Papers

Learning contact-rich dexterous manipulation from human demonstrations provides a scalable source of interaction data, yet transferring such skills to robots remains challenging due to unreliable human–robot correspondence. Existing human-to-robot transfer methods typically rely on visual appearance or motion similarity, which may associate similar motions with different contact states and force patterns. Tactile dynamics provide interaction-aware cues to distinguish manipulation processes with similar motions but different contact states. We introduce tactile-guided optimal transport (TacOT), a framework for human-to-robot contact-rich manipulation. TacOT leverages action–tactile dynamic time warping to identify human–robot demonstration correspondences with consistent interaction dynamics and uses these correspondences to guide soft optimal transport alignment in a shared policy representation space. This enables human demonstrations to provide contact-rich supervision for robot policy learning without requiring predefined frame-level human–robot pairing. Across four real-world dexterous manipulation tasks, TacOT improves closed-loop success rates over action-guided OT by up to 17 points on in-distribution tasks and 20 points under targeted human-to-robot out-of-distribution transfer. Further analyses show that tactile-guided correspondence selects demonstration pairs with more consistent contact dynamics and produces latent representations that better reflect interaction-state evolution. These results demonstrate that tactile dynamics provide an effective semantic signal for establishing reliable human-to-robot correspondence in contact-rich dexterous manipulation.

Multi-modal object re-identification exploits complementary RGB, near-infrared (NIR), and thermal-infrared (TIR) observations to retrieve target objects. However, existing methods commonly employ visual encoders optimized for global image-text alignment and select tokens using learned importance scores. Such designs fail to preserve fine-grained identity cues or explicitly account for token redundancy, resulting in underrepresented local evidence and duplicated tokens that lead to noisy and costly cross-modal interaction. To address this gap, we propose TRIM-ReID, a compact framework that unifies dense feature extraction, intra-modal token reduction, and inter-modal aligned interaction. Specifically, semantically rich and spatially coherent patch features are extracted by Dense Identity Representation (DIR), which leverages DINOv3 to preserve fine-grained identity information. We then introduce Token Diversity Mining (TDM) to identify complementary local evidence and construct compact modality-specific token sets by suppressing repetitive patches while preserving informative diversity. Retained tokens are subsequently fused by Modal Relational Interaction (MRI) to enable effective information exchange across modalities, while a triangular alignment loss explicitly regularizes their joint relationships to maintain cross-modal semantic consistency under independent token selection. Extensive experiments on RGBNT201, RGBNT100, and MSVR310 demonstrate that TRIM-ReID achieves state-of-the-art performance.

Quantum networks will rely on intermediate nodes to distribute entanglement, but the same nodes may learn which users are being connected. We introduce a minimal two-call routing primitive in which a user, Alice, establishes an entangled link with one of several clients while testing whether the provider, Bob, has disturbed a coherently encoded request. The construction adapts the hidden-order mechanism of quantum private queries: one call contains a direct request, the other a coherent superposition of the desired and null addresses, and Alice keeps their order secret. A private matching qubit held by the selected client restores the coherent test state after a favorable Bell-measurement outcome, so that Alice can test the returned query without any further operation by Bob. The claim is deliberately limited. The protocol shows that a routing request can carry a physical cheating test, and it detects an explicit destination-reading attack with nonzero probability. It does not yet provide a quantitative information-disturbance tradeoff, composable security, or certification of an unmeasured output against arbitrary attacks. We also characterize the exact perfect-pass limit, in which Bob's residual state, averaged over the private matching states, is independent of the destination.

We give checkable positive-definiteness criteria for the limiting neural tangent kernel (NTK) and high-probability finite-width gradient-descent guarantees for scalar- or vector-valued physics-informed neural networks (PINNs) with linear constraints. The constraints may be strong-form differential rows of any fixed finite order, including coupled systems, with any linear initial or boundary conditions; weak-form residual and boundary functionals; or finite-measure nonlocal observations such as integral, nonlocal-diffusion, and Dirac-data rows, all in any dimension. For each class, positive definiteness of the limiting NTK is equivalent to a rank condition on a finite coefficient, functional, or moment matrix of the fixed design: a certificate computed before training that detects structural zero modes. The model hypotheses are those of an ordinary two-layer PINN, a smooth nonpolynomial activation with bounded symmetric initialization, and are met by standard choices such as $\tanh$ with uniform initialization. Given a certificate, explicit width and step-size conditions ensure, with high probability, that the empirical NTK retains at least half the limiting gap and that full-batch gradient descent decreases the training loss geometrically at every iteration. Controlled experiments check the certificates and the finite-width mechanism.

Comprehensive 3D scene understanding for autonomous driving requires modeling geometry, semantics, and motion. However, camera-based occupancy and scene flow prediction are sensitive to unreliable spatial and temporal aggregation, caused by semantically incompatible image features, misaligned historical observations, and incomplete voxel structures. To address this issue, we propose SelectOccFlow, a selective spatiotemporal aggregation framework that progressively refines contextual evidence across image, temporal, and voxel domains. To obtain semantically compatible image evidence, we design Semantic-Guided Sampling (SGS) to regulate feature sampling with semantic priors. Since reliable image evidence alone cannot resolve temporal inconsistency, we then present State-Conditioned Temporal Aggregation (SCTA) to selectively retrieve historical evidence according to voxel states. To further enhance the structural completeness of voxel representations, we introduce Extent-Aware Spatial Aggregation (ESA), which exploits directional structural support to refine foreground geometry. Experiments on OpenOcc demonstrate that SelectOccFlow achieves a state-of-the-art OccScore of 44.9, improving the previous best by +4.2%. It also maintains competitive occupancy performance on Occ3D-nus and improves the mean OccScore under nuScenes-C corruptions by +11.1%, demonstrating improved robustness to visual corruptions. The source code will be made publicly available at https://github.com/muchen1021/SelectOccFlow.

In this paper we devise and analyze Lie Trotter splitting schemes for non-autonomous evolution equations, following two approaches: 1) A fully non-autonomous scheme, motivated by the potential use of explicit solutions of some subproblems. 2) A splitting scheme relying on the concept of Magnus integrators, where we freeze the operators at the beginning of each time interval as this suffices for first order convergence of the Lie-Trotter integrator. We derive first order convergence in a general setting and apply the analysis to the Bloch-Torrey equation of magnetic resonance imaging, where a decomposition into three subproblems is well-motivated by the physical role of the underlying operators. Numerical tests, complementing the time stepping scheme by a finite element space discretization, illustrate the theoretical findings.

Doc2Query, a popular document expansion technique, leverages sequence-to-sequence models to generate relevant queries, effectively addressing the "vocabulary mismatch" problem in information retrieval. However, these models often suffer from generating either hallucinations unrelated to the document or repetitive content already present in the document. Training sequence-to-sequence models to produce high-quality, novel, and relevant tokens remains a significant challenge. To address these issues, we introduce a novel approach, QGDPO, that employs direct preference optimization (DPO) to guide the generation process. We first fine-tune a base sequence-to-sequence model and subsequently utilize a relevance model to score its predictions. Based on these scores, we construct pairs of winning and losing predictions as relevance preferences for the DPO training. Furthermore, we enhance our pipeline by using the relevance model to filter out poor predictions, retaining only the most relevant generated content for indexing. QGDPO effectively eliminates 50% of irrelevant predictions comparing against Doc2Query baselines, while the relevance filter removes an additional 14.61%. This feature has been successfully deployed to production on Walmart.com for full traffic, with a substantial improvement in relevance and user engagement.

Feed-forward 3D Gaussian Splatting (3DGS) increasingly aggregates multi-view evidence with heavy learned 3D networks. We propose LoCoSplat (Local-Context Splatting), motivated by the observation that a Gaussian is a local primitive: once depth is predicted, what the 3D stage must add (scale, rotation, opacity) depends on the point cloud around each anchor, and a fixed local average of that neighbourhood is enough to supply it, no heavy network required. LoCoSplat realises exactly this average: it splats a 16-d linear projection of the point features into a fine and a coarse grid and reads both back at each anchor with a 0.14M-parameter pointwise MLP; with no learned 3D network and no dynamic sparse computation, its whole encoder runs as one fp16 CUDA graph. On RealEstate10K, LoCoSplat outperforms every prior feed-forward method on PSNR, SSIM, and LPIPS at 6, 12, and 24 views, with a margin that widens as views densify (+3.3 PSNR over VolSplat, the prior voxel-aligned state of the art, at 24 views) and grows further under zero-shot transfer to ACID and fine-tuning on ScanNet. It reconstructs a 6-view scene in 33 ms on one NVIDIA RTX PRO 6000 GPU, the fastest of seven feed-forward methods and $4.2\times$ faster than the previous state of the art, trains $2.7\times$ faster ($5.9\times$ at 24 views), and uses $6.7\times$ less inference memory.

We present a method for the data-driven inference of Hamiltonian and Lie-Poisson dynamical systems from vector-field samples, in which the Hamiltonian and the underlying Poisson tensor are estimated jointly rather than assuming the structure is known. The Hamiltonian is represented as a linear combination of random feature maps with fixed random internal weights, so its estimation is a linear regression problem for any given Poisson structure. We treat three cases of increasing generality: a known Poisson tensor, an unknown constant invertible (symplectic) structure matrix, and Lie-Poisson and affine Poisson structures, with structure matrix linear in the state. In the constant case, the Hamiltonian and structure matrix are recovered jointly from an eigenvalue problem that avoids the trivial zero solution due to the scaling symmetry of Hamiltonian systems. For Lie-Poisson and affine Poisson systems we alternate between an initial Hamiltonian estimate from a generalized eigenvalue problem for an approximate first integral, and a constrained nonlinear least-squares fit of the structure constants that enforces the Jacobi identity and, for affine structures, the cocycle condition, via continuation in a penalty parameter. Orthogonal matching pursuit selects a compact, well-conditioned subset of random features. Because the learned Hamiltonian is a sum of single-feature terms, the model admits an explicit Poisson/symplectic splitting integrator, which we use to validate the learned systems via long-time Poincare sections. We demonstrate the approach on the four- and five-dimensional Lorenz-86 model and a six-dimensional Kirchhoff rigid-body system, recovering Hamiltonians, structure constants and Casimir invariants to high accuracy, and showing the inferred Poincare sections converge to the correct invariant-tori topology as the number of random features increases.

We study the involution-preserving properties of the Active Flux method on unstructured triangular meshes. Two model problems are considered: the two-dimensional linear acoustic system, which satisfies a curl involution on the velocity field, and the two-dimensional Maxwell equations with an in-plane magnetic field, which satisfy a divergence involution. Both systems can be written in conservative form. The question we address is how a single discretization can satisfy the conservative form and the involution at the same time. Within a discrete Active Flux framework on simplices, we derive explicit formulas for a discrete curl operator that is exactly preserved by the scheme for acoustics, and for a discrete divergence operator preserved for the Maxwell system. The analysis applies to general simplex meshes and to any polynomial degree, including extensions of the method beyond third order. Numerical experiments with the third-order scheme confirm the theoretical results. Highlights: 1-Active Flux on unstructured triangular meshes preserves discrete involutions, 2-Explicit discrete curl operator preserved by the scheme for linear acoustics, 3- Explicit discrete divergence operator preserved for 2D Maxwell equations, 4- Results hold on general simplex meshes,% and for any polynomial degree 5- Third-order numerical experiments confirm the theoretical results. \end{itemize}

Identical utterance choices can arise from different communicative causes, and identical interpretations can leave different traces in what a listener learns. Rational Speech Act (RSA) models treat interpretation as inference over speaker meaning, but standard one-shot RSA does not intrinsically distinguish these causal update targets. We develop a coupled active-inference model of dialogue in which a listener's likelihood for an utterance is the policy distribution attributed to the speaker, placing the speaker's expected free energy within the listener's variational free energy. Each utterance is therefore both evidence about and an intervention on a partner. The model separates four updates that RSA approaches typically collapse: inference about the partner's current pragmatic state; learning of partner-specific parameters; prospective evaluation of clarification or repair; and a policy prior shaped by habit and a context-sensitive price of time. Under one-step, exact-inference restrictions, the model recovers the RSA speaker and listener, with RSA as the restricted single-turn limit of the coupled process. Outside these restrictions, the updates obey distinct rules and timescales, predicting which adaptations persist, remain partner-specific, or transfer. Worked examples show audience design reversing after clarification, self-confirming misunderstanding in which both interlocutors have low free energy while disagreeing about reference, and rational closing before uncertainty is resolved. Consistent with critiques of equating natural language with communication, the model treats communication as a downstream use of linguistic structure and recasts production and interpretation as coupled inference: speaking is both an intervention on a partner and an epistemic action that samples evidence for the speaker's model of that partner.

This paper proposes any-scale object detection using arbitrary-scale super-resolution for continuously rescaling object images, while general multi-scale object detection uses discretely rescaled appearance representations. However, a naive usage of super-resolution produces many false-positive detections if many super-resolution images are independently fed into an object detector. Our method suppresses these false positives by predicting scale proposal maps, each of which represents a set of pixels appropriate for each super-resolution scale.

Many wireless control tasks require repeated selection of a single action from a finite feasible set under stringent latency and reliability constraints. While large language models (LLMs) have recently emerged as general-purpose decision engines, their autoregressive generation mechanism is not naturally aligned with such bounded control problems. This paper investigates System-One models, which directly learn probability distributions over explicitly defined decision spaces, as a lightweight alternative for wireless decision-making. We formalize their decision structure and learning objective, identify their applicability across physical-layer control, radio resource management, mobility, network slicing, and network operations, and evaluate their practical behavior through representative wireless case studies. Using Jev as a System-One implementation, we benchmark decision quality and client-observed latency against generative LLMs and conventional baselines. In receive-antenna selection, Jev delivers up to an 8.5x reduction in median response latency relative to the evaluated LLMs, although this gain comes with a loss in decision quality compared with stronger task-specific alternatives. More notably, in intent-conditioned RAN slicing, Jev achieves utility comparable to the evaluated LLMs while providing more than a 3.5x reduction in median response latency. Complementary evidence from edge-service orchestration further shows that faster decisions do not necessarily translate into lower end-to-end service latency. These results expose a fundamental quality/latency tradeoff and position System-One models not as replacements for numerical optimization, but as a promising decision interface for latency-sensitive, bounded, and intent-driven wireless control.

Reinforcement learning with verifiable rewards (RLVR) has been shown to improve the reasoning capability of large language models (LLMs) across diverse reasoning tasks. However, group-based RLVR methods, such as GRPO, assign a uniform advantage to all tokens within rollouts of the same outcome. While existing works refine credit assignment of GRPO based on local signals such as token locations or entropy, they often fail to capture the global semantic novelty of a reasoning behavior relative to the current policy. In this work, we propose a hierarchical credit assignment approach for group-based RLVR methods, called HarA, which identifies and encourages semantically novel reasoning behaviors during RLVR. HarA represents each sampled rollout as a distribution over the hidden states and locations of tokens, and computes the Fused Gromov-Wasserstein (FGW) barycenters of all rollouts with the same outcome, capturing the internal reasoning patterns in the latent space under the current policy. The semantic novelty of a reasoning element can then be measured by its contribution to the FGW distance between the current rollout and the barycenter. While solving the FGW formulation is expensive, we introduce an anchor-guided linearization that turns it into a Wasserstein formulation solvable via the Sinkhorn algorithm efficiently. By reweighing token-level advantage of group-based RLVR methods based on the novelty signals, HarA highlights novel reasoning behaviors at flexible granularities to encourage fine-grained LLM exploration. Extensive experiments across three group-based RLVR methods show that our plug-and-play method effectively enhances the exploration of LLMs, outperforming existing methods across diverse reasoning benchmarks.

Reservoir computing (RC) is a machine learning technique for data-driven modeling of dynamics for forecasting and control, primarily studied in computer science and physics literature with promising applications in neural network learning, physical computing, and neuroscience. Why reservoirs learn and how to choose good reservoir architectures are considered important open questions. We show that the RC problem is mathematically an extension of a classic problem in systems and control theory, the Kazantzis-Kravaris-Luenberger (KKL) observer design problem. As a consequence, many of the questions considered open for RC stand to benefit from a large body of theory in the mature KKL literature, non-exhaustively including on questions of embedding, transverse stability, local and global uniqueness guarantees, and effective data-driven solution constructions. Elaborating on this connection, we show that the surprising forecasting ability of reservoirs is in fact a direct consequence of the well-known observer internal model principle, derive an upper bound on the prediction error over a fixed forecast horizon, and provide a partial explanation for why linear readout training in RC works reasonably well. This work illustrates how classical ideas from systems and control can provide strong theoretical backing and open new questions for modern machine learning methods.

Inter-robot communication is a key enabling technology in cooperative robotics applications. While wireless communication is ubiquitous, simple, and effective, it has inherent limitations: the receiving robot cannot easily localise the source of an incoming signal, clock synchronisation is challenging, and signals are broadcast to all nearby devices rather than targeted recipients. Optical communication provides a complementary channel that is inherently directional and spatially localised, directly addressing these limitations. Event cameras are bio-inspired dynamic vision sensors that respond to changes in image intensity with high temporal resolution, high dynamic range, and low latency, making them well-suited as receivers for high-rate optical communication in cooperative robotic systems. In this paper, we propose the Asynchronous Tracking and Optical Communication (ATOC) system, which integrates LED smart-beacon modulation, event-based detection, optical tracking, and event-data demodulation in a single pipeline. By simultaneously tracking and demodulating multiple LED smart beacons, ATOC transforms a conventional visual marker into a robust communication channel suitable for a wide range of high-impact robotics applications. We validate ATOC in a suite of laboratory studies and two 'applications': a smart-city demonstration and a 3D motion-capture system.

ShadowMiner v1 is a system that automatically discovers research problems and generates hypotheses from AI papers. It is a nine-stage pipeline. It structures documents into a knowledge graph and finds graph gaps in it - structural blind spots in research. These graph gaps are included in the LLM generation prompt. Each generated hypothesis is then verified by checking whether it is already covered by existing research, scoring its quality, and checking that the facts it relies on are accurately drawn from its sources. This report does not propose a new generation or evaluation technique. It describes our experience of implementing and applying ideas from prior work, and measuring whether each one actually contributed.

Machine learning-based weather prediction (MLWP) has achieved strong performance in global weather forecasting. Recent Hierarchical Equal Area isoLatitude Pixelation (HEALPix)-based methods use the HEALPix (HP) grid to avoid area distortion near the poles of conventional latitude-longitude (LL) grids. However, existing HP-based approaches often use pointwise mapping methods and process HP pixels within separate base faces or local windows. Consequently, the mapping may introduce reconstruction errors and cross-face communication depends on handcrafted boundary handling or shifted windows. We propose the Spherical Spiral Scanning Network (S$^3$N) to address both limitations. First, L2Proj provides a bidirectional method for mapping atmospheric fields between the LL and HP grids through an $L^2$ projection of their continuous finite-element representations. Second, the Attention-Guided Quad-Spiral State-Space Scanning (AQSS) block uses cross-latitude attention to guide selective state-space updates along four global pole-to-pole spiral paths. This design enables continuous information propagation across HP base-face boundaries without additional boundary-processing mechanisms. Experiments show that S$^3$N achieves better results at 4-, 7-, and 10-day lead times, and exhibits slower error growth in long-range forecasting.

Calibrating microscopic traffic models for digital twins is an expensive black-box optimization problem: tuning car-following and lane-changing parameters requires a full simulation run, affording only a tight budget per recalibration window. Matching raw trajectories yields a rugged objective that sparse surrogates cannot learn, reducing sequential acquisition to near-random probing. We present FLAT, which couples what to optimize with where to sample next. An eight-dimensional behavioral fingerprint smooths the parameter-error landscape, making the objective learnable from a few dozen samples; annealed lower-confidence-bound (LCB) acquisition then spends each remaining run where it most reduces error. The surrogate, interchangeable among a Gaussian process (GP), random forest (RF), or multi-layer-perceptron (MLP) ensemble, plugs into the same LCB loop. Across six heterogeneous real-world scenes, FLAT-GP achieves the lowest scene-averaged behavioral error, winning 6/6 scenes against SPSA, GA, and CMA-ES and 5/6 against TPE under the matched budget. Some baselines need up to 4.4 times more simulations to match. Ablations show objective choice shifts final behavioral error by 81% on average, removing sequential LCB raises the six-scene mean by 20%, and surrogate choice shifts it by at most 4.2%, confirming gains trace to objective geometry and sequential allocation rather than surrogate capacity.

Recent advances in generative world models have increased interest in digital models that reproduce both the appearance of real objects and their response to physical interaction. Three-dimensional reconstruction techniques, including 3D Gaussian Splatting, capture detailed surface geometry and appearance from images and videos. However, extending these representations beyond plausible animation to mechanically interpretable models for constitutive behavior, boundary conditions, and inverse parameter identification remains less explored. In this work, a differentiable Lagrangian-coupled 3DGS-smoothed particle hydrodynamics (SPH) model is proposed for forward simulation and inverse analysis of deformable solids. The observed object is first reconstructed from multi-view calibrated visual dataset as a 3DGS rendering model. An envelope-based procedure then generates an independent SPH support for the solid-mechanics model, avoiding the direct use of rendering primitives as mechanical particles. A reference-configuration Lagrangian transfer maps SPH deformation to Gaussian positions and covariances, thereby coupling the physical model and the image observation model while preserving a differentiable computational path. The SPH formulation supports linear elastic, hyperelastic, and Kelvin–Voigt viscoelastic responses, together with fixed, free, and Robin-type boundary conditions. Numerical studies validate the SPH response against finite-element results, assess accuracy and efficiency against a conventional model using Gaussian centers as surface SPH particles, and demonstrate forward simulations on beam, bridge, and liver-shaped examples. Inverse analyses further estimate constitutive and boundary parameters from rendered deformation observations, including noisy cases, demonstrating the feasibility of the proposed model for mechanics-based parameter identification from image data.

Remote pose estimation of noncooperative Unmanned Aerial Vehicles (UAVs) from imagery is critical, as they cannot be influenced or instrumented in advance. Deep-learning-based approaches offer a promising solution; however, their development is constrained by the cost and difficulty of acquiring large-scale real-world datasets with accurate pose labels. Synthetic imagery provides an alternative, but models trained on synthetic data must overcome the synthetic-to-real domain gap to generalize to real-world imagery. This work investigates the inherent synthetic-to-real generalization capability of a Vision Transformer (ViT)-based model for monocular UAV pose estimation. The proposed approach employs a self-supervised DINOv2 backbone and is trained exclusively on labeled synthetic imagery while being evaluated on labeled real-world imagery. Pose ambiguity-aware strategies are incorporated during training and inference to address ambiguities arising from the projection of a three-dimensional target onto a two-dimensional image plane and from target symmetries. An $\alpha$-$\beta$ filter is further integrated during inference to improve pose estimations. To assess the model under operational requirements, it is evaluated in terms of Mean Angular Error (MAE) and inference time, both before and after filtering, using a real-world dataset containing 77,077 labeled UAV images. Before filtering, the model achieves an MAE of $19.18^{\circ}$ and an inference time of $13.25$ ms, whereas after filtering, these values are $8.74^{\circ}$ and $13.42$ ms, respectively.

Delayed Streams Modeling (DSM) enables LLM-based streaming automatic speech recognition (ASR) by aligning acoustic and text streams on a common timeline. DSM adds the padding token <p> and the word-start token <w> to the LLM vocabulary and predicts them together with normal text tokens using the same softmax. We first show that <w> can be removed while maintaining competitive recognition performance. Based on this result, we propose Factorized DSM (F-DSM), which separates the waiting probability for <p> from the distribution over the original LLM vocabulary. This factorization removes ASR-specific tokens from the text prediction space and allows the large-vocabulary softmax to be skipped on waiting steps. Experiments on the Corpus of Spontaneous Japanese and LibriSpeech show that F-DSM achieves better recognition performance than DSM. It also greatly reduces GPU memory use while maintaining similar training throughput, provides a small inference speed improvement through softmax skipping, and reduces the degradation in text-only perplexity observed with DSM.

This paper studies intrinsic path following for fully actuated rigid-body attitude dynamics over a broad class of attitude paths, namely arbitrary smooth embedded closed curves on $\mathsf{SO}(3)$. We construct a tangential–transverse state decomposition on a tubular neighborhood using normal and pullback bundle geometry. The resulting second-order dynamics admit a smooth feedback transformation that independently assigns tangential and transverse covariant accelerations. A transverse proportional–derivative law renders the path-following manifold invariant and locally uniformly exponentially stable on an explicit forward-invariant neighborhood, while leaving the tangential dynamics freely assignable. This allows independent regulation of progression along the path, stabilization of a selected attitude on the path, or tracking of a time-varying motion. For a representative fiber path, the tubular coordinates are obtained in closed form up to the natural cut-locus obstruction, and simulations illustrate point stabilization and trajectory tracking.

Reliable language models should resist unsupported user pressure while effectively using objective contextual information. However, models may exhibit sycophancy by yielding to unsupported user pressure or contextual stubbornness by failing to update their answers when relevant contextual information warrants revision. Evaluating interventions for these failures separately can obscure whether mitigating one failure exacerbates the other. To assess this trade-off, we introduce CoPE-Bench with six conditions per question: a neutral baseline, correct or incorrect user pressure, contextual information consistent with or conflicting with the neutral answer, and a joint condition combining incorrect claims with conflicting contextual information. To regulate the influence of user pressure and contextual information, we propose SPAE (Suppressing Pressure, Amplifying Evidence), a training-free framework that uses the model's own judgments to identify relevant tokens, suppressing user pressure and amplifying contextual information through token-level attention steering. On average across five backbones, SPAE reduces pressure following by 18.8 percentage points and increases joint-condition updating by 5.5 percentage points relative to the strongest baseline in the main comparison. In two-turn dialogue, it improves joint-condition updating by an average of 13.2 percentage points over the strongest prompting baseline. The source data and codes can be found at https://github.com/03Grant/sycophancy-and-stubbornness.

Correction-based offline preference pipelines commonly treat model failures only as rejected responses under the original prompt. This supervision is incomplete for boundary failures: responses that violate the given instruction yet coherently satisfy a nearby intent or constraint setting. We introduce Bidirectional Preference Synthesis (BPS), a data-construction method for standard Direct Preference Optimization (DPO) that makes this missing prompt dependence explicit. For each validated boundary failure, BPS keeps the conventional forward pair under the original prompt and adds a reverse pair under a synthesized achieved prompt, so the same response is rejected where it is wrong and chosen where it is right, without changing the DPO objective, training a reward model, or requiring online sampling. On Qwen3-4B-Instruct-2507, BPS preserves original-side pairwise ranking while raising achieved-side ranking accuracy from 6.8% to 62.3% on held-out crossed anchors, with a similar shift under a Kimi-K2.6 cross-teacher probe. A blind human audit supports the intended reverse preference direction, and downstream evaluations show the clearest separation from Forward-DPO in multilingual multi-turn instruction following, with consistent capability-retention patterns on agentic, tool-use, and code checks.

Satellite and street-view imagery see different parts of a damaged building: roofs from above, facades and debris from the ground. Average accuracy hides whether fusion methods use this complementary evidence. We study the conflict cases, the samples on which independently trained single-view models disagree, as a diagnostic subset. On paired collections from the 2025 Eaton wildfire and Hurricanes Ian and Milton, conflicts make up 9.6-33.2% of the test data. Fusion improves conflict accuracy over either view, yet an oracle that selects the correct single view per conflict remains 0.19-0.32 above the best evaluated fusion method. A linear fusion gate over building-visibility features, calibrated confidences, and disagreement recovers part of this headroom where ground photographs are aimed at the building: on Eaton it beats end-to-end fusion by +0.072 conflict accuracy p < 10^-4 and calibrated averaging by +0.051 p = 0.0001; on the panoramic hurricane datasets it is indistinguishable from end-to-end fusion. Cropping panoramas toward the building raises the conflict gain of fusion, whereas random crops of the same size do not. The spatial density of conflicts, computed without labels, correlates with tile-level wildfire damage Spearman r = 0.615, 25 tiles. Conflict cases thus show how much complementary evidence fusion uses, which view it carries under each hazard and capture setting, and where current methods fall short.

Pretrained flow models are now widely used as generative priors in science and vision, where inference-time guidance enables test-time constraints without retraining. Existing methods use projection, posterior sampling, or iterative optimization of the generative trajectory. We propose LyapuFlow, an alternative based on Lyapunov feedback control. At each sampling step, LyapuFlow predicts the terminal sample towards which the current flow is evolving, and evaluates the constraint violation on this prediction. Then, we compute the minimum-norm control that satisfies a prescribed Lyapunov decrease condition. The resulting control remains inactive when the uncontrolled dynamics already reduce the constraint violation at the prescribed rate. Otherwise, it provides a corrective update within a feedback trust region that prevents the control from dominating the pretrained dynamics. We demonstrate LyapuFlow in both data and latent spaces, outperforming alternatives spanning different mechanisms for test-time constraint enforcement in scientific machine learning and image inverse problems.

Transformer models such as BERT and Vision Transformer~(ViT) achieve strong performance via densely parameterized attention backbones. However, the least significant bits~(LSBs) of their 32-bit floating-point weights can be abused as covert channels to conceal malicious payloads, posing a serious threat to the AI model supply chain. We propose \GS (\GSabbr), a lightweight, post-training, zero-data sanitization method that completely replaces the declared mantissa-LSB channel with a Gray-code-guided low-transition sequence. Complete payload-independent overwrite, whether keyed or public, makes the sanitized target bits independent of the embedded payload and gives that declared channel zero capacity. Gray coding supplies overwrite structure, while a keyed per-tensor phase supplies pattern diversity. Benchmarked against seven post-training defenses on four Transformer model presets and two real-world malware payloads, \GSabbr maintains sub-$1\%$ accuracy impact and achieves $49.96\pm0.66$ percentage-point Recovery Reduction (RR) under five implemented attacker variants. Because pre-defense recovery is effectively $100\%$, RR near 50 percentage points corresponds to post-sanitization bit accuracy at binary chance. Its main empirical advantage is stable near-chance sanitization with substantially smaller weight-distribution shift than the evaluated near-chance baselines PatternMask (PM) and Post-Training Quantization (PTQ).

Efficient long-video understanding with vision-language models (VLMs) is often framed as selecting informative frames or visual tokens at a fixed native resolution. We show that per-frame resolution can instead be traded for denser temporal coverage, while front-end decoding latency depends on the size of the candidate pool rather than the final token budget. An empirical study across multiple VLMs and long-video benchmarks yields three findings: dense low-resolution sampling outperforms sparse native-resolution sampling at matched token budgets; resolution-sensitive tasks benefit from selected high-resolution frames; and front-end decoding dominates wall time for hour-long videos. Motivated by these findings, we introduce LoHi, a training-free, single-pass framework that combines a dense low-resolution video stream with sparse high-resolution image frames through the VLM's native video and image pathways. LoHi-Anchor selects high-resolution frames using codec I-frame metadata, while LoHi-SemDiv uses query relevance and visual diversity over CLIP features. Across three long-video benchmarks, LoHi improves average accuracy by 10.6 percentage points over the native-resolution baseline at a matched token budget and by 5.2 percentage points over the strongest prior efficiency method. It also reduces front-end decoding latency by up to 7x on hour-long videos. Project page: https://sixundong.com/projects/lohi

Output-only safety monitoring sees only the end of a model's computation, yet the model computes its answer before emitting it: what it is internally poised to say is safety-critical. Jacobian-space (J-space) readouts, linear maps from hidden states to the output vocabulary via the model's input-output Jacobian, have been proposed as a window on behaviorally accessible internal content, and a first safety protocol (JADR) showed that danger recognition is readable there. What remains unknown is how this accessible content relates to the latent-space safety geometry studied by representation engineering, and how safety training shapes it. We introduce two quantitative bridges: (i) a transport-amplification profile $A_\ell$, measuring how strongly a latent safety direction is carried toward output space by each layer's Jacobian, and (ii) a paired base-vs-tuned protocol that attributes accessibility to training provenance. Across four small model pairs (135M to 1.5B, three lens-fit seeds), tuned checkpoints concentrate recognition in upper-middle layers, and at 0.5B DPO installs refusal while J-space recognition drops to chance: training can widen the accessibility gap exactly where behavioral safety looks best. A deployment audit closes the loop: monitor-aware GCG-style suffixes suppress the prompt-side monitor at zero behavioral cost at every scale; only a learned monitor rung resists its own adaptive re-attack; the training-time defense fails its re-attack at every penalty weight; steering shows the amplification profile is descriptive, not causal; and continuous-prefix optimization fails where discrete search succeeds. Safety monitoring, training, and evaluation must operate on accessibility itself, not on outputs alone.