Papers

When an agentic prover works on an open problem, there is no proof assistant to fall back on: its verifier and lemma library are ultimately language models judging model outputs. We instrumented such a system end to end and analyzed $51{,}754$ traced observations across three full runs ($186$ hours, \$$5{,}694$). We find three connected failure modes. First, the three-model verifier requires unanimity and treats parse or API failure as non-approval; in $10$ of $12$ verification events, one member returned no parseable output or an API error, making acceptance arithmetically impossible without surfacing an error. Second, when the ensemble did function, one verifier approved $3$ attempts that GPT rejected, each claiming to resolve the open problem; a single-verifier design would therefore have announced a solution three times. Third, because nothing could be approved, every review was a refutation, yet the lemma extractor mines reviews as well as proofs: $24$ of $93$ lemmas ($26%$) were extracted from rejected arguments with their refutational context removed. Taken together, these findings show that without external verification, supervision is itself a critical trust boundary: systems must distinguish abstention from rejection, preserve useful disagreement, and preserve the provenance and polarity of information before it becomes future context.

This paper proposes a label-free physics-informed neural network (PINN) for non-associated Mohr-Coulomb plasticity with incremental strength reduction for the reliability of spatially variable $c$-$\phi$ slopes. The PINN is trained with a virtual-work loss on the finite-element mesh and solves many random fields at once on a graphics processing unit. With a solver budget fixed on the homogeneous slopes, it reproduces the Abaqus factor of safety within 2 to 3% root-mean-square on 52 random fields and locates the failure mechanism in the same weak zones, with the plastic strain in the band under-resolved. A neural operator trained only on the PINN solutions predicts the factor of safety, displacement path and failure mode of a million fields in seconds. No finite-element solution enters the pipeline: Monte Carlo fields, importance samples and operator training data come from the PINN, and Abaqus serves only to verify the result. Monte Carlo simulation, the first-order reliability method, importance sampling and the operator give consistent failure probabilities down to $10^{-3}$ for the slope studied, but a 3% scatter in the factor of safety inflates a probability of $10^{-3}$ by up to a factor of two, and the absolute values refer to one mesh. A heterogeneous Bishop analysis gives failure probabilities 2.6 and 8.3 times lower although its critical circles follow the PINN mechanism, because a small difference in the factor of safety is amplified in the tail. The operator also yields the failure mechanism and displacement response conditional on failure.

Time-series foundation models (TSFMs) enable zero-shot forecasting through large-scale cross-domain pretraining, while retrieval augmentation further improves their performance by leveraging historical information. However, existing methods typically correct TSFM forecasts using the ground-truth futures of similar historical windows, which contain both predictive components already captured by the foundation model and sample-specific random fluctuation that is difficult to transfer. In contrast, recurring systematic model bias within prediction errors more directly characterizes the failure modes of a frozen TSFM and therefore provides more valuable correction signals. Effectively exploiting such model bias, however, poses two challenges: prediction errors at different numerical levels are difficult to compare due to scale differences, and the recurring bias must be extracted from prediction errors contaminated by random fluctuation. To address these challenges, we propose GRAM, a general retrieval-augmented framework for frozen TSFMs. GRAM first introduces an Amplitude Memory Module (AMM) that scales prediction errors by amplitude and aggregates them into retrievable prototypes. It then employs a Prototype Graph Module (PGM) to model relations among prototypes to aggregate consistent bias information while suppressing random fluctuation. During online forecasting, GRAM retrieves and expands prototypes relevant to the current query and generates per-horizon corrections to refine the original TSFM forecast. Experiments across multiple datasets and foundation models demonstrate consistent forecasting improvements.

Empathetic spoken dialogue is a sophisticated cognitive process that requires not only recognizing emotions but also inferring a user's latent mental states to provide appropriate support. However, current SpeechLLMs often treat empathy as a direct input-to-response mapping, leading to "superficially warm" but emotionally hollow interactions. In addition, since empathy relies on a multi-stage process with strong inter-step dependency, errors at any intermediate step can cascade through subsequent steps and lead to inappropriate responses, while existing training paradigms lack mechanisms to precisely localize and improve such errors. In this work, we propose EchoChat, a unified framework that reformulates empathetic spoken dialogue as a structured cognitive reasoning process integrating perception, mental-state reasoning, and response generation. To support this paradigm, we first construct EchoDialogue-400K, an acoustically rich dataset for multi-stage empathetic supervision. During the SFT stage, we strengthen acoustic grounding through proposed Acoustic-Anchored Attention (AAA). During the RL stage, we further introduce a novel stage-aware optimization objective with Step-Decomposed Credit Assignment (SDCA) to localize reasoning errors and mitigate cascaded error propagation. In addition, we introduce EchoEval, an expert-annotated benchmark for multi-dimensional empathy evaluation. Extensive experiments demonstrate that EchoChat achieves state-of-the-art performance in perception, reasoning, and response alignment. Project page: https://github.com/dingdongwang/EchoChat

Fluent guidance is not the same as useful intervention. LLM tutors are typically trained to generate the next teacher utterance, implicitly assuming that every student turn warrants a response. However, our experiments indicate that this conflates tutoring capability (what to say) with intervention necessity (whether to say it). We introduce DICE, a framework that decouples intervention decisions from response generation by first selecting an explicit pedagogical action. To calibrate this action selection policy, we define Intervention Value (IV), a rollout-grounded counterfactual metric that compares each action against non-intervention. IV shows that many prescribed interventions provide little or no marginal benefit. We further introduce DICE-Bench, a multi-variant math tutoring benchmark with skill-preserving problem variants for session-level evaluation. Using IV-weighted and KL-regularized policy optimization, DICE learns to intervene selectively while preserving tutoring effectiveness. In simulated tutoring sessions, DICE reduces the over-intervention rate to near zero while guiding students to correct solutions in approximately 3-4 fewer turns on average than existing Socratic tutoring baselines.

AI agents based on foundation models (FMs) have demonstrated strong capabilities to perform complex open-ended tasks. However, they face some common challenges in practice: (a) agent behavior can deviate drastically even for semantically similar tasks, leading to catastrophically propagated errors; (b) high cost and latency due to FM calls, repeated in full whenever a task recurs with different inputs; (c) FMs' limited context and instruction following capability confine how well agents manage the ever-growing execution context and follow complex plans. We introduce $\textbf{A}$gent $\textbf{B}$ehavior as $\textbf{C}$ode $\textbf{Agent}$ (ABCAgent), which uses a symbolic program (e.g., Python code with potential neural functions) to fully specify the agent's behavior at runtime, with a powerful FM agent editing that program for flexibility. Behavior is thus specified without premature variable binding, and its execution is deterministic. We evaluate ABCAgent on six agent benchmarks, two of which we construct to test how well a derived program generalizes to variants of the task it was written for. ABCAgent matches a model-matched neural agent on GAIA and augmented GAIA, and surpasses it where robustness and long control flows matter: 98.3% against 97.3% on GSM-Symbolic ($p = 0.001$), 71.9% against 47.4% $\mathrm{Pass}^4$ on the telecom domain of $\tau^2$-bench ($p = 0.0001$), and more records written correctly at every loop length on our control-flow-augmented WorkArena benchmark. For more parametric task families, ABCAgent is also significantly superior in efficiency. Without authoring a new program, ABCAgent solves 92.6% of GSM-Symbolic instances and 20.1% of augmented GAIA variants, which yields $5.2\times$ lower latency and $7.0\times$ lower cost on GSM-Symbolic, 19% lower cost on augmented GAIA, and $9.5\times$ lower agent latency on $\tau^2$-telecom.

Adaptive control is usually developed through Lyapunov stability, excitation, and parameter convergence, whereas neighboring fields describe learning using convexity, information geometry, variational principles, and stochastic thermodynamics. This paper develops a constrained common language for these viewpoints. For linearly parameterized models, the excitation Gramian is simultaneously the Hessian of accumulated prediction loss and, under Gaussian observations, Fisher information up to scaling. Persistent, finite, and partial excitation become uniform, finite-horizon, and restricted temporal curvature; memory methods retain previously acquired curvature. On a closed convex parameter set, deterministic adaptation is a projected gradient/Onsager flow, with tangent and normal cones recovering projection, while composite learning supplies data-dependent symmetric dissipation. Reflected Langevin dynamics, no-flux Fokker–Planck evolution, and constrained free-energy flow provide the stochastic counterpart. The framework distinguishes information supplied by data from confinement supplied by regularization or hard constraints and connects familiar adaptive-control structures to modern variational and information-theoretic language.

An announced boundary reveals the partition before the first affected outcome, and a reset there serves one of two audits. The per-time audit reads conformal reliability, the expected distance of the threshold's coverage from nominal at miscoverage 0.1. The frequency audit reads frequency calibration, whether a closed window's error rate matched it. The reset rule, the post-boundary empirical quantile, attains the minimax per-time rate for independent scores. On bounded scores a recursion reset at the boundary attains a frequency gap of order one over the count when its floor adds a bounded total. At the 50th forecast after a large synthetic break, the reset rule's per-time gap stands below that of five adaptive methods, four from their authors' code, with every paired band below zero. After an electricity market reform, on 68 unseen scores, the band stands below zero against two, includes zero against two, adaptive conformal inference among them, and is a point at zero against one. After monetary policy decisions its pointwise band stands above zero against all five on 10 of the first 34 days. Of 9 announced boundaries one clears the smallest coverage difference its window detects, where every method's per-time gap at the 50th forecast exceeds the nominal miscoverage. Over the first 50 forecasts the reset recursion leads the frequency audit after every synthetic break.

Let $n=2^d$. We prove that the worst-start total-variation mixing time of the full labelled Thorp shuffle satisfies, for every fixed $0<\varepsilon<1$, \[ t_{\mathrm{mix}}(\varepsilon) \le C_* d^{5/3}+O_\varepsilon\!\left(d^{4/3}\log(d+2)\right), \qquad C_*=5.1536951454\ldots . \] Time is measured in ordinary Thorp shuffles. This improves the previously published full-deck upper bound $O(d^3)$ of Morris. The geometric starting point is elementary: label the $n=2^d$ positions by binary strings. One physical Thorp shuffle is a random swap layer on one coordinate matching of the $d$-dimensional cube, followed by a deterministic cyclic rotation of the coordinates. In a rotating frame that deterministic rotation disappears, so $d$ successive physical shuffles become one sweep: one independent random swap layer in each coordinate direction. The proof then combines five ingredients. First, a collision-graph representation and negative dependence give pointwise smoothing for ordered marked cards. Second, an entropy argument is sharpened by moving an endpoint moment up to the critical exponent $\beta_c=\log_2(3/2)$, yielding an entropy-preparation clock $s^2/(2\log(3/2))+O(s\log s)$. Third, exact multiplicities in the ordered $h$-tuple permutation representation amplify marked-card bounds into Fourier operator and energy estimates on low representation levels. Fourth, entropy truncation produces a positive mixture whose regular component retains the marked-card budget. Finally, only two long blocks need this entropy regularization: a shorter third block is used solely for additional low-level smoothing. Optimizing the three block lengths gives the constant above.

Frontier language models reliably refuse to explain how to cause physical harm. We ask whether this refusal survives when the same models control a robot. We introduce RoboHarm, a physical AI safety benchmark of harmful robot instructions, and test seven frontier models on it. We hold each model fixed and vary only the interface, from text to Code as Policies to vision agents that command end-effector poses or joint angles. In a MuJoCo simulation twin of a bimanual YAM arm robot, refusal falls from 73% of text requests to 18-25% once the model controls the robot, even when it is told what the objects are. On the physical robot, six models with end-effector control refused 18% of 612 trials against 19% in sim, and three with joint control refused 41% of 252 against 31%. GPT-5.6 Sol refused every text request and none of its trials on the real robot. Case studies of the robot transcripts show that models miss the hazard, name it without inferring its consequences, or act despite recognizing it, and their refusals often hinge on uncertainty about what an object is rather than on the harm of the action. Safety alignment learned in text thus fails to generalize to action, leaving current models with a poorly understood safety gap in the physical world.

Self-driving laboratories (SDLs) are transforming chemical and materials discovery through closed-loop automation, yet automated infrastructure for physical manipulation of soft, deformable matter remains beyond current robotic platforms. A critical instance is autonomous droplet transport on an open surface, where contact-angle hysteresis, capillary pinning, and surface heterogeneity produce partially observable dynamics that pose significant challenges for classical model-based controllers. We introduce the first robotic platform for closed-loop autonomous liquid droplet navigation on an open, unconfined surface using model-based reinforcement learning. A two-axis tilting board coated with a thin silicone oil film drives the droplet, while an overhead camera provides real-time feedback. A learned policy was trained on just 50 to 150 physical episodes depending on geometric complexity, without simulation or analytical models. Beyond performance alone, the platform demonstrates three capabilities of interest to the SDL community: it robustly transfers zero-shot to unseen geometries; it autonomously discovers an oscillatory depinning strategy to free the droplet when it sticks; and it completes its full training pipeline in under 90 minutes. These results extend reinforcement-learning manipulation from rigid microrobots to deformable soft-matter systems for next-generation SDLs.

Quantifying language distance among closely related languages remains a core challenge in quantitative linguistics. Our previous work [1] introduced QuanLing (Quantitative Linguistics via Pretrained Language Models), a quantitative framework combining language distance metrics (sentence embedding distance, tokenization fragmentation rate) with language property analysis (MLM prediction probability), validated on North Germanic (Danish, Norwegian Bokmål, Swedish). This paper extends QuanLing to Western Romance–French, Portuguese, Spanish, Italian–testing cross-branch applicability with the same metric family and aggregation protocol as our North Germanic study, adapted for four languages (English anchor, quadruplet construction). Using 150 four-language parallel sentences, we compute LaBSE sentence embedding distances, tokenization fragmentation rates from four monolingual BERT tokenizers, and mBERT masked language model mutual intelligibility. Results show that Portuguese–Spanish are closest (LaBSE distance 0.0229), French–Italian most distant (0.0338); LaBSE and mBERT rankings agree on 4 of 6 pairs, confirming cross-model robustness. Western Romance shows a wider absolute distance span than North Germanic (0.011 vs. 0.008) but comparable relative ratios (1.48 vs. 1.67), consistent with longer divergence time. French exhibits notably higher MLM predictability (36.12% top-1 accuracy vs. 29.28% for Italian), reflecting its orthography–phonology decoupling. This cross-branch validation provides further evidence for QuanLing's generalizability beyond a single language branch.

Safety-critical Model Predictive Control (MPC) formulations based on Control Barrier Functions (CBFs) often require multiple tuning parameters and may become conservative or difficult to keep feasible, particularly for high-relative-degree constraints. To address these limitations, in this paper we draw inspiration from the recently introduced Taylor-Lagrange Control (TLC) method, and propose an MPC formulation based on a Taylor expansion with Lagrange remainder (MPC-TLR). To enable digital implementation, we develop a numerical approximation of the Lagrange remainder under zero-order-hold control. The resulting safety constraint includes a tunable margin to account for approximation errors and balance safety, feasibility, and conservativeness. Comparisons with MPC using high-order CBFs (MPC-HOCBF), discrete-time high-order CBFs (MPC-DHOCBF), and direct safety constraints (MPC-DC) on a unicycle obstacle-avoidance problem show that MPC-TLR yields safer trajectories than MPC-DHOCBF and MPC-DC, while requiring fewer tuning parameters and being less conservative than MPC-HOCBF. Its broader control-input coverage under the adopted discretization can improve feasibility and control performance for short prediction horizons.

We provide potential function proofs of the competitive ratios of DoubleCoverage for the hard $k$-taxi problem on HSTs and general weighted trees of bounded depth. Buchbinder, Coester, and Naor (2023) obtained these ratios using time-reverse dual fitting and noted that they did not know a pure potential proof beyond $k=2$. We observe that the minimum matching distance between the online and offline taxi configurations can be expressed as an integral over the rooted tree of the absolute difference between their taxi counts below each point. This representation lets us analyse DoubleCoverage over short movement intervals and apply the resulting estimates to potentials that remain unchanged when both serving taxis are relocated from pickup to destination.

Knowledge graph embedding (KGE) methods represent entities and predicates in continuous vector spaces to infer missing knowledge. Despite strong benchmark performance, their predictions often lack principled reliability guarantees, limiting their use in high-stakes applications. Moreover, uncertainty arises throughout the KGE pipeline, from incomplete or probabilistic input knowledge to stochastic training and prediction. This thesis systematically investigates three sources of uncertainty in KGE: knowledge uncertainty, arising from incomplete, noisy, or probabilistic input knowledge; algorithmic uncertainty, induced by randomness in model training; and predictive uncertainty, concerning the reliability of model outputs. To address algorithmic uncertainty, the thesis demonstrates that models trained under identical settings can produce substantially different predictions and introduces a voting-based aggregation framework to mitigate this instability. To quantify predictive uncertainty, it adapts conformal prediction to KGE, constructing answer sets with distribution-free coverage guarantees and extending them to provide predicate-conditional reliability guarantees. To support reasoning under knowledge uncertainty, it develops statistically valid prediction intervals for confidence-scored triples and an embedding-based approach to approximate probabilistic reasoning over statistical ontologies with formal soundness guarantees. Together, these complementary, model-agnostic methods provide a practical and theoretically grounded approach to uncertainty in KGE, advancing beyond predictive accuracy toward reliable and uncertainty-aware knowledge graph reasoning.

Melt-pool monitoring is central to qualifying metal additive manufacturing (AM), yet no public event-camera benchmark exists for this domain. Event cameras report per-pixel brightness changes with microsecond timing instead of reading full frames, giving the temporal resolution AM transients demand at a fraction of the data rate. We present SynAM-E (Synthetic AM Events), the first public multi-source simulated event-camera benchmark for metal-AM melt-pool monitoring: 85 physics-calibrated event shards from 15 sources across 8 institutions, with public baselines and fixed cross-machine evaluation splits. On a single-machine case study, event-spatial monitoring matches dense-frame accuracy (0.874 versus 0.863 macro-F1), and the absolute intensity that events discard adds only +0.006 under fusion. On the NIST Additive Manufacturing Metrology Testbed (AMMT) build, a near-sensor event-rate counter recovers a raw-frame-confirmed 528.7 Hz intensity oscillation at ~380 times less sensor readout than the frame stream requires. A compact 93 k-parameter spiking model runs at 15 times lower modeled inference energy for a 0.073 macro-F1 cost. Every cross-source task includes a built-in trust test against camera identity shortcuts: process-type classification passes while material classification remains confounded by camera band, a corpus-structural limitation the release documents and the trust test exposes.

Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities in visual understanding and multimodal reasoning, yet they remain fundamentally limited in Human Action Feedback Generation. Existing methods infer coaching feedback directly from visual observations, producing generic advice, limited interpretability, and physically implausible hallucinations. In contrast, expert human coaches diagnose performance through explicit biomechanical reasoning over joint kinematics, posture, and body dynamics. We introduce BoT-Feedback, a framework that grounds MLLM reasoning in structured biomechanical evidence. Our key contribution is Biomechanics of Thought (BoT), a four-stage reasoning framework that progressively identifies the action, localises the critical body regions, analyses quantitative biomechanical differences between expert and student performances, and synthesises interpretable coaching feedback. To support this reasoning process, we develop a plug-and-play Biomechanical Data Parser (BDP) that converts videos into structured biomechanical descriptors and an alignment strategy that temporally matches expert and student motions. We further introduce BiomAF, a benchmark containing paired teacher-student videos, 3D skeletons, biomechanical attributes, and expert-coaching annotations. Experiments across twelve open- and closed-source MLLMs demonstrate that grounding reasoning in biomechanical evidence consistently improves feedback quality, interpretability, and robustness while substantially reducing biomechanical hallucinations. BoT-Feedback improves the average expert evaluation score from 2.07 to 2.95 (+40%), enabling compact open-source MLLMs to approach the performance of substantially larger proprietary systems for explainable action feedback generation.

Traditional data pipelines are notoriously brittle, often failing due to upstream schema drift, API contract changes, or website DOM modifications. Present observability tools only raise alerts but for human engineers, resulting in a high Mean Time to Repair (MTTR) and operational fatigue. In this paper we propose AegisFlow (Agentic Engine for Intelligent Self-healing and Graph-driven Operations for Workload remediation), a novel agentic framework that closes the loop between detection and resolution. AegisFlow uses a Watchdog agent to collect runtime telemetry and has a Repair agent to automatically create, test and deploy code patches based on Large Language Models (LLMs). The framework presents the non-intrusive execution model called Parallel Shadow Patching, a non-intrusive execution model based on the Monitor, Analyze, Plan, Execute, Knowledge (MAPE-K) loop to generate and verify patches in digital twin environments. Through experimental testing, we have evaluated AegisFlow across five common failure scenarios, and see 98.1 percent improvement in MTTR (from an average of 170 minutes per patch to 3.2 minutes) and a patch success rate of 92 percent . In particular, the system is successful in dealing with changes in the JSON schema (96 percent ) and punctuation drift (98 percent ), and is least successful in Shadow DOM cases (85 percent ). AegisFlow frees up about 98 percent of data engineering on-call time from firefighting and reallocates it towards innovation. The framework is deployment agnostic consisting of a system that can be deployed in a plugin fashion into an existing pipeline orchestration system with minimal uplift to the existing system.

Egocentric bimanual hand pose estimation is important for virtual interaction, wearable control, and rehabilitation, but visual observations are often degraded by self-occlusion, hand-hand contact, and object manipulation. We propose EVFormer, a multimodal framework that combines the current RGB frame with the preceding 200 ms of bilateral wrist surface electromyography (sEMG) to estimate 44 finger and wrist joint angles. EVFormer separately encodes visual spatial features and sEMG temporal features, enables cross-modal information exchange through sequential bidirectional cross-attention, and integrates the two modalities using feature-wise gated fusion. We evaluate EVFormer in a single-participant feasibility study using one synchronized public EgoEMG recording with chronologically separated training, validation, and test splits. On 296 test samples, EVFormer achieves a mean absolute error of 11.482 degrees, compared with 13.228-13.610 degrees for vision-only, sEMG-only, late-fusion, and training-mean baselines. This corresponds to relative error reductions of 13.20% compared with the vision-only model and 14.23% compared with late fusion. EVFormer also achieves the lowest error in four of the five evaluated gesture classes. These results provide preliminary evidence that feature-level interaction between egocentric vision and sEMG can improve bimanual hand pose estimation. Further evaluation across participants, recording sessions, sensor placements, and real-world interaction conditions is required to establish the generalizability of the approach.

Fuzzing is effective at finding bugs in deep learning (DL) compilers, but existing fuzzers repeatedly trigger faults they have already uncovered. Current fuzzers diversify the generated programs, and downstream tools deduplicate bug reports post hoc, but neither stops a fuzzer from generating programs that re-trigger known defects. We present Reprise, a DL compiler fuzzer that suppresses reports of known defects during test generation. Reprise uses a deliberately lightweight generator that builds graphs in a unified intermediate representation (UIR) with local operator signatures. Instead of diversifying programs, Reprise distills each discovered defect into a semantic graph pattern that captures the operators, value constraints, graph context, and data flow required to trigger it. During graph synthesis, it regenerates any node that completes a known pattern, so known defect triggers are avoided before compilation and execution. We evaluate Reprise on three DL compilers: TVM, PyTorch Inductor, and ONNX Runtime. Against its unguided variant, patterns written for the dominant known defects reduce crash reports by 90.2% on TVM (254 to 25) and by 93.8% on PyTorch Inductor (594 to 37), with comparable branch coverage and 1.0-17.9% fewer executed tests. These results come from one run per configuration and concern the defects the patterns were distilled from. Reprise also found 25 previously unknown bugs across the three compilers, 24 of them within one month, of which 7 have been fixed and 12 more confirmed by developers. The replication package has been provided at [11].

Indirect prompt injection (IPI) hides adversarial instructions in content that large language model (LLM) agents read at runtime. As agents compose heterogeneous capability units, including Tools, MCP servers, and Skills, the carriers of injection multiply, and defenses built to recognize attack patterns fall behind them. We instead shift defense from covering attack patterns to one stable point: whatever the carrier and however the injection propagates, harm materializes only at the execution boundary, where the agent turns internal state into an external action or released output. Safety there turns on two conditions, both settled by the trusted task rather than by the run: whether the proposed effect is authorized, and whether the runtime information reaching it is endorsed by that task. We present APEX, an active defense that enforces both at this boundary from a single authorization contract compiled before untrusted execution: evidence-gated prevention admits an effect only when the contract justifies it, while deception-based exposure makes unendorsed use reveal itself before the effect commits. Protection therefore follows from what the task permits rather than from how an attack is built, and applies uniformly across capability units without attack-specific policies or taint tracking. Against 13 baselines, APEX attains 0% attack success on five of six benchmarks and 0.56% on the sixth, holds 0% under adaptive attacks on all three capability-unit types, and remains effective across defender backbones. Code is available at https://github.com/ZhengXR930/APEX_official/tree/official.

World action models jointly learn visual predictionand robot actions, providing a way to use observations ofscene evolution for policy learning. Their video and actionlosses, however, provide no explicit target for the geometricconsequences of a demonstrated action sequence. Moreover,visual features taken after temporal attention can contain futureobservations, making them unsuitable as the sole current visualinput to an auxiliary predictor. We introduce ACG-WAMand its auxiliary objective, the Action-Conditioned GeometricJoint-Embedding Predictive Architecture (ACG-JEPA), whichpredicts geometric features at several horizons from the currentobservation and intervening actions, using the future slot of afrozen VGGT encoding of each current and future image pairas the target. We apply this supervision from the head and wristcameras to a shared visual embedding before temporal mixing,and remove the teacher and auxiliary modules at inference.On 50 RoboTwin 2.0 tasks, ACG-WAM achieves 93.46%success in clean scenes, with the best randomized success(92.68%) and mean across both settings (93.07%) among thecompared methods; across three tasks on a real robot, itachieves 85.00% success and 91.67% partial completion score,exceeding Motus by 10.00 and 9.17 percentage points, respec-tively. Code:https://github.com/RoboOpus/ACG-WAM.Website:https://RoboOpus.github.io/ACG-WAM.

Large language model (LLM) agents have demonstrated strong capabilities in interactive environments, yet their ability to continually evolve from experience remains limited. Although fine-tuning enables adaptation, its dependence on parameter access and high computational costs restrict its flexibility, especially for large-scale and closed-source LLMs. External memory offers an alternative by allowing agents to accumulate experience without modifying model parameters. However, existing methods mainly focus on experience representation and organization, while the acquired knowledge remains tightly coupled with specific tasks and contexts, limiting generalization. A key challenge is how to transform concrete interactions into abstract and reusable knowledge that guides future decisions beyond individual experiences. To address this challenge, we propose SAGA (\underline{S}elf-evolving \underline{A}gents through Experience-\underline{G}rounded \underline{A}bstraction), a framework for experience-grounded knowledge abstraction and utilization in LLM agents. SAGA progressively transforms interaction trajectories into episodic descriptions, reusable procedures, and principles with explicit applicability conditions, while maintaining links to execution evidence. Retrieved principles are instantiated into task-specific guidance and used to refine candidate actions through corrective feedback and resampling. This creates an execution–abstraction feedback loop, where accumulated knowledge guides future interactions and new experiences continuously update hierarchical memory. Experiments on ScienceWorld and ALFWorld demonstrate improved task performance, with ablation studies highlighting the importance of contextual instantiation and action regulation for leveraging principle-level knowledge.

Rotary Position Embedding (RoPE) encodes token positions by rotating each two-dimensional channel of the query and key vectors at a channel-specific frequency, making the attention logits invariant to a common shift of positions. However, this rotation is periodic, and it leads to position aliasing where relative positions separated by a full rotation period become hard to tell apart. To address this, we propose WavePrune, which restricts each channel to its first rotation period. We show that it removes the distractions in attention maps created by position aliasing and improves overall long-context performance. Specifically, WavePrune raises the HELMET score on four of five models we test without any extra tuning (e.g., 35.7 -> 40.0 on Qwen3-8B). When pretraining models from scratch, WavePrune also achieves lower validation loss at extrapolated lengths than pretraining without it. Because WavePrune restricts each channel to a sliding window, it induces a fine-grained sparsity that our hardware-aligned CUDA kernels exploit for 1.15x prefill and 1.24x decoding speedups over FlashAttention-2 at 32K context. Together, these results show that RoPE's periodic structure, widely regarded as essential, is largely redundant beyond the first rotation period.

In many review workflows the verdict is the only thing retained. The passages behind it are not marked, because that annotation costs far more than recording the decision. We measure how much of that evidence a small language model can recover when it is post-trained on the verdicts alone, with no human evidence labels at any stage. On ContractNLI the human evidence spans are held out until evaluation. Matching the recorded verdict and agreeing with those spans are not the same thing: across six systems the two scores are only weakly related and rank the systems differently, so accuracy is a poor guide when the citations have to be reviewable. Label-only training on the bare verdict reaches accuracy 0.896 and span F1 0.564. Rejection sampling, which keeps a generated trace only when its verdict matches the record and then picks one by an automatic source-grounding score, reaches 0.797 and 0.556, against 0.747 and 0.493 before training. Verbatim citation rises from 0.597 to 0.729 under label-only training and to 0.701 under rejection sampling. One seed on one corpus cannot say which method is better, but both improve the evidence without anyone annotating it.

We present DistScene, a framework for single-image compositional 3D scene generation by jointly modeling the environment and individual objects. Unlike existing methods that represent scenes primarily as collections of objects, we model the environment as an explicit scene component to provide geometric context for object placement. Specifically, we introduce Scene-Frame Generation, which jointly generates separate environment and object components in a shared coordinate frame, allowing their geometry and relative placement to be learned together. Then we introduce Object-Centric Refinement to refine each object in a local frame with scene context. Finally, we develop Object-to-Scene Distillation to transfer pretrained object-generation priors to scene generation through automatically composed and rendered synthetic scenes. Evaluations on indoor and outdoor benchmarks demonstrate improved scene-level spatial coherence over the evaluated baselines. Project page: https://coolbeam.github.io/DistScene/

We study the fundamental problem of computing Personalized PageRank (PPR) on an undirected and unweighted graph $G$. We focus on the $\eps$-error guarantee introduced by Andersen, Chung, and Lang [ACL; FOCS 2006 $\&$ Internet Math 2007]. Their classic local push method computes an approximate PPR vector satisfying the ACL $\eps$-error guarantee in $O((\alpha\eps)^{-1})$ time, where $\alpha$ is the teleportation parameter. In this paper, we eliminate the dependence on $\alpha$ and present an algorithm that achieves the same error guarantee in $O(\eps^{-1-o(1)})$ time. Consequently, we obtain a nearly optimal algorithm for PPR computation under the ACL error guarantee, and a nearly linear algorithm for local graph clustering with complexity independent of $\phi$.

Object-goal navigation (ObjectNav) in multi-floor scenarios presents a challenge due to sparse rewards caused by long-horizon decision-making. In this paper, we propose a diagnostic study based on a modular framework with an effective learnable policy to analyze failure factors in multi-floor scenarios. To achieve an effective policy for diagnosis, we design the hierarchical factorization policy that deconstructs a single global policy into an intra-floor exploration policy and an inter-floor switching policy. To providing an effective initialization for Reinforcement Learning (RL), the lightweight intra-floor policy is learned by distilling the exploration logic of Visual Language Models (VLMs). Under idealized assumptions, we show that the factorized policy is theoretically equivalent to a single global policy at the policy-representation level. Experiment results indicate that perception performance and stair climbing stability are the primary bottlenecks in multi-floor navigation.

Large speech language models have demonstrated strong capabilities in unified cross-modal understanding and generation, yet paralinguistic cues, especially emotion, remain difficult to preserve. Existing systems typically rely on entangled acoustic representations, which allow the underlying language model to depend excessively on recovered lexical content instead of grounding its behavior in acoustic-prosodic evidence. We address this limitation with EMODE, an emotion-aware speech language model built around Dynamic Para-Semantic Experts (DPSE). DPSE decomposes continuous speech features into semantic and paralinguistic pathways, routes them dynamically, and fuses them before integration into the language model. To turn this structural decomposition into functional specialization, EMODE is trained with a three-stage curriculum consisting of semantic warm-up, paralinguistic activation, and joint refinement, guided by Orthogonal Expert Guidance (OEG), Semantic-to-Acoustic Alignment (SAA), and Gating Diversity Regularization (GDR). Experiments on SER test, empathetic response evaluation, and the newly constructed bilingual MEPA benchmark show that EMODE improves the balance between lexical fidelity and emotional sensitivity, strengthens affect-grounded response generation, and exposes the value of explicit para-semantic factorization for robust cross-corpus emotion understanding.

Humans inherently understand the physical world through an active process. When sensory evidence is insufficient to infer physical properties, we naturally interact with the environment by deciding what information is missing, how to acquire it, and when sufficient evidence has been obtained. In stark contrast, existing multi-sensory robot systems mainly integrate sensory inputs rather than actively acquiring missing evidence through interactions. In this work, we introduce ROMA, an LLM-based system for Real-World Object-Centric Multi-Sensory Active Perception. ROMA integrates vision, audio, tactile, and force sensing into a reasoning-interaction-feedback loop. The model identifies missing evidence and determines the target objects, interactions, and modalities, while a physical interface executes the selected interactions and collects the multi-sensory feedback. To support this capability, we construct ROMI-2K, a large-scale real-world multi-sensory object interaction dataset covering nearly 2,000 objects and 6 atomic interactions with synchronized sensory feedback. Building on these data, we develop a two-stage training framework that aligns sensory modalities and equips the LLM to assess evidence sufficiency, select informative interactions, and reason over the multi-sensory feedback. We further characterize active perception as perception chains, where acquired evidence guides subsequent interactions and reasoning, and establish ROMA Bench to evaluate single-attribute, long-horizon multi-attribute, and intent-driven active perception. Experiments show that ROMA can actively acquire missing evidence and solve complex, long-chain multi-sensory perception tasks that existing methods struggle to handle, laying a strong perceptual foundation for active multi-sensory embodied agents.