horizon adaptive application in which conditional leakage costs add.

Movable agents, such as autonomous vehicles and robots, should be able to autonomously adapt their pre-trained object recognition models to new environments without human annotations. While existing test-time adaptation (TTA) methods address label-free model adaptation, they typically treat test data as independent snapshots, without constructively exploiting a rich, unique source of supervision: the significant variance in prediction quality as an agent observes the same object from different distances and viewpoints. To capitalize on this, we propose MoCaFe, a model-agnostic and hyperparameter-insensitive paradigm that leverages motion-induced prediction variance for unsupervised adaptation. Specifically, we first develop mobility-calibrated filtering of pseudo-labels, which constructs reliable pseudo-labels by fusing cross-view predictions using inverse-variance weights. The filtered pseudo-labels then guide the selection of “hard samples” for effective learning. To further suppress pseudo-label corruption, we introduce p-MoCaFe, an optional extension that uses a public dataset to assign weighting factors to data samples, yielding an estimator of the clean loss with bounded bias. Extensive experiments on autonomous driving (nuScenes, KITTI) and embodied-agent datasets show that MoCaFe significantly outperforms state-of-the-art TTA schemes, achieving up to approximately 20% accuracy gain.

A fundamental challenge in cognitive science and neural computation is the binding problem. Compositional generalization can be achieved by factorizing representations into roles, the structural slots of a representation (e.g., serial position or syntactic function), and fillers, the contents occupying them (e.g., a phoneme or a word), which can be recombined freely to represent novel combinations. How neural systems implement role–filler binding, however, remains unknown. One of its simplest cases is phoneme sequencing, where phoneme identity (filler) must be bound to position (role) so that /gum/ and /mug/ are represented differently. A proposal from NLP, the `onion hypothesis' (Csordas et al., 2024), holds that the direction of the neural population state encodes token identity, while its magnitude encodes position. It was formulated for random token sequences; we extend it to natural language and test it in neural models and the human brain. We first evaluated it in an encoder-decoder trained to repeat English words, using representational geometry analyses and causal interventions. Magnitude alone proves incomplete: position is encoded by magnitude together with a systematic rotation of the state direction, a scheme we formalize as the `spiral hypothesis'. We then analyzed intracranial recordings from the auditory cortex of 20 participants listening to short sentences. Across superior temporal gyrus electrodes, the change in population response between consecutive phonemes decreases in magnitude as position advances, and its direction is aligned within consonants and within vowels but near orthogonal across the two classes, supporting the magnitude coding shared by both accounts. These results provide a geometric account of how neural systems bind phoneme identity to position, and show how neural-model hypotheses can yield testable predictions about human neural computation.

Reinforcement learning (RL) for large language model (LLM) agents remains challenging in long-horizon tasks with sparse outcome rewards, where assigning credit to individual actions is difficult. Methods such as GRPO assign the same trajectory-level advantage to all actions, while critic-based methods require learning an additional value function that can be costly and unstable. We introduce LLM Value Function RL (LVRL), which uses the pretrained knowledge and in-context learning ability of an LLM to estimate policy-aware state values without finetuning a separate critic. LVRL conditions the value model on group context and calibrates its predictions using known values at the initial and terminal states, enabling step-level credit assignment from estimated value differences. We further extend this idea to learning from an expert policy through on-policy distillation (OPD), introducing LLM Value Function OPD (LVOPD), which focuses teacher supervision on actions associated with decreases in estimated value. We evaluate our approach across diverse agentic benchmarks: AppWorld, BabyAI, and SWE-bench Verified. Across policy scales (4B to 32B), LVRL consistently and substantially outperforms both GRPO and PPO. On AppWorld and SWE-bench Verified, LVRL yields absolute improvements of up to 13.6 and 8.6 percentage points over GRPO. Notably, on tasks requiring extended exploration like BabyAI, LVRL bridges the gap from total failure to 75.0% success. Furthermore, LVOPD yields consistent improvements over standard distillation baselines on AppWorld. Finally, Monte-Carlo analysis confirms that our frozen critic accurately identifies semantic pitfalls and recoveries, proving that LVRL extracts robust, policy-aware step-level credit from sparse terminal outcomes.

Scientific claims shared through social media platforms frequently reference research findings without explicitly linking to the corresponding scientific publications. Automatically retrieving the correct scientific source for such claims is a challenging multilingual retrieval problem involving semantic matching, paraphrasing, noisy social media text, and cross-domain language variation. In this work, we present a multilingual two-stage retrieval framework for CLEF 2026 CheckThat! Task 1 on Source Retrieval for Scientific Web Claims. Our system combines fine-tuned dense retrieval models with cross-encoder reranking for improving multilingual scientific document retrieval. We evaluate multiple transformer-based dense retrieval architectures including multilingual-E5, BGE, and Contriever under different metadata configurations. We further investigate multilingual joint training, retrieval depth analysis, metadata enrichment, and language balancing strategies for improving retrieval effectiveness. Experimental results demonstrate that metadata enrichment consistently improves retrieval effectiveness, while fine-tuned cross-encoder reranking provides additional gains in retrieval precision. Our final multilingual system achieves MRR@5 scores of 0.6222 for English, 0.5112 for German, and 0.6254 for French, with an overall average MRR@5 score of 0.5862.

When told that robins have sesamoid bones, people are far more prone to extend the property to sparrows than to whales. This behavior, called category-based property induction, is well documented in humans, but whether and how it develops in language models trained only to predict the next token is unclear. To shed light on this, we construct a hierarchical taxonomy of synthetic entities grouped by the properties they share. We then inject facts from this latent taxonomy into the continued pre-training of Pythia and OLMo 2 models from 70M to 7B parameters, and track the development of inductive behaviors. We find that taxonomically aligned induction emerges from next-token prediction alone: a property learned for one entity transfers more strongly within its taxonomic category than across categories. Internally, entities that share a deeper category become more similar in activation space, and clustering recovers the full taxonomic tree from entity representations. Contrary to the usual expectation that scale helps, we find that induction weakens with model size when the data budget is fixed, suggesting that larger models memorize individual facts rather than compress them into shared structure. Induction also strengthens with more pre-training before the taxonomy injection, transfers to novel in-context properties, and can be overridden by a latent taxonomy implied by the prompt. Together, these results show that a model's inductive generalization reflects the latent organization of its training data, with implications for data curation and for predicting how far a learned fact will spread.

Behavioral foundation models (BFMs) provide humanoids with reusable repertoires of whole-body behaviors; in real-world humanoid control, the same behavioral intent may require fundamentally different motions under different terrain constraints. Covering such variation with terrain-matched demonstrations requires enumerating a combinatorial space of behaviors and terrain geometries. We introduce P-BFM, a perceptive behavioral foundation model that factorizes humanoid control into terrain-invariant intent and terrain-conditioned realization. A forward-backward representation maps motion references, target states, and reward functions into a shared intent space without terrain information, while a perceptive policy learns how to realize each intent under the surrounding terrain geometry. We further identify a complementary limitation of purely latent control: behavioral latents can preserve local motion patterns while losing precise global progress. We therefore introduce Explicit Root Instructions (ERI), which separates global transport (e.g., heading and planar velocity) from behavioral intent, while leaving local whole-body adaptation to the perceptive policy. This decomposition leads to a characteristic capability absent from conventional BFMs: one fixed intent induces different, terrain-appropriate behaviors, including changes in step height, stride, posture, foothold placement, and contacts, without terrain-specific demonstrations. Across challenging terrains, P-BFM improves motion fidelity, root-control precision, and perturbation recovery over strong baselines. A single frozen policy further supports motion tracking, goal reaching, reward inference, teleoperation, and terrain-aware recovery, and transfers zero-shot to a Unitree G1 using onboard depth sensing.

Resolving correct citations can be a time-consuming task for research writers. We address local citation recommendation (LCR), where the goal is to retrieve appropriate references for a given sentence. Web tools generally depend on the existing citation graph, which may bias results toward citation count or more recent publications. LCR research using full paper texts has been limited by the lack of publicly available data. We present a new dataset, the Local Citation Recommendation Corpus (LCRC), containing 14,735 scientific statements derived from astrophysics review papers, each with one or more citation targets whose full text is accessible via API. Using this dataset we systematically compare retrieval methods and develop a citation recommendation system that avoids citation graph bias while achieving 8.0% improvement in Recall@25 and 6.9% improvement in NDCG@25 over dense retrieval baselines without model fine-tuning. To support reproducibility, we make our code and dataset publicly available.

When several predictions use the same internal representation, information about some targets can remain easier to recover than information about others. We study how this specialization forms, persists, and changes in minimal artificial learners. With reconstruction errors weighted equally during learning, changing which target is observed precisely selects the favored target and redirects an already specialized learner. Removing current support can reduce a formed bias, whereas state-following update feedback can maintain it. We then compare learners formed through different observation and update conditions but matched on specified reconstruction measurements. Both change under the same brief opposing intervention, yet their responses differ. Resetting Adam largely removes the ten-update difference; exchanging its second moment reverses which route is more susceptible. New interventions show that equalizing the mean second-moment scaling reduces but does not remove the difference: its coordinate-wise pattern also matters. With fresh Adam or calibrated SGD, smaller route differences emerge later, whereas 500 common opposing updates bring the selected allocation endpoints close. A further control produces response differences with identical acquisition and equal mean loss weights but jointly different weight variability and realized sequence. Acquisition-selected formation also extends to a symmetric three-target setting. These interventions distinguish specialization's formation, present support, and time-dependent latent susceptibility.

Hard-margin SVM training is the distance problem between two convex hulls, which the Triangle Algorithm (TA) solves with a certified duality gap. We propose guarded block transfers: $k$ Mitchell–Demyanov–Malozemov (MDM) transfers per scan are aggregated into one exact line search, guarded by a comparison with the single best transfer at a cost independent of the number of points $n$, and replaced by an away step when the leading transfer is capacity-clipped. On the Minkowski-difference polytope, where the objective is $1$-strongly convex, we prove global linear convergence with explicit constants and no combinatorial swap-step bounds; the per-step contraction is $k$-independent and only the drop-count bound grows with $k$, so aggregation is provably safe while the saving in scans is measured, not proved. Since the TA certificate is the Frank–Wolfe gap divided by the current distance, gap-safe screening applies, and we bound the iteration after which it has removed every zero-weight point off the optimal faces by $O\!\left(\tfrac{k}{\rho}\log\tfrac{1}{\tau}\right)$ up to logarithmic factors in the geometry, $\rho$ the rate constant and $\tau$ the optimal-face margin. On five LIBSVM benchmarks, with fairer baselines than our submission, the solver matches the baselines to four or five digits wherever both converge, ties LIBSVM on sparse-support hard margins and beats our blended pairwise conditional gradients (BPCG). But our full-cache sequential minimal optimisation (SMO) code ties or beats it on kernel $L_2$ margins, LIBLINEAR's dual solver is up to $4.5\times$ faster on the separable $L_2$ margin, and our solver is one to three orders of magnitude slower than LIBLINEAR's Newton solver on dense-support $L_2$ margins and than LIBSVM on a nearly touching kernel hull. Its edge is the certificate, not speed.

Sorting is one of the most basic algorithmic problems. Adaptive sorting algorithms exploit features that make an input easier to sort, often captured by entropy measures. Many such measures have been studied, raising two natural questions: how do they relate to one another, and are there algorithms that are adaptive to them? We complete a hierarchy of relations between previously studied entropy measures and show that the classical measure Reg is a source of this hierarchy. We prove that an existing data structure yields a Reg-adaptive sorting algorithm, thereby achieving the adaptive bounds of every measure in the hierarchy. Following McIlroy's observation that Reg has a mathematically inconvenient definition, we study a slightly weaker but simpler measure, reg. For inputs without duplicates, we show that O(n + log reg) = O(n + log Reg), so this simplification preserves the adaptive bound. Finally, we do algorithms design and introduce SproutSort, a reg-adaptive algorithm in the offline binary search tree (BST) model. SproutSort achieves an access cost of O(n + log reg), making it the most adaptive algorithm in this model to date. We also introduce SkeletonSort, a simple randomised pointer-machine algorithm that sorts distinct keys in expected O(n + log reg) time, offering a considerably simpler implementation of a sorting algorithm this high in the adaptive hierarchy.

Sequential statistical procedures are increasingly embedded in scientific systems in which observations arrive incrementally from heterogeneous sources. In such systems the statistical update is only one part of the computation: observations must also be acquired, aligned across inputs, ordered, delivered, persisted, and replayed, and results must be passed to systems that display them or act on them. Sequential methods are nevertheless missing from the general software toolkit. They run inside proprietary experimentation platforms and commercial statistical packages, or exist as offline functions over complete data, and the batch fit interface that dominates machine-learning libraries has no place for a procedure that listens, keeps state, and decides when to stop. We propose an architecture that separates statistical inference from data topology, execution runtime, run lifecycle, persistence, and external action. Its central abstraction is the Sequential Procedure: a state-transition function that maps an inferential state, a valid statistical input, and a statistical random state to a new state and a scientific output. The procedure does not define how observations are acquired, paired, scheduled, stored, or acted upon. We specify the architecture semantically rather than through an object model, derive seven conformance invariants that any implementation can be tested against (among them synchronous/asynchronous, checkpoint, replay, and batch equivalence), and describe a dependency-free reference implementation in which all seven hold bitwise. Lorden's 2-SPRT for the Kiefer-Weiss problem serves as the reference procedure; the same contract accommodates sequential estimators, Bayesian filters, change-point detectors, and stateful predictive models.

Modern cloud platforms generate high-dimensional, noisy telemetry from thousands of microservices. Unlike classical anomaly detection benchmarks, cloud anomalies are short-lived, irregular, and largely non-periodic, making timely detection challenging. We present an empirical benchmark of deep learning models for anomaly detection using a real-world IBM Cloud Console dataset. We evaluate GRU, Transformer, TimesNet, and Autoformer architectures using the Numenta Anomaly Benchmark (NAB), a timing-aware metric that accounts for early detection and false alarms. TimesNet achieves the highest NAB scores due to strong early-detection sensitivity. Transformer models also detect anomalies early but incur higher false-positive rates. An improved GRU autoencoder, incorporating robust scoring and persistence rules, achieves competitive performance with a more stable false-alarm profile. Overall, the modest performance differences suggest that real cloud anomalies are dominated by short-term deviations rather than persistent periodic structure, emphasizing the importance of timing-aware evaluation and robust post-processing.

Real-world datasets often exhibit severe class imbalance, degrading the performance of deep learning models, particularly on rare classes. While imbalance is well studied in binary and multiclass settings, it is substantially more challenging in multilabel problems due to complex label co-occurrence. In clinical applications such as ECG classification, this challenge is coupled with a second critical challenge: the need to prioritize high-risk conditions, regardless of their prevalence, to mitigate the high cost of false negatives. We introduce the Multilabel Class-Prioritized Probabilistic Sampler (ML-CLAPPS), a learnable sampler that optimizes a distribution over training instances to control the model's exposure to different classes. This approach addresses multilabel imbalance while enabling prioritization of significant conditions via a user-defined importance vector (IV). Empirical evaluation on a real-world multilabel ECG image dataset demonstrates that ML-CLAPPS achieves state-of-the-art performance. We show that it aligns class exposure with the specified IV, and that increasing a class's importance leads to corresponding improvements in its classification performance, enabling explicit prioritization of clinically critical conditions.

Recent work shows that ordinary Reed–Solomon codes over prime fields admit polynomial-size lists up to list-decoding capacity on every evaluation set. We show a contrasting phenomenon over binary extension fields. For every fixed rate $0<R<1$, there exists $\varepsilon_R>0$ such that, for all sufficiently large $h$, the full-field Reed–Solomon code over $\mathbb{F}_{2^h}$ has a received word with $n^{\Omega_R(\log n)}$ codewords within relative distance $1-R-\varepsilon_R$. To our knowledge, these are the first superpolynomial ordinary Reed–Solomon list-size lower bounds at fixed positive rate and a fixed positive agreement gap above the rate. Thus capacity-approaching polynomial list size for Reed–Solomon codes is not uniform over the underlying field, and the obstruction occurs at every fixed rate. As a puncturing consequence, the same holds for primitive Reed–Solomon codes; the obstruction is also robust under controlled dense puncturing. The proof begins with an exact correspondence that turns certain binary Bose–Chaudhuri–Hocquenghem weight-enumerator coefficients into sizes of single Reed–Solomon lists. A classical binary weight family gives the initial quarter-rate construction, while a quadratic-form enumeration supplies its superpolynomial multiplicity. A common-kernel quotient/descent argument and common-root padding then extend the construction to all rates. We also determine the optimized agreement envelope achieved by the construction.

General-purpose multimodal foundation models are increasingly capable of directly controlling robots, yet improving their manipulation capabilities through autonomous interaction remains challenging. A key challenge to such self-improvement is learning from failures in complex manipulation tasks involving cross-stage dependencies, where an action may succeed locally but cause a later stage to fail. To address this challenge, we introduce ReTrace, a self-improvement framework for robotic manipulation through autonomous physical interaction and retrospective reasoning, without provided demonstrations or model-parameter updates. ReTrace retrospectively traces downstream failures to earlier decisions, proposes targeted corrections, and evaluates them through subsequent interactions. Across attempts, ReTrace progressively consolidates validated corrections into a reusable manipulation skill that guide subsequent execution. We evaluate ReTrace on four real-world manipulation tasks using a Franka Research 3 robot. On the two tasks with strong cross-stage dependencies, GPT-6 Astra-low improves from 0% to 80% success on plug insertion and from 0% to 100% on ball pouring, while GPT-6 Sol-xhigh improves from 0% to 10% and 0% to 90%, respectively. The learned skills also transfer across model and reasoning configurations and reduce inference-time decision cost in most evaluated settings.

Bitemporal remote sensing change detection requires representations that preserve spatial detail while capturing contextual information relevant to changes between acquisitions. Visual test-time training offers an alternative to explicit token-to-token attention by adapting a compact inner model to each image. This letter describes ChangeT$^3$, a vanilla Siamese configuration combining the existing hierarchical ViT$^3$-Tiny encoder with the multiscale ChangeFormer decoder. The two image streams share learned encoder parameters, while their fast weights are computed independently from their respective spatial features. Each encoder block retains the inherited global simplified SwiGLU and local adaptive depthwise-convolution branches. Temporal comparison is performed by the decoder through learned fusion of paired features at four resolutions, followed by coarse-to-fine aggregation and binary prediction. This separation provides a direct setting for examining image-conditioned inner updates without introducing an additional temporal interaction module. Trained from scratch on LEVIR-CD+ with the same decoder, recipe, and budget as a ChangeFormer control, ChangeT$^3$ reaches an F1 of 80.61$\pm$0.90 against 78.84$\pm$0.21 for ChangeFormer (three seeds each); as in the standard protocol for this dataset, checkpoints are selected on the test split. Trained with both inner updates disabled, the same encoder reaches 78.63$\pm$0.20. Neither model transfers zero-shot to SYSU-CD, DSIFN-CD, or WHU-CD, and ChangeT$^3$ is the slower of the two at small inputs despite similar parameter and operation counts.

The introduction of long chain-of-thought (CoT) reasoning and external tools has substantially improved the performance of multimodal large language models (MLLMs) on general and reasoning tasks. However, existing fast-slow reasoning paradigms mainly regulate textual reasoning length, while visual computation introduced by external tools or visual tokens is often treated as an auxiliary component. This separation prevents MLLMs from jointly optimizing visual perception and textual reasoning. To address this issue, we propose L2T (adaptive Look, flexible Thought), a unified reasoning framework that jointly optimizes the form and depth of computation. L2T first discretizes tool-produced visual states into visual tokens, bringing visual perception and textual reasoning into a shared autoregressive space. Based on this unified representation, L2T introduces a modality-aware routing reward to select appropriate reasoning modes and a difficulty-aware fast-slow strategy to adapt reasoning depth according to task complexity. Experiments on general perception and reasoning-intensive tasks demonstrate the effectiveness of L2T. Compared with explicit tool-calling approaches, L2T maintains competitive performance while achieving 1.7–3.6 times higher throughput.

Data flow underlies deep learning (DL) framework execution, including automatic differentiation and operator-fusion optimization. Bugs involving data flow can cause incorrect results or unexpected behavior. Existing testing methods represent models primarily by operator topology, omitting runtime tensor values and properties and operator input requirements. They therefore cannot distinguish data flows in models with similar topologies, even when these differences affect framework behavior. We introduce the Tensor Flow Graph (TFG) to capture runtime tensor states and operator input requirements. Its nodes represent operators and their input requirements, while its edges record tensor values and properties. Building on the TFG, we propose FlowMuT, a mutation-based testing method grounded in data flow. Mutations are represented as constrained subgraph replacements, with mutation operators derived from prior work and an analysis of mutable data flow objects. In each round, FlowMuT builds a TFG for the seed model, pairs mutation operators with TFG sites, and filters illegal candidates using data flow constraints. It uses LinUCB to select a candidate based on coverage and impact, executes the mutant, updates selection feedback, and uses each accepted mutant as the next seed. We evaluate FlowMuT on PyTorch and MindSpore using representative models from diverse application scenarios. FlowMuT submitted 55 bug reports for cases missed by all baseline methods; developers confirmed 41 reports and fixed the bugs described in 7 reports. FlowMuT achieves 43.4% higher data flow diversity than the strongest baseline and a legal mutant rate of 96.7%, compared with 86.4% without constraints. These results show that testing from a data flow perspective is practical and effective.

Generating coherent videos grounded in a unified underlying world remains a major challenge, particularly under large viewpoint changes, where existing methods often struggle to maintain stable representations of individual objects. To this end, we introduce Totem, an object-centric 3D memory framework for world-consistent video generation. Totem first represents each selected object as a persistent 3D memory aligned to a shared world frame, assigning it a dedicated state. Based on these memories, we design a dual-stream 3D conditioning architecture that combines object memories with scene geometry for target-view synthesis. Moreover, these object memories constrain the model's attention with 3D structural information, thereby preserving object surfaces and relative spatial layouts. Extensive experiments on world-generation and object-centric benchmarks demonstrate that Totem generates videos depicting a more object-consistent 3D world under challenging camera trajectories, while supporting versatile applications.

Recovering visually faithful content from damaged video bitstreams is important for reliable multimedia communication and storage. Existing methods exploit local residuals and non-local RGB references of the input video sequence, but both sources can become unreliable after decoding, while processing additional references increases receiver computation and memory. This paper studies multimodal bitstream-corrupted video recovery (MBSCVR) using RGB and depth information, where pre-transmission depth-based geometry aids reconstruction when the received appearance is corrupted. We propose a novel MBSCVR framework that selects and effectively utilizes a single selected non-local RGB reference per recovery window to achieve higher performance. It integrates a reliability-aware multimodal selector that combines valid appearance with depth compatibility before reference encoding. Depth-Conditioned Reference Adaptation (DCRA) introduces selected RGB evidence into feature completion and content generation, while frame-aligned depth conditions completion queries. On the complete BSCV benchmark with oracle masks, MBSCVR uses one selected reference to achieve 32.30 dB on YouTube-VOS and 28.93 dB on DAVIS, surpassing full-reference B2SCVR by 0.27 and 0.33 dB and improving SSIM on both subsets. Matched receiver-forward measurements show 36.8% lower latency and 40.8% lower peak memory. These results demonstrate how compact geometric guidance and adapted appearance processing improve the quality–resource trade-off of multimedia recovery.

Large language model (LLM) that retrieve internal documents within an organization must decide who may receive them. Existing approaches based on Contextual Integrity ask the model to make that decision, so access depends on probabilistic reasoning. We present Privacy Conductor, an architecture that removes the decision from the model: a deterministic gate enforces a policy written by an authority, and the model receives only what the requester may already see. On a synthetic clinical corpus of 2,520 requester–document flows, the gate reproduces its policy exactly, so the remaining risk lies in the document labels it reads. As label source is of primary importance, we compare LLM and human generated labels. Across seven configurations of six models from two model families, LLM-written labels degrade the policy in every case. Models assign a document's sensitivity tier correctly on 78–86% of documents, but reproduce its access-control list on as few as 23%. Relabeling the same documents five times with an identical request changes 13 to 124 access decisions per configuration, a result that does not depend on our own labels. On 214 flows from HIPAA court cases, the most accurate configuration (87.8%) still permits 11 of the 20 disclosures the law prohibits. LLMs can describe what a document contains, but not decide who should receive it. Hence, we recommend privacy policy should be authored by accountable personnel and enforced outside the model.

We study the problem of fair allocation when agents have dichotomous submodular preferences over groups of, rather than individual, items. These preferences naturally extend existing models, and have a natural interpretation in settings such as course allocation. We show that this seemingly minor extension results in computational intractability barriers: computing a utilitarian or Nash-optimal allocation with item complements is NP-hard. However, when agents' complementary bundles have size at most $k$, there exist poly-time algorithms that compute $\frac1k$-$\texttt{maxUSW}$ allocations and approximate Nash-welfare-maximizing allocations. While PO+EF-1 allocations may not exist in the complementary preference setting, it is possible to compute $\frac1{k+1}$-approximately $\texttt{maxUSW}$ EF-X allocations in polynomial time; furthermore, these allocations offer each agent $i$ a utility of at least $\frac{\texttt{MMS}_i - k}{k+1 + \frac1n}$, where $\texttt{MMS}_i$ is agent $i$'s maximin share guarantee. While the standard dichotomous binary submodular setting admits simple sequential techniques for computing fair and efficient allocations, our results for the complements setting require more complex tools from combinatorial optimization.

Oral health is vital for overall well-being, yet many users brush different tooth surfaces for insufficient or excessive durations. Toothbrushing monitoring systems can help improve brushing habits by reminding users of under-brushed and over-brushed surfaces, but existing approaches often rely on external sensors, increasing cost and reducing practicality. We observe that the intrinsic motor signals of electric toothbrushes can reflect the brushing surfaces and durations. When brushing different tooth surfaces, the brush head experiences different constraints from the surrounding oral cavity, changing the motor's driving currents and position feedback. Based on this observation, we propose FreeBrush, a toothbrushing monitoring system that uses the intrinsic motor signals of electric toothbrushes to recognize brushed tooth surfaces and estimate brushing durations without any external sensors. We prototype the system on a commercial off-the-shelf electric toothbrush and conduct a comprehensive evaluation with 18 participants. FreeBrush achieves 92.8% accuracy for 16-surface recognition and 92.0% accuracy for brushing duration estimation. The system further shows robustness across different toothpastes, brush heads, vibration modes, usage scenarios, and long-term use, and is capable of running on edge devices. By recognizing brushed tooth surfaces and estimating the brushing duration, FreeBrush can help users identify under-brushed and over-brushed surfaces and build balanced brushing habits.

Planning and revision are intended to bring generated code closer to the task specification. Yet a mistaken interpretation can persist across the plan, implementation, and tests - a pattern we call interpretive lock-in. We investigate this risk on 30 DataFrame tasks derived from TPC Benchmark DS, comparing Frozen One-Pass, Test-Guided Revision, and PlanAudit-Guided Revision. Test-pass counts rose from 22/30 to 29/30 and 30/30, yet no generated solution matched its stored reference output. Evaluator controls returned the expected outcome in 57/57 cases. Same-engine SQL re-execution reproduced 28/30 references; all 77 test-passing runs on these tasks still failed reference agreement. PlanAudit-Guided Revision improved code scores on 3/30 tasks and test coverage on 11/30, but the median paired change was zero for both measures. Failure analysis revealed tests that omitted requirements, adopted candidate-defined schemas, used nondiscriminating fixtures, or repeated the implementation's incorrect logic. Fresh generation on ten targeted tasks produced 0/30 reference matches and 25/25 false passes among test-passing runs. These findings are diagnostic rather than causal: the pipelines differ in feedback, grounding, call budgets, and editing permissions, and the study uses one model and one primary generation per eligible task. Because pre-adjudication human labels were not retained, inter-rater agreement cannot be reconstructed and the mechanism analyses remain exploratory. Overall, increasingly successful self-generated tests did not establish semantic correctness, even after excluding unreproduced references.

Automated moderation systems are deployed at scale, yet they remain opaque to the users and moderators affected by their decisions, and their coverage is almost entirely confined to English. For Urdu written in the Nastaliq script, no such system exists, and the annotated resources it would be built on are drawn largely from Twitter/X, where character limits constrain how users write. Interpretability has received little attention, leaving it unclear what such models actually learn. This study addresses these gaps in three steps. First, we curate and publicly release a corpus of 7604 Nastaliq Urdu Facebook comments, extending an earlier corpus by correcting text-level inconsistencies and re-examining the original labels. Second, we fine-tune four Urdu-capable transformers under a common grid-search protocol and compare them against twelve feature-classifier baselines on identical data partitions. The best configuration, Multilingual Representations for Indian Languages (MuRIL), reaches 87.05% accuracy and 87.06% macro F1, improving on the strongest baseline by 5.52 percentage points, a margin that exceeds the 95% confidence interval on test accuracy; the two leading transformers differ by less than the test set can resolve, and performance is stable across five optimizers. Third, we compute word-level SHAP attributions adapted to right-to-left rendering. Attribution mass concentrates on high-frequency slurs but extends to syntactically neutral function words and to political vocabulary characteristic of the source pages, indicating that aggregate performance may rest partly on lexical and topical regularities rather than a generalizable representation of offensiveness.

Flow matching is a common generative recipe for robot foundation models. Vision-language-action models (VLAs) generate action chunks, while world-action models (WAMs) additionally generate future visual states. Both start generation from a fixed Gaussian prior, despite temporal continuity between past and future states and actions. This requires costly training and iterative sampling, especially for WAMs with an additional visual branch. We introduce $\Delta$-flow, which constructs the source from recent history with mild noise rather than from a fixed Gaussian prior. This moves the source closer to the target, reducing transport to a small residual $\Delta$ that can be learned efficiently and enables single-step inference. We integrate $\Delta$-flow into the VLA policy $\pi_{0.5}$ and the WAM policies ImageWAM and Fast-WAM. On LIBERO, ImageWAM reaches 90% success in one epoch rather than six, and Fast-WAM reaches 95% in two epochs rather than eight. With LoRA fine-tuning, $\pi_{0.5}$ exceeds its baseline's five-epoch performance after one epoch. Single-step inference gives $1.54\times$, $2.49\times$, and $4.45\times$ speedups on $\pi_{0.5}$, ImageWAM, and Fast-WAM, respectively. It preserves action smoothness, achieving a $4.5\times$ reduction in action jerk compared with one-step Gaussian sampling. $\Delta$-flow reaches state-of-the-art generalization performance on LIBERO-Plus. Action-history sources can be applied to pretrained WAM policies without retraining. Replacing the Gaussian action source at inference largely preserves task success with a single sampling step and can slightly improve it. We further study how action representation and visual information in action learning affect history-to-future prediction in $\Delta$-flow. These results show that $\Delta$-flow is simple yet effective, improving the efficiency and generalization of robot foundation models.