Papers

Approximate Message Passing (AMP) algorithms are attractive as they are computationally efficient and simultaneously admit a precise characterization in terms of the low-dimensional "state-evolution" recursion. In this work, we establish quantitative universality for AMP with centered rank-one sensing matrices $Z_i=(x_ix_i^\top-I_d)/\sqrt d$, where $x_i$ are independent standard Gaussian vectors. These matrices arise in quadratic regression and learning quadratic neural networks. Their normalized vectorizations have isotropic covariance in dimension $p=d(d+1)/2$, but strongly dependent coordinates. For linear observations with Gaussian noise and prescribed smooth, bounded spectral denoisers, we compare rank-one sensing with a covariance-matched isotropic Gaussian ensemble. At a fixed iteration horizon, with $n\leq Cp$ and uniform bounds on the normalized signal energy, initialization, coefficients, and first four denoiser derivatives, normalized signal overlaps, cross-time overlaps, and smooth linear spectral statistics differ by $O(d^{-1/2})$ in every fixed $L^m$. This yields universality of the normalized mean-squared error, fixed spectral moments, and empirical spectral distributions. Whenever the Gaussian observables admit a joint state-evolution limit, the same predictions hold under rank-one sensing. The proof uses columnwise Lindeberg replacement, leave-one-out analysis, smooth spectral expansions, and high-moment concentration. For prescribed smooth spectral denoisers, the result establishes the AMP universality conjectured by Maillard et al. (arXiv:2408.03733).

The corpus callosum (CC) is a major white-matter structure and a well-established marker of brain aging, but most studies quantify it using scalar summaries that discard its boundary geometry. We present a Riemannian shape-space framework \cite{fletcher2013geodesic-regression} for analyzing age-related morphological change in the midsagittal CC, applied to the OASIS-1 cohort. Each contour is represented by $128$ landmarks and embedded into Kendall shape space, where translation, rotation, and scale are removed. We derive a multivariate geodesic regression with exact Riemannian gradients and use the fitted age-velocity field to localize age-related deformation to five anatomical sub-regions. In the cognitively normal cohort ($n = 252$), geodesic regression outperforms the Euclidean linear benchmark ($R^2 = 0.1355$ vs.\ $0.1216$). Regional energy is anterior-dominant: the Genu carries $46.1\%$ and the Anterior Body $23.3\%$ of total age-related shape change, together accounting for $\sim 69\%$ despite comprising only $\sim 41\%$ of landmarks. Signed projections confirm the ordering (Genu $r = 0.562$; Anterior Body $r = 0.507$). In contrast, age explains less than $0.5\%$ of shape variance in Alzheimer's disease ($n = 88$), indicating that the disease disrupts the healthy aging trajectory. A tangent-space classifier achieves an age-group AUC of $0.791$ from the 2D contour alone, exceeding a recent volumetric benchmark ($0.67$).

In the field of deep learning, has seen significant advancements; however, shadow removal remains a persistent challenge owing to the variable sizes and colors of shadows influenced by lighting conditions. This study proposes a novel shadow feature refinement network (SFR-Net), which leverages supervised learning, feature refinement loss, and knowledge distillation to enhance shadow removal performance. A dedicated post-processing algorithm is further introduced to restore natural color consistency in the generated shadow-free images. We evaluated our method on two public datasets: the adjusted image shadow triplet dataset (ISTD+) and the shadow removal dataset (SRD), which demonstrate strong generalization capabilities under diverse conditions. On ISTD+, our model achieved a root mean square error (RMSE) of 3.4627 and structural similarity index measure (SSIM) of 0.9382 across the entire image. On SRD, it recorded an RMSE of 4.3781 and an SSIM of 0.9341. These comprehensive results show that our approach performs competitively across both shadow and non-shadow regions while setting a promising direction for robust and perceptually natural shadow removal. Code is available at https://github.com/DongHyun99/SFRNet.

Partnered human-humanoid interaction couples locomotion with continuous physical contact. A humanoid needs to coordinate with a person's motion while responding to interaction forces and maintaining stable and natural movement. We present CoDance, a framework for learning reactive and compliant human-humanoid interaction from video. We study partnered dancing as a challenging instantiation, where a humanoid coordinates its footsteps with a moving partner and maintains continuous two-hand contact. Given a single video of two human dancers, CoDance retargets their motions into a robot reference and a moving partner. We introduce a multi-link compliance augmentation that transforms the kinematic demonstration into force-aware training data by adapting the robot reference under structured forces at both hands. Policies trained on this data follow the observed partner while preserving the demonstrated locomotion style and responding compliantly to physical interaction. In simulation, the policies adapt their footsteps to changes in the partner and reproduce approximately 80% of the wrist displacement encoded by the augmented demonstrations. On a physical humanoid, CoDance enables sustained two-hand dancing with a human partner including repeated transitions between forward and backward motions.

Large language model watermarking embeds detectable statistical signals during decoding, but the resulting changes to token probabilities can degrade generation quality. This trade-off is particularly important for code, where small changes in token selection can break syntax or alter program behavior. Existing code watermarking methods mitigate this risk through entropy-based insertion or syntax-aware token selection, but they do not directly construct the watermark over the set of continuations admitted by the current grammar state. We propose Grammar-Guided Code Watermarking with Green Temperature (GTCW), which integrates grammar-constrained decoding with probability-aware watermarking. At each decoding step, GTCW restricts the candidate set to grammar-admissible tokens and partitions this support into keyed green and red subsets. At eligible high-entropy positions, green temperature reweights the green tokens according to the model's relative preferences, strengthening the watermark signal while retaining the grammar constraint. Across five models and five benchmarks spanning four programming languages, GTCW achieves a mean AUROC of 73.61%, compared with 67.83% for the strongest baseline, while maintaining a mean Pass@1 of 59.18% versus 59.58% for unwatermarked generation. Our implementation is available at https://github.com/hyundong98/GTCW .

Test-time compute can improve mathematical reasoning, but can short-budget runs predict how mathematical reasoning scales with additional compute? We introduce a Discovery–Execution (DE) framework that predicts the aggregate held-out scaling curves through a convolution of strategy discovery and conditional execution. From independent short-budget attempts and oracle-sketch-conditioned runs, the framework estimates cumulative success along held-out reasoning trajectories under alternate compute allocations. We evaluate four models on 35 fresh Olympiad problems and non-geometry problems from IMO-ProofBench Advanced. Under the DE framework, near-saturated execution predicts geometric scaling, as observed for the GPT models. For Claude Opus 4.8, incorporating measured execution substantially improves held-out forecasts over geometric extrapolation across one- and two-arm allocations. As a secondary application, regularized DE (R-DE) decisions to continue or restart yield lower average regret than the best model-specific retrospective policy. Together, these results show that measuring conditional execution provides information about longer reasoning that short-budget success rates do not always capture.

AI agents increasingly rely on tools exposed through the Model Context Protocol (MCP) to complete user tasks. Hundreds of thousands of MCP servers are listed across marketplaces, yet they are organized only by coarse, marketplace-specific server categories. This makes it difficult for agents and users to identify tools for a given operation, find functional alternatives, and assess how those alternatives differ. We present MCPacific, the largest tool-level, cross-marketplace map of the MCP ecosystem. MCPacific collects 368,754 MCP server listings corresponding to 124,267 unique servers across 17 marketplaces, statically extracts 1,328,233 tool specifications from these servers in seven languages, and organizes them into a hierarchical functional taxonomy of 58,915 capabilities. We construct the taxonomy through an iterative LLM-driven design-test-refine process and map the full corpus to it using calibrated embedding routing. Our study reveals that MCP extends well beyond developer tooling, with 85% of tools serving other domains. Functional alternatives are widespread but unevenly distributed: 98.5% of tools have at least one alternative, yet nearly a quarter of capabilities are supported by only one tool. Functionally comparable tools also differ in security alerts, code complexity, and project maintenance, with complexity differing by more than 2.5x in 41% of comparable tool pairs. Finally, presenting candidate tools through the taxonomy rather than a flat list improves task completion rate across all four evaluated models, with gains of up to 12 percentage points in Pass@0.75 for crowded candidate sets.

Analog in-memory computing is a promising platform for on-device execution of large language models because it performs matrix–vector multiplications (MVMs) in memory and in parallel, reducing data movement. However, limited digital-to-analog converter precision, input noise, and finite conductance states can degrade model accuracy, while full-model retraining to address these effects can be costly. We develop an optimizer-agnostic, parameter-efficient adaptation method based on Low-Rank Adaptation (LoRA), keeping the pretrained weights stored on analog arrays fixed while training the LoRA weights to adapt to downstream tasks and hardware non-idealities. Reliable adaptation requires handling errors in both forward and backward MVMs and physical weight updates. We use input reshaping to reduce input-induced MVM errors and update accumulation to retain small updates before programming them to finite-state analog devices. Across Llama-3.2-1B-Instruct and Llama-3-8B with both Muon and AdamW, input reshaping improves analog LoRA fine-tuning under noisy MVM computation. Update accumulation separately preserves sub-threshold updates and substantially improves adaptation under finite-resolution programming, including configurations with as few as 20 conductance states. Additional experiments show consistent held-out negative log-likelihood improvements across noisy analog settings.

The PPAD-completeness of computing Nash equilibria in bimatrix games has driven extensive research into polynomial-time algorithms for $\epsilon$-approximate Nash equilibria. Determining the strongest approximation guarantee achievable in polynomial time is a fundamental question in the complexity of approximate Nash equilibria. Despite sustained efforts, breaking the $1/3$ barrier has long been regarded as a difficult open problem. In this paper, we give an affirmative answer by proposing a simple, symmetric deterministic polynomial-time algorithm that has no tunable hyperparameters and achieves an approximation guarantee of approximately $0.30954+\delta$. The main idea is to enrich the optimization framework of Tsaknakis and Spirakis with building blocks drawn from our work on strategy mixing and automated approximation analysis, yielding a new algorithmic structure that provides a conceptual explanation for the improved guarantee.

Pluggable type systems extend the basic type system of a programming language by introducing additional type hierarchies to specify more complex properties and dependencies between variables. We consider the Constant Value Checker, a pluggable type checker for Java implemented using the Checker Framework, which provides various pluggable type systems and corresponding checkers. The Constant Value Checker lets users specify restrictions on the values a primitive integer or boolean variable can hold using a set or range of constant values. However, the current implementation of the Value Checker makes use of a limited set of syntactic typing rules that, in practice, often fail to verify well-typedness of more complex code. In this paper, we introduce an extension to the Checker Framework using Satisfiability Modulo Theories (SMT) solvers to address these limitations by creating corresponding first-order logic formulas for expressions whose type checks fail with the current syntactic typing rules and determining their well-typedness based on the satisfiability of these formulas. Furthermore, we introduce new dependent types for specifying allowed values with expressions that can depend on other variables in the program.

Rossman, Servedio, and Tan proved that the polynomial hierarchy is infinite relative to a random oracle. This paper strengthens this celebrated result by considering strong separations, witnessed by immune languages. The main result of the paper shows that with respect to a random oracle the levels of the polynomial hierarchy separate with immunity. Specifically, with probability one over a random oracle $A$, for every $k\geq 1$ there is a language in $\Sigma_{k}^{P,A}$ that is immune to $\Pi_{k}^{P,A}$, and, symmetrically, a language in $\Pi_{k}^{P,A}$ that is immune to $\Sigma_{k}^{P,A}$. We thus extend classical results due to Bennett and Gill and Vereshchagin. We accomplish this by developing a "rare-or-wrong" approach to strong separations. We highlight the flexibility of the approach by strengthening other separations with respect to a random oracle from the complexity-theoretic literature.

We develop and analyse a conforming piecewise affine finite element method for the Grad–Mercier equation, a nonlocal elliptic model for magnetic equilibria whose source depends on the global rank of a flux function. The principal difficulty is that the unique continuum equilibrium has a maximum plateau of prescribed positive area. The solution therefore samples the rank law only above the plateau mass, while perturbations that separate equal values may distribute this mass over several numerical ranks and thereby alter the non-strict rank source. We exploit the continuum well-posedness structure to replace the rank law by an equivalent clipped, nonnegative law, and replace Lebesgue measure by a mass-lumped nodal measure. Equal nodal values are assigned a common rank interval, thereby retaining the full lumped mass of each tied level. The resulting weighted rearrangement gives a strictly convex discrete energy and hence a unique finite element equilibrium. On meshes satisfying a discrete comparison principle, the discrete equilibrium is nonnegative and bounded above by a computable barrier. A Poisson-preconditioned rearrangement iteration converges on each fixed mesh and admits computable a posteriori energy and energy-norm bounds. As polygonal finite element domains converge to the physical cross-section, the discrete equilibria converge strongly in $H^1_0(\Omega)$ and uniformly on $\overline\Omega$ to the unique Grad–Mercier equilibrium. Reconstructed equilibrium sources converge weak-* in $L^{\infty}(\Omega)$ to the original nonlocal source.

Frozen speech foundation models (SFMs) make downstream adaptation efficient: the backbone can stay fixed while a task learns layer fusion and a lightweight classifier. Full adversarial fine-tuning is a standard route to robustness, but generating adversarial examples and updating the backbone for every task sacrifices that efficiency. We ask whether robustness can instead be learned before future tasks are known. For a frozen backbone and linear classifier, robustness can be understood through the interaction between representation stability and decision-boundary margin. This leads directly to our design: we stabilize representations across the hidden layers, rather than only the final layer, while preserving clean representations; after clean adaptation selects the layer mixture, we keep it fixed and enlarge only the classifier margin, without downstream adversarial examples. We evaluate Wav2Vec2, HuBERT, and WavLM Large on four tasks under adaptive 30 dB attacks. Across 12 backbone-task pairs, hierarchical robustification improves robust accuracy by 46.4 pp, while margin refinement adds 4.0 pp for 1.1 pp of clean accuracy. Code and configurations are available at https://github.com/arefmousavi/hierarchical-robust-sfm.

Rubric-based reinforcement learning (RL) provides interpretable rewards for aligning large language models (LLMs) by evaluating responses against query-specific evaluation criteria. To construct rubrics at scale, a straightforward approach to LLM-based rubric generation is to prompt an LLM to generate a rubric directly from the query. However, rubrics directly generated by LLMs are vulnerable to reward hacking, since omitted or underspecified criteria allow the policy to obtain high rubric rewards with low-quality responses. Existing LLM-based rubric generation methods improve the granularity and coverage of the generated criteria but do not proactively guard against reward hacking. To address this limitation, we propose RubricArmor, an adversarial framework that exposes and mitigates potential reward hacking at the rubric generation stage before it occurs in subsequent RL. Specifically, RubricArmor performs adversarial evolution, in which an attack step and a repair step alternate over multiple rounds. The attack step simulates the reward hacking of the policy by constructing adversarial responses that satisfy the current rubric but fail to properly complete the task. The repair step then revises the rubric to detect the response defects exposed by the attack step while preserving other valid criteria. Extensive experiments demonstrate that RubricArmor outperforms competitive rubric generation baselines and translates into more effective downstream rubric-based RL.

Sign language-text alignment remains a fundamental challenge for text-driven sign language understanding. Existing methods predominantly rely on appearance-heavy RGB representations, which entangle motion semantics with visual variations and lead to ambiguous motion-language grounding. In this paper, we reformulate sign language-text alignment in a structured kinematic space and propose a kinematics-centric framework that adopts 3D SMPL-X motion as the primary representation. By explicitly modeling the kinematic dynamics of signing in a unified motion space, our approach reduces reliance on appearance signals and yields more semantically consistent representations. To capture the compositional nature of sign language, we introduce a gloss-guided local alignment mechanism that leverages gloss temporal spans as weak supervision to decompose continuous motion into coherent segments and establish fine-grained motion-text correspondences, thereby reducing ambiguity in localizing word-level semantics in continuous signing. Furthermore, we develop a visual distillation strategy, where RGB signals serve as privileged supervision during training to provide complementary contextual cues, while being completely removed at inference time. Extensive experiments on standard benchmarks demonstrate that our method achieves state-of-the-art bidirectional retrieval performance on CSL-Daily and competitive results on PHOENIX-2014T. These results highlight the effectiveness of kinematic representations and explicit local grounding for sign language-text alignment.

Large Language Model (LLM) inference has become the dominant workload in modern AI systems, requiring serving infrastructures to maximize throughput while meeting strict latency Service-Level Objectives (SLOs). Since state-of-the-art LLMs exceed the compute and memory capacity of a single GPU, inference is commonly distributed across multiple GPUs using tensor parallelism (TP), pipeline parallelism (PP), or hybrid parallelism (HB). However, selecting the most effective parallelism strategy remains challenging due to complex interactions among computation, communication, pipeline utilization, sequence length, batch size, and model architecture. Existing approaches largely rely on empirical evaluation and provide limited analytical insight into the trade-offs among these strategies, particularly across the distinct prefill and decoding phases of inference. In this paper, we present a unified analytical framework for modeling distributed LLM inference under TP, PP, and HB. The framework decomposes end-to-end latency into computation, inter-GPU communication, and pipeline bubble overhead, and derives analytical models that capture TP collective communication, PP point-to-point communication, and pipeline utilization as functions of hardware, model, and workload characteristics. The model further characterizes the differing execution behavior of prefill and decoding, explaining why PP-oriented configurations favor compute-intensive prefill while TP-oriented configurations reduce decoding latency by eliminating pipeline bubbles. Experiments with modern LLMs on multi-GPU platforms validate the model and confirm the fundamental compute-communication trade-off across parallelism strategies. The framework provides practical guidance for parallelism selection, capacity planning, and optimization of future LLM serving systems.

A large language model (LLM) agent solves long-horizon tasks through many reasoning-action turns, with one verification signal at termination. Deployed agents face streams of related tasks, making their trajectories a natural resource for improvement. In-context adaptation agents store reflections, memories, or skills as text, so reuse depends on retrieving the right experience and on a frozen policy executing it. We study Online Agentic Test-Time Training (OaTTT), which trains the LLM's weights on its own execution trajectories during deployment. The agent executes each task once, in one pass over the stream, and the executed trajectory with its verification result is the only learning signal for weight updates that persist across tasks. Directly imitating or reinforcing the generated tokens of this single attempt destabilizes the policy. We introduce ASCENT (Agentic Self-distillation for Cross-task EvolutioN at Test-time), which instead self-distills verified experience. A stable version of the LLM, its frozen initial copy, receives the verified trajectory as privileged information and predicts next-token distributions along it with this hindsight. Distilling them into persistent LoRA fast weights updates the agent for later tasks, without an external reference solution or stronger teacher. By further removing invalid-action turns, ASCENT distills enhanced privileged experience for more efficient execution. We characterize its population target and the limits of sparse outcome selection. Across ALFWorld, WebShop, and AppWorld at varied model scales, ASCENT improves task success and interaction efficiency as experience accumulates, outperforms online adaptation methods, and transfers to held-out scenes, showing that an agent can consolidate verified experience into its weights without a separate training phase or memory retrieval. Project page: https://artificer-ai-lab.github.io/ASCENT

Generative artificial intelligence (GenAI) is increasingly embedded in university students' academic work, yet student engagement is often examined through adoption, frequency of use, or general perceptions, with less attention to how it is shaped by linguistic and sociotechnical conditions. This comparative qualitative study examines how university students access, incorporate, and evaluate GenAI across three contrasting higher education settings in China, Japan, and Mongolia. Semi-structured interviews were conducted with 42 undergraduates from one university in each setting, and the data were analyzed through qualitative interpretation supported by co-occurrence network and correspondence analyses using KH Coder. Across the three samples, GenAI was incorporated into academic practices spanning exploration, production, and refinement. However, access pathways differed across the three contexts. Participants in China navigated between domestic and global GenAIs, those in Japan predominantly used global GenAIs through Japanese, and those in Mongolia more frequently shifted to English when Mongolian outputs were perceived as less satisfactory. During their use of GenAI, students evaluated its accuracy, privacy implications, potential for dependence, and effects on critical thinking. The findings suggest that technological access alone does not ensure equivalent educational usability across linguistic and sociotechnical contexts. Building on the empirical findings and the analytical perspectives, an empirically informed integrative framework is developed in which student GenAI engagement is understood as situated, mediated, enacted, and evaluated. The framework conceptualizes realized affordances as educational possibilities that become practically usable through the interplay of sociotechnical conditions, linguistic resources, academic practices, and evaluative judgments.

An agent harness is the code that organizes context, maintains state, and coordinates tool calls for a language model. We study how to improve the harness under a limited evaluation budget while keeping model weights fixed. Our method, MESH-Harness, organizes each harness into functional modules with explicit role-specific interfaces, allowing alternative implementations of each module to be substituted and recombined. It uses shared module representations and full-covariance LinUCB to score candidate combinations based on predicted performance and exploration value. Mixed-start coordinate ascent selects complete configurations for evaluation without enumerating the combinatorial space. Validation traces then guide local code edits, and the resulting candidates are incorporated into fixed-capacity role-specific pools for subsequent recombination. On text tasks, retrieval-augmented mathematical reasoning, code generation, and interactive scientific tasks, MESH-Harness outperforms Meta-Harness by 5.70, 7.01, 2.00, and 5.00 points, respectively, under matched candidate-evaluation budgets. Iterative harness optimization improves MESH-Harness by 5.63-7.79 points over its first-round configurations. For the reported configurations, aggregate test-time cost is 44.2% lower than that of Meta-Harness, while total cost including search is 14.6% lower. These results show that combining module-level design reuse with feedback-driven compositional search can systematically improve agent harnesses while keeping overall optimization cost under control.

Motion-deblurring datasets are commonly synthesised by averaging $N$ consecutive high-frame-rate frames and labelling the result with the central frame. We show that this label is unbiased for every capture timing only when the window is odd and the frames' sample durations are equal. An even window shifts every label by a fixed fraction of the blur length, even under perfect timing. Unequal durations are subtler: their mean misalignment is zero, so dataset statistics cannot reveal them, yet when the offset cannot be read from the blur they convolve the supervision rather than adding noise to it. Auditing 51 smartphone clips recorded at a nominal 240 fps, we find 19 captured near 176 fps, a behaviour recorded only in the container's timing tables. In a controlled test the parity choice costs a fitted linear deblurring filter far more than these timing irregularities do, and interpolation labels, unlike blur labels, can be repaired with the true frame times. We provide a tool that checks both conditions without decoding, and will release the clips and their timing tables.

We ask whether a neural network can learn a structure-preserving map from a compressed latent space to a Hilbert-space representation. Using an 8-dimensional autoencoder bottleneck on MNIST and $n$-qubit product-state targets from PCA-based angle encoding, we report four findings. Although the target angles are generated by a nonlinear sigmoid transformation of the latent projections, the resulting mapping is well approximated by a linear function over the observed latent distribution: linear regression from $z$ to the true target angles achieves $R^{2} = 0.98$, while regression to the MLP's recovered angles achieves $R^{2} = 0.91$. The learned map's primary direction is strongly aligned with the target-induced direction, with cosine similarity $0.989$, while remaining nearly orthogonal to the input's principal direction, with cosine similarity $0.002$. The map is genuinely rank-4: removing any singular direction degrades inner-product preservation by $2.5$–$5.9\times$ despite a singular-value spectrum with two dominant and two small values. The learned subspace does not coincide with the PCA basis used to construct the target, and different random seeds recover the same primary direction but diverge in higher ranks. Finally, kernel ridge regression with an RBF kernel outperforms a tuned MLP (IP error $0.0144$ vs.\ $0.0197$), suggesting that for approximately linear structure-preserving mappings, classical kernel methods may be a simpler and more effective alternative.

Most deep learning weather models assign a fixed set of variables and pressure levels to predefined channels, limiting transfer across atmospheric field configurations. This dependence on a fixed field set limits the transferability of trained models across atmospheric field configurations. We propose FlexCast, a field-adaptive weather forecasting model that uses a single set of parameters to produce identity-aligned forecasts for variable-cardinality subsets drawn from a 69-field ERA5 registry. Specifically, a metadata-conditioned adapter the first encodes variable identity, pressure level, and field type and combines them with spatial features. Then, shared rank-16 projec?tions are modulated by metadata-dependent gates to produce field?specific features, while masked set fusion aggregates the available fields into a fixed-width representation. Subsequently, a multiscale U-Transformer processes the fused atmospheric features, while an identity-aware query decoder produces forecasts for the requested fields. Finally, FlexCast learns a standardized six-hour increment and applies it recursively to generate forecasts at longer lead times. Experiments on the 2020 ERA5 test set demonstrate that FlexCast operates across varying field configurations. Compatible cross-field context is associated with lower forecast errors, whereas mismatched context increases them.

Indirect prompt injection causes LLM agents to follow commands embedded in external data. A probe may distinguish instructions from data without identifying a state edit that changes the next action. We study this gap through counterfactual role probes, component-wise activation patching, and separate interventions on AgentDojo trajectories. Role decoding survives changes in content and format. In controlled Qwen tests, it precedes strong tool-choice effects from patches along an independently estimated role direction. On AgentDojo, directions estimated from hijacked and resisted training trajectories reduce attack success at pre-action and injected-span positions, but have little effect at random positions. In longer Qwen trajectories, single-position edits become less effective at later layers; span-wide and repeated edits reduce attack success on the same evaluation set. Removing the learned channel subspace preserves role decoding, yet effective intervention directions transfer poorly across the tested channels. These findings distinguish a readable role signal from an effective behavioral intervention: depth matters in controlled tool choice, while position and context also matter in attack trajectories.

We describe the BossouChimpanzee video dataset, a unique long-term visual record of wild chimpanzees at an outdoor laboratory for field experiments in Bossou, Guinea, spanning three decades (1988-2018) and comprising over 1,200 hours of continuous video recordings collected through collaborative fieldwork and research. In this paper, we outline the history and scientific contributions of the experimental paradigm and video archive, provide key statistics and details on the structure of the main video dataset, and release an initial ~74h snapshot, BossouChimpanzee70h, covering 23 identified individuals focused on chimpanzee individual and action recognition, ahead of the full video resource. This dataset represents a valuable resource for cognitive and behavioural research in ethology and a rich benchmark for training and evaluating machine learning models on audiovisual data from the wild.

What happens to information a pretrained model already encodes when post-training no longer rewards using it? The common language of representation compression conflates three fates: information may be erased, rerouted away from the decision while still represented, or rescaled to occupy less variance while still represented and used. We make these fates identifiable in models whose pretraining recovers Bayesian belief states. A reward that reads only a coarse function of the hidden state defines an exact reward-null kernel. The kernel lets us separately measure whether the information remains recoverable, whether decisions causally depend on it, and how much activation variance it occupies. Theory says what is protected: KL-anchored reinforcement learning preserves the reference policy's log-odds among equally rewarded outputs, supervised and unanchored objectives carry no such constraint, and spectral compression implies neither erasure nor loss of use. In controlled worlds, post-training mostly reroutes or rescales reward-null information and leaves it decodable. Without an anchor decisions can stop using it although the representation survives, and with one they keep using it. Erasure appears only under prolonged weight decay, for distinctions that neither reward nor next-token prediction can see. Open language models show the same dissociation: in-context belief geometry stays decodable under late-layer spectral compression, and within-class behavior depends on the anchor. Post-training thus selects a causal quotient of the pretrained belief state: the reward defines decision-equivalence, the anchor and the state update protect part of what it ignores, and optimization decides whether the rest is erased, rerouted, or rescaled.

In recent years, federated stochastic bilevel optimization has attracted increasing attention due to its wide range of applications in machine learning. To reduce the computational overhead associated with second-order Hessian and Jacobian matrices, several first-order methods have been proposed. However, existing methods typically impose restrictive assumptions on the lower-level function, suffer from a strong dependence on the condition number in their convergence rates, and require different learning-rate scales for variables across the upper- and lower-level problems, limiting their practical applicability and complicating hyperparameter tuning. To address these challenges, we propose a stochastic doubly smoothed gradient method for nonconvex federated stochastic bilevel optimization problems, which decouples the learning rates of upper- and lower-level variables and does not require a strongly-convex lower-level loss function. We establish rigorous theoretical guarantees for the proposed algorithm, demonstrating an improved convergence rate of $O(\kappa^{15/2}/\epsilon^5)$ and a communication complexity of $O(\kappa^{4}/\epsilon^3)$, where $\kappa$ denotes the condition number and $\epsilon$ represents the solution accuracy. Notably, these bounds exhibit significantly better dependence on the condition number $\kappa$ than those of existing methods. Extensive experiments validate the effectiveness of our algorithm.

Recent advances in 3D Gaussian Splatting (3DGS) have achieved remarkable performance in novel view synthesis, yet deploying both static and dynamic Gaussian representations on resource-constrained mobile devices remains challenging due to heavy storage, redundant primitives, and costly per-frame computation. We present Mobile-4DGS, a unified lightweight framework for high-fidelity real-time static and dynamic Gaussian rendering on mobile platforms. For compact appearance modeling, we introduce a Monte Carlo Specular Energy Aggregator that compresses high-order radiance residuals into the first-order Spherical Harmonics (SH), together with an Attribute-Conditioned SH Enhancement module whose predicted offsets are pre-baked before inference. We further propose a Multi-View Alpha-Based Densification and Pruning strategy to suppress redundant primitives while maintaining multi-view consistency. For dynamic scenes, we develop a compact explicit 4D representation by constructing second-order Gaussian motion, learnable temporal support, and a binary static-dynamic partition, enabling continuous-time modeling without runtime deformation networks. Based on this partition, a Depth-Order Certificate selectively reuses previously committed depth orders to reduce re-projection, sorting, merging, and index-buffer updates during playback. Extensive experiments on static and dynamic scenes demonstrate that Mobile-4DGS substantially reduces storage and rendering overhead while maintaining competitive visual quality, enabling real-time 3D and 4D Gaussian Splatting on mobile devices. \textcolor{magenta}{\href{https://xiaobiaodu.github.io/mobile-4dgs-project/}{Code has been released: https://xiaobiaodu.github.io/mobile-4dgs-project/}}.

An intermediate result computed by one model can be reused by other models to perform their tasks. Existing work mainly focuses on practical execution, leaving a theoretical gap in optimizing reuse decisions. This optimization faces two challenges: quality uncertainty, because the effect of reuse on task quality is uncertain across models, and coupled scheduling, because tasks need to share the cost of preparing reusable results. These challenges compound each other: quality must be learned online, but the shared preparation structure makes scheduling NP-hard even with known quality, breaking the key assumption in existing methods. We formulate cross-model computation reuse as an online decision problem and develop the Quality-Aware Reuse Scheduling (QARS) algorithm to address it. For quality uncertainty, QARS learns task-dependent reuse quality from selected, possibly delayed feedback and uses optimistic estimates to guide decisions. For coupled scheduling, it jointly chooses which results to prepare and which tasks should use them, adapting scheduling accuracy to the remaining quality uncertainty. For the considered problem, our analysis separates learning and optimization error in regret and quantifies the tradeoff between scheduling accuracy and computation. Completing the quality-aware stopping rule yields $\widetilde O(\sqrt{T})$ regret while preserving feasibility. Experiments demonstrate the effectiveness of QARS in optimizing cross-model reuse, reducing the combined cost of computation and quality loss by up to 18.0%, and mean regret by 63.9% over the strongest scheduling baseline.

Automated agentic workflow optimization relies on costly evaluations, making it essential to allocate a limited evaluation budget effectively. Multi-parent fusion can reuse designs from previously discovered workflows, but identifying promising parent combinations requires learning from limited fusion feedback. We introduce DAGO (Directed Acyclic Graph Optimization), a contextual-bandit-guided framework that learns which parent workflows to fuse under a limited evaluation budget. DAGO formulates each candidate parent combination as an arm, represented by pretrained embeddings of its constituent workflows' code and prompts. A diagonal LinUCB policy learns a shared reward model across arms and balances exploitation of arms with high predicted offspring quality against uncertainty-driven exploration. After an arm is selected, an LLM generates a child workflow through summary-guided fusion, and the child's validation score serves as the reward for updating the bandit. A shared directed acyclic graph maintains discovered workflows and their multi-parent lineage, providing an expanding pool of parents for subsequent arm proposals. Across six benchmarks covering mathematical reasoning, code generation, and question answering, DAGO achieves the highest macro-average score among the evaluated baselines. Under matched validation-evaluation budgets, it improves over AFlow from 80.3 to 81.7 while reducing aggregate search expenditure by 11.2%. Ablation studies show that LinUCB-guided arm selection outperforms both random selection and its exploration-free variant, supporting the value of feedback-driven selection and exploration-exploitation balance.

Large language models remain vulnerable to jailbreaks, and automated red teaming is the standard way to find jailbreaks in large language models at scale. Current methods either draw more samples at test time through search, rewriting, and tree expansion, or train a stronger attacker offline with reinforcement learning. Both share a limitation: once an attack on a specific target behavior begins, the attacker's weights are frozen. Any signal it gathers about the behavior stays in its context window and is discarded afterward. The attacker never adapts its proposal distribution mid-attack, so success depends almost entirely on the sampling budget, and under a budget affordable at scale, many behaviors remain unbroken. We propose Red-TTT, which updates the attacker's parameters during the attack on each behavior. At each round, the attacker samples a group of candidates, scores them against the victim's replies, and takes a policy-gradient step before drawing the next group, so what it discovers about the current victim is consolidated into weights rather than accumulated as context. We also adapt the training objective to red teaming, where success is judged by the single best sample rather than the average. Red-TTT requires only sampling access to the victim and integrates into existing attack pipelines with no other changes. Against the Best-of-N baseline, Red-TTT raises attack success rate from 55.9% to 72.4% on average at a budget of 120 samples, improving over the baseline in every configuration and cracking many behaviors previous method cannot. The code is available at https://github.com/SaFo-Lab/Red-TTT