Most parameters of a Mixture-of-Experts (MoE) model belong to its experts, and standard training keeps every expert gradient in GPU memory until the optimizer step. This memory grows with the number and size of the experts, so it rises as MoE models scale. We introduce Sonar-MoE, which updates each expert tensor during the backward pass and frees its gradient right away. It brings optimizer-in-backward from dense models to batched MoE expert tensors, so the GPU never holds more than one layer of expert gradients. When we fine-tune OLMoE-1B-7B and Granite-3.1-3B-A800M in bf16 with two expert-kernel backends, Sonar-MoE lowers peak memory by 19% to 22%. Its updates are bitwise identical to standard training without expert gradient clipping, and held-out loss stays within 0.003 of all baselines. Each step also takes 17% to 23% less time than standard training with global-norm clipping. Updating whole batched tensors rather than individual experts is what keeps Sonar-MoE fast, as it avoids a 20% to 72% slowdown while using the same memory. Sonar-MoE matches optimizer-in-backward over all parameters in memory and speed while still clipping the dense parameters.
Long-horizon LLM agents make multi-step decisions, yet outcome-reward-only reinforcement learning provides little direct supervision for individual steps. We introduce AttnSR, a step-level reward augmentation method that derives reward allocations from policy attention extracted during training replay, without a separately trained reward model or additional rollouts. AttnSR traces backward through the policy model attention graph to estimate the relevance of earlier steps to the final response. The terminal step, measured relative to a running baseline, determines whether these rewards reinforce or penalize the selected steps. In a controlled Qwen2.5-7B-Instruct study, AttnSR consistently outperforms the PPO baseline across all evaluated checkpoints on ALFWorld, raising final held-out success from 91.2% to 97.1%. The ablation experiment shows that the benefit depends on how intermediate rewards are assigned, rather than simple accumulation of step rewards over time. AttnSR also improves ScienceWorld at every checkpoint while cutting environment interactions by 29.9%. These results support policy attention as a practical prior for step-level reward allocation in long-horizon agentic RL. Code is available at https://github.com/fairyshine/AttnSR .
Cheng Yu, Jim Chi, Aaron J. Snoswell, Orestis Papakyriakopoulos
The tension between news personalization and information diversity has traditionally been understood as a matter of different content reaching different people, with diversity evaluated by the range of sources, topics, and perspectives to which audiences are exposed. Generative AI outdates this understanding, since it transforms personalization in two ways: how journalists adapt news to audiences and how the AI tools journalists use personalize content production to their own preferences. Drawing on 16 semi-structured interviews with media professionals across public service, commercial, and state-affiliated news settings, we examine how these practices reshape editorial judgment and the conditions under which news diversity is produced. We find that participants negotiate AI integration by maintaining authority over meaning while delegating aspects of content presentation. Yet the tools they adapt for their own work also mediate which sources and perspectives become available, introducing diversity-relevant decisions that remain largely invisible to organizational oversight. While participants anticipate diversity risks in audience-facing personalization, they predominantly interpret diversity-imposed limitations in their own tools as technical problems. As a result, decisions that shape news diversity increasingly occur before publication and across multiple versions of a story, where neither evaluation of published items nor human approval can readily observe them. Journalists thus remain formally responsible for outcomes shaped by decisions they cannot fully see or control. Our findings highlight that safeguarding news diversity requires looking beyond published content to the professional practices, infrastructures, and distributions of control that shape it.
We propose FluxC, a zero-shot codec that uses Flux as a compression prior. Although DiffC has shown that pretrained diffusion models can serve as zero-shot codecs, we ask how the choice of coding path and reconstruction method affects rate--distortion--perception behavior. We analyze FluxC along three axes: coding cost and bit allocation, reconstruction method, and low rate distortion. At matched signal-to-noise ratio (SNR), we show that the path choice does not affect the ideal information rate, but changes how bits are allocated across time steps. For reconstruction, we compare ancestral sampling, one-step ODE reconstruction, and many-step ODE reconstruction. Our analysis shows that ODE-based reconstruction can achieve lower distortion than ancestral sampling, especially in the low rate regime. Experiments on Gaussian sources and natural images support this prediction. Moreover, FluxC performs competitively with, and in some cases outperforms, recent neural image compression methods, suggesting that Flux can serve as a principled zero-shot codec.
Petru Neague, Quinten Stokkink, Martijn de Vos, Rafael Pires, Mathis Randl, Anne-Marie Kermarrec, Johan Pouwelse
Retrieval-augmented generation (RAG) is a standard approach to enhance the reliability of large language model (LLM) outputs by retrieving documents relevant to a query and supplying them in the LLM prompt. Today's RAG systems collect, index, and serve these documents from centralized stores. In practice, however, documents are owned by organizations, devices, and individuals who cannot or will not surrender custody of them to a central party, e.g., because of privacy, copyright or regulatory compliance. Existing modern decentralized retrieval techniques either reintroduce a coordinating authority, or rely on a global data structure that requires continuous network-wide coordination (e.g., DHTs). We present ZeroRAG, a decentralized routing layer for document retrieval in RAG that performs semantic query routing across an overlay of peers with- out global data structures or centralized orchestration. Documents remain with the peers that own them, and each peer keeps only lightweight local state: a bounded set of neighbor advertisements and a cache of recently routed queries. Using these, ZeroRAG steers each query through a best-first walk that progressively rewires the overlay toward semantically relevant neighbors, so recall improves over time as the network self-organizes. We evaluate ZeroRAG using three real-world workloads and two decentralized baselines. We find that ZeroRAG outperforms baselines in terms of recall and required network messages, approaches the recall and LLM answer accuracy of a centralized document store, self-heals after catastrophic peer failures, and scales to 100 000 peers.
Engagement features developed for English are routinely applied to Arabic, but unvocalized Arabic changes what a lexical feature measures. We introduce AraEngage, a query-defined sample of 10,000 public Arabic Facebook news posts from 2,451 pages in 57 countries (2023-2024) that pairs engagement counts with interpretable linguistic features. We use it to examine which interpretable Arabic indicators are associated with follower-normalized engagement, and how far those associations can be trusted given publisher context, the definition of engagement, and lexical ambiguity. Religious expressions, message length, sentence length, and lexical density are the leading correlates in a Random Forest model ($R^2$ = 0.38 on a post-level split) interpreted with SHAP. Unlike English news, positive sentiment is associated with higher engagement in this Arabic news sample. Message length shows a length paradox: unadjusted raw engagement is highest for long posts, whereas the model's attributions for follower-normalized engagement favor short posts. A linguistic audit identifies why standard lexicon features are fragile in unvocalized Arabic: imperative lists contain homographs such as the form read as "express!" that is also the frequent preposition "via", hedge lists conflate the possibility marker qad + imperfect ("may") with the emphatic qad + perfect ("has indeed"), and religious lists mix devotional formulas with place names in political news. These are plausible sources of measurement error that may attenuate the near-zero associations of imperatives and hedges. All findings are observational associations with operationalized indicators. We release the derived linguistic features, engagement counts, refined lexicons, and analysis code; raw posts are not redistributed, in line with CrowdTangle terms.
Simple cycles encode feedback and alternative routes in networks and are important for routing problems, the convergence of dynamical processes on networks, message passing and belief propagation, or even static network characterizations such as structural balance in signed networks. Being able to assess and sample from the cycle structure of a graph is thus of interest in many analysis contexts. However, due to the combinatorial growth of the cycle space with graph size, the computational cost for such sampling is high. Therefore, we here introduce a sampling framework that enables us to sample from the cycle structure without enumerating all cycles. Specifically, the method draws uniform spanning trees and collects the fundamental cycles created by their non-tree edges, such that each tree provides an informative subset of simple cycles. Because cycles are included with unequal probabilities and are dependent within each subset, we formulate an inclusion-probability-based Horvitz-Thompson estimator, which we average over random subsets, a procedure we call Monte Carlo Horvitz-Thompson. For this, we sample multiple Horvitz-Thompson estimators and introduce a method to further subsample the individual estimators for increased computational performance. For sampling cycles, the estimator is unbiased if exact cycle inclusion probabilities are known. To extend the method to larger networks, we derive an efficient mean-field approximation for the inclusion probabilities and characterize the graph regimes in which it remains accurate. Experiments on synthetic and real-world networks demonstrate the computational scaling and statistical performance of the framework. We then illustrate the utility of Monte Carlo Horvitz-Thompson estimation by estimating cycle counts and cycle-level frustration statistics in signed spin systems. An implementation is available in the py-raccoon package.
Text-guided image editing with diffusion models struggles with spatial edits (e.g. manipulating objects, changing viewpoints, manipulating occlusions) that depend on the 3D structure understanding of a scene. Marigold and follow-up work demonstrate that a pretrained diffusion backbone already harbors strong geometric priors that can be activated for depth estimation, yet how to redirect these priors toward editing remains unresolved. Through per-layer gradient probing, we discover that the editing and depth objectives are compatible in early transformer layers but actively conflict in late layers, explaining why naive depth-conditioned editing either ignores geometry or degrades visual quality. We present StitchEdit, which first activates the backbone's geometric capability via a lightweight depth-denoising adapter, then selectively fine-tunes this adapter for editing using a probe-guided per-layer learning rate schedule that preserves depth knowledge where it helps and prunes it where it conflicts. On the proposed StitchEditBench, a benchmark of 150 geometry-dependent edits spanning spatial transformation, occlusion manipulation, and viewpoint change, StitchEdit consistently surpasses both instruction-based and depth-conditioned baselines on perceptual quality and geometric fidelity.
Jiaoyang Ruan, Xin Gao, Yinda Chen, Hengyu Zeng, Liang Du, Guanghao Li, Jie Fu, Jian Pu
While Diffusion Large Language Models (dLLMs) offer structural advantages for global planning, efficiently verifying that they arrive at correct answers via valid reasoning traces remains a critical challenge. In this work, we propose a geometric perspective: Reasoning on the Manifold. We hypothesize that valid generation trajectories reside as stable attractors on the high-density manifold of the learned distribution, whereas invalid paths exhibit off-manifold drift. To operationalize this, we introduce Bidirectional Manifold Consistency (BMC), a training-free, unsupervised metric that quantifies the stability of the generated sequence through a forward-masking and backward-reconstruction cycle. Empirically, we demonstrate BMC's versatility across the full reasoning lifecycle: (1) in Diagnosis, it serves as a robust discriminator of solution validity without ground truth answer; (2) in Inference, it enables rejection resampling to effectively concentrate computational resources on complex reasoning tasks; and (3) in Alignment, it functions as a dense geometric reward that transforms sparse outcome supervision into fine-grained guidance, empowering models to self-evolve beyond standard baselines. Our results establish intrinsic geometric stability as a robust indicator of correctness for dLLMs.
A child who learns to add small numbers adds numbers of any length. LSTMs and Transformers trained on short sequences fail on longer ones. We trace the difference to what a learner reads when it decides. A position, a count or a time keeps taking new values as the input grows: it is a coordinate, and a learner that reads one meets in every long input values its experience never covered. Whether two things are the same has the same two answers at every size: it is a coincidence, and a learner that reads only coincidences finds in a long input nothing that short inputs lacked, so finite experience fixes its rule for every size. The memory carries the size of the input and the rule stays finite, and this determines the memory a generalizing brain can have: a memory that grows while what the learner reads of it stays finite has to look the same from every place, so it is made of cycles, and the past is reached by asking what happened when the cycles last stood where they stand now. The cycles are the grid modules, the cells that detect their return the place cells, the records the hippocampal index; a prefrontal controller moves the cycles and never sees them. Reading only coincidences costs no power: what finite experience fixes in such a learner is what Turing machines compute in step with their input, which is why a circuit that evolved to find the way home can learn arithmetic and grammar. In a network of neurons this machine of cycles is the dynamics seen through its symmetry, and symbols are its invariants. Trained on strings up to 40 symbols, the machine is at least 99% correct, on average over lengths 41 to 500, on nine of fifteen benchmark tasks, five of them tasks that no LSTM or Transformer solves; given new cycles after training it reverses strings of 15,015 symbols without error and fails beyond that on the curve the theory predicts; and its controllers run unchanged on neurons.
We present a novel approach for model checking fault-tolerant distributed algorithms: instead of building large, monolithic algorithm models, we propose a modular approach where the user decomposes the algorithm into modules, verifies them individually, and replaces them with a synchronous abstraction. To enable our technique, we introduce a new system model that combines aspects of threshold automata and global synchronization protocols. We investigate the expressiveness of our new model, and present a (bounded) model checking procedure as well as a method for checking the correctness of the abstraction step. We implemented the technique and report on promising experimental results on a case study
Reasoning trajectories generated for the same problem often arrive at conflicting answers. These disagreements reveal alternative intermediate decisions that can be reused to explore new solutions. We introduce Adaptive Semantic Branching (ASB), a training-free inference procedure that turns cross-trajectory disagreement into targeted candidate generation. ASB first groups independently sampled trajectories by extracted final answer, selects a representative from each group, and aligns their reasoning steps with a monotonic dynamic program. A persistent-divergence objective identifies the point at which initially similar reasoning paths develop different continuations. For each selected directed pair, ASB retains the source trajectory's textual prefix, extracts a short reasoning window from a competing donor trajectory, and generates a fresh solution conditioned on both. A diversity-aware policy allocates up to three additional generations according to the answer-cluster structure. In a 500-problem MATH evaluation with Qwen2.5-7B-Instruct, ASB increases oracle candidate coverage from 445/500 (89.0%) with seven initial samples to 455/500 (91.0%), recovering correct candidates for 10 of the 55 initially uncovered problems. The method uses 576 additional generation calls across the evaluation set, averaging 8.152 calls per problem. We provide a precise specification of the branching operator, alignment objective, generation policy, and experimental configuration.
Lightweight methods like QuickCheck have proven to be successful in functional languages and beyond, as they let users specify properties as Boolean functions directly in code. We can replicate some of their benefits for (bounded) proofs via symbolic execution. If we can execute arbitrary functions symbolically, we can specify properties just as easily. Unfortunately, existing symbolic execution tools for Haskell are restricted to a subset of the language that excludes many real-world applications like IO; they also have a hard time scaling to larger programs, making them a poor fit for general purpose lightweight verification. One reason behind these limitations is that existing techniques only offer fixed symbolic semantics for SMT base types. This makes it hard to cover more programming constructs and libraries; it also makes it difficult to fine-tune symbolic representations for better performance. We present SymFC, a new symbolic execution method for Haskell that offers configurable semantics via embeddings. An embedding of a function in SymFC is just another Haskell function that can use arbitrary code, as well as SMT primitives. Embeddings make it easy to support new programming constructs and fine-tune symbolic representations for speed. To integrate embeddings into Haskell's type system, we make use of Haskell's coercions, which let us transparently convert between a Haskell type and the type of its embedding. We realize embeddings in SymFC, an all-paths symbolic execution engine for Haskell, which lets us symbolically represent arbitrary types and values in GHC core. We implement SymFC as a GHC plugin, and evaluate it by checking properties of Haskell's container library, and the iproute library. We build embeddings for IORef, Haskell's Ptr/ForeignPtr interface, and the hardware library Clash, and our evaluation reports a 6.25x speedup on G2's own benchmarks.
Multi-agent systems (MAS) combine multiple agents to solve complex tasks, and recent MAS scale to more than a thousand agents. At this scale, decentralized MAS remove the central orchestrator that assigns sub-tasks, and the agents, which we call workers, self-organize through their own choices of sub-tasks and collaborators on a shared board. As workers keep adding participation and collaboration records, however, the information explodes: workers spend more and more effort on reading the records of other workers, and self-organization loses efficiency. In the statistical view, we represent the execution history of an MAS by two continually updated networks, a collaboration network among workers and a participation network between workers and task types. We then propose AgentCue, a method built on a new latent space model that places workers and task types in a shared latent space. At each checkpoint, that is, after a fixed number of sessions of the workers, the estimated parameters of the latent space model form a compact and interpretable statistical summary, and workers use the statistical summary to choose sub-tasks and collaborators without repeated searches of the shared board. We also provide theoretical results on model identifiability and estimation accuracy for continually updated networks. Experiments on three benchmarks for cross-document reasoning, mathematical theorem proving, and software patch generation show that AgentCue improves the efficiency, accuracy, and interpretability of self-organization.
Language models can produce similar probabilistic predictions after observing outcomes, reading summaries, receiving instructions, or undergoing activation interventions. Whether these routes induce equivalent belief states remains unclear: agreement in current predictions does not establish how a model will respond to subsequent evidence. We investigate this distinction in a Bernoulli inference task with an analytic Bayesian reference. Across three open-weight language models, we prepare states through four textual routes and two activation-level interventions, measure both next-outcome continuations and explicit probabilistic estimates, and append identical new evidence from each prepared state. We quantify the persistence of the initial belief through its effective evidence weight, the number of new observations it is worth. We find that the format of evidence decides which readouts express its content: outcome sequences primarily affect continuations, whereas summaries and stated rates primarily affect explicit estimates. Moreover, effective weights differ substantially across routes, with summaries receiving weights that scale weakly with their stated observation counts and elicitation questions changing inferred weights by orders of magnitude. In a one-shot steering setting, an intervention that reproduces a natural record's immediate prediction loses most of its effect under subsequent evidence. These results show that agreement in current readouts is insufficient to establish equivalence in belief updating: how a belief is acquired and how it is read jointly determine its subsequent dynamics.
Xiyan Su, Finn Rasmus Schäfer, Hojun Lim, Frank Diermeyer, Johannes Betz, Markus Lienkamp
World action models (WAMs) for autonomous driving couple the driving policy and the visual world model in a single generative network whose objectives are fixed post-training. Conditioning on a new driving objective requires retraining. We study energy guidance as a general mechanism and present GuidedWAM, a model-agnostic framework for injecting driving objectives into frozen WAMs at sampling time. Our framework constructs each objective as differentiable energies on the sampler's intermediate clean-action estimate, which needs only forward access to the frozen model, and is stabilized by sampling-time guidance scheduling, per-energy normalization, and velocity-bounded correction. On the NAVSIM v2 hard-split, guidance improves all three frozen, publicly released WAMs by 1.8 to 3.6 EPDMS, averaged over three sampler seeds, and adds less than 10% to the sampling time per step. These results position energy guidance as a practical post-training control channel for WAMs.
Duo Chai, Zizhen Liu, Shuhuai Wang, Songwei Pei, Cheng Liu, Huawei Li, Shangguang Wang
Large language models (LLMs) are highly compute- and memory-intensive, posing significant demands on high-performance GPUs. At the same time, advances in GPU technology driven by shrinking transistor sizes and lower operating voltages have made these devices increasingly susceptible to soft errors, particularly in radiation-prone environments such as satellite computing scenarios. While prior work has examined GPU soft errors, most studies have focused on general-purpose applications or conventional neural networks primarily used for vision tasks such as classification and detection. However, vulnerability analysis of modern LLMs remains limited, despite their rapid adoption in diverse application scenarios. Given the unique characteristics of LLMs, their resilience to soft errors may differ substantially from earlier models. To bridge this gap, we conduct the first instruction-level fault injection study of LLM inference to analyze the factors affecting LLM vulnerability. Our approach reveals reliability characteristics from multiple perspectives, highlighting the effects of model architecture, parameter scale, and task complexity. These findings provide new insights into LLM reliability and inform the design of more effective fault tolerance mechanisms.
Recent advances in Gaussian splatting have enabled efficient and photorealistic head avatar reconstruction, yet monocular video-based or image-based methods still suffer from under-constrained geometry and texture, leading to artifacts such as blurring, stretching, and temporal instability. We present EDGAR (Efficient Diffusion-Driven Gaussian Avatar Refinement), a conditional diffusion framework designed to enhance the quality of Gaussian avatars by correcting artifacts while preserving identity fidelity. To train EDGAR, we construct EDGAR-SET, a hybrid triplet dataset composed of real-captured and synthetic-rendered samples, which simulates realistic Gaussian avatar failure cases by pairing degraded renderings with ground-truth and reference images. EDGAR refines degraded frames through a fusion block that injects identity features extracted by the self-supervised Sapiens model into a one-step diffusion backbone, enabling effective artifact removal without sacrificing efficiency. Our EDGAR model incorporates a Mixture-of-Experts (MoE) architecture to address diverse degradation types and severity levels. The refined multi-view renderings are subsequently used as pseudo-targets to optimize a single underlying Gaussian avatar, so that the final avatar can be rendered without applying EDGAR to every output frame. Experiments on three different Gaussian avatar reconstruction methods, each evaluated under three different settings, show improvements in both perceptual and reconstruction measures over the raw outputs and an SDEdit-style refinement baseline. In addition, cross-view evaluation demonstrates improved calibrated consistency, while tests under large yaw and open-mouth expressions verify robustness under extreme settings. These results collectively indicate a useful operating regime for recoverable appearance artifacts. The code is available at https://kwanyun.github.io/edgar/
Rewriting Induction (RI) is a method to prove inductive theorems, originating from equational reasoning. By using Logically Constrained Simply-typed Term Rewriting Systems (LCSTRSs) as an intermediate language, rewriting induction becomes a tool for program verification, with inductive theorems taking the role of equivalence predicates. Soundness of RI depends on well-founded induction, and one of the core obstacles for obtaining a practically useful proof system is to find suitable well-founded orderings automatically. Using naive approaches, all induction hypotheses must be oriented within the well-founded ordering, which leads to very strong termination requirements. This, in turn, severely limits the proof capacity of RI. Here, we introduce Bounded RI: an adaption of RI for LCSTRSs where such termination requirements are minimized. Traditionally, RI can be used not only to prove equivalence, but also to establish ground confluence. Moreover, for ground confluent TRSs, RI can be extended to a system for disproving inductive theorems. We will show that this is also possible with Bounded RI.
Reliability management of industrial fleets depends on early fault warnings, yet lifecycle data across similar assets are highly imbalanced. A few assets accumulate run-to-failure histories, while newly installed or rarely failing assets provide only normal-condition data. This paper asks how diagnostic knowledge can be transferred from a data-rich source asset to a data-scarce target asset when no fault data have been observed on the target. We formulate cross-asset early fault detection as a zero-shot domain adaptation problem, in which zero-shot refers to the absence of target fault samples, and propose a framework that requires only normal data from the target. The framework comprises (1) a statistical process control step that labels the source asset run-to-failure data with an exponentially weighted moving average chart, reducing manual annotation to verifying flagged segments, (2) an Adapted Domain-Adversarial Neural Network (ADANN) that combines domain-adversarial training and a Maximum Mean Discrepancy penalty with an Adapted Classification-Domain (ACD) loss, which anchors the classifier normal region to the target own normal data, and (3) an ahead-of-time labeling scheme with a prediction horizon that targets incipient faults. Under an exact-classification assumption, a theoretical analysis characterizes the domain-adversarial term for a normal-only target and the alignment state in which source fault features remain separated from target normal features. On a public run-to-failure bearing dataset, ADANN attains the highest mean F1-score among four transfer baselines, with a nominally significant precision gain over the strongest baseline, and its robustness is examined under synthetic noise.
Dominant Resource Fairness (DRF) is an influential policy for sharing multiple resources among jobs with heterogeneous demands. It combines sharing incentive, strategy-proofness, envy-freeness, and Pareto efficiency, making it particularly attractive for shared computing systems. DRF has been implemented in major resource-management frameworks, including Apache Mesos and Hadoop YARN. DRF is also nonclairvoyant: it requires no advance knowledge of job sizes, making it practical when processing requirements are difficult to predict. This motivates understanding its completion-time performance relative to an optimal schedule that knows all job sizes. However, elementary arguments can only give an $Ω(\sqrt D)$ lower bound and an $O(D)$ upper bound for total completion time, where $D$ is the number of resource types. In this paper, we close this gap by proving that DRF is $Θ(\sqrt D)$-competitive. More generally, for every fixed real $q\geq1$, we show that DRF is $Θ(D^{1/(q+1)})$-competitive for the $\ell_q$ norm of completion times, with larger $q$ placing greater emphasis on late-finishing jobs. These bounds hold with arbitrary release times. We also give a refined analysis of total completion time for small $D$.
Esmeralda S. Whitammer, Sameera Lanka, Pranav Goel, Nebojsa Jojic
As neural language models approach human performance on NLP benchmark tasks, their advances are widely seen as evidence of an increasingly complex understanding of syntax. This view rests upon a hypothesis that has not yet been empirically tested: that word order encodes meaning essential to performing these tasks. We refute this hypothesis in many cases: in the GLUE suite and in various genres of English text, the words in a sentence or phrase can rarely be permuted to form a phrase carrying substantially different information. Our surprising result relies on inference by iterative shuffling (IBIS), a novel, efficient procedure that finds the ordering of a bag of words having the highest likelihood under a fixed language model. IBIS can use any black-box model without additional training and is superior to existing word ordering algorithms. Coalescing our findings, we discuss how shuffling inference procedures such as IBIS can benefit language modeling and constrained generation.
Shumao Wu, Kongcheng Zhang, Jianqiang Huang, Baisheng Lai
Agent trajectories often contain both successful decisions and failed attempts, yet optimization with sparse trajectory-level rewards struggles to identify which critical decisions should be reinforced or suppressed. We observe that meaningful distinctions among on-policy trajectories can be distilled into natural-language process rubrics and used as process supervision signals. Based on this insight, we propose Online Rubric Self-Distillation (ORSD), which leverages self-derived process rubrics for dense token-level supervision, thereby augmenting outcome-centric reinforcement learning. Specifically, the current policy independently generates multiple rubric bundles at each step according to the same group of trajectories, extracting candidate process rubrics from cross-trajectory behavioral differences. A frozen verifier then evaluates all rubric items against each trajectory in the group, and selects one candidate bundle based on bundle-level alignment between rubric satisfaction and task success. ORSD treats the selected rubric bundle as privileged information during training, then compares the log-likelihoods of the original tokens with and without the rubric. The resulting likelihood gap provides token-specific process credit, which is added to the original trajectory-level GRPO advantage through additive advantage shaping. ORSD thus bridges trajectory-level outcome supervision and fine-grained process supervision, enabling token-level credit assignment using process signals. Extensive experiments on ALFWorld and WebShop show that ORSD accelerates policy improvement and boosts final task performance for both Qwen3-1.7B and Qwen3-4B under the same environment-interaction budget. Our code is available at https://github.com/AnnaXChun/ORSD.
Haoyu Han, Jack Benarroch Jedlicki, Jiahui Lei, Heng Yang
Accurate 4D reconstruction remains challenging: jointly estimating camera motion, scene geometry, and non-rigid deformation entails highly nonconvex optimization, making global optimality difficult to guarantee. We present XM4D, a framework for accurate and efficient 4D reconstruction built upon certifiably globally optimal geometric optimization. By separating the estimation of camera motion from non-rigid deformation, we formulate both stages as structured quadratically constrained quadratic programs (QCQP) that admit semidefinite programming (SDP) relaxations and verifiable global optimality certificates. For camera motion estimation, we lift static point tracks into 3D using foundational depth prediction models and formulate a scaled bundle adjustment problem that jointly estimates camera poses, depth scales, and static landmarks. By analytically eliminating translations and landmarks, we reduce the problem to a QCQP over scaled rotations. For non-rigid reconstruction, we represent dynamic motion using an embedded deformation graph (EDG) and show that its tracking, rigidity, and temporal regularization objectives similarly reduce to a rotation-only QCQP. Both problems admit SDP relaxations that we solve efficiently with a GPU-accelerated Burer-Monteiro solver based on Riemannian optimization. The recovered geometry further initializes dynamic Gaussian splatting for photometric refinement. Extensive experiments across seven benchmarks demonstrate accurate camera motion estimation on both static and dynamic scenes, efficient optimization that scales to thousands of frames, and state-of-the-art novel-view rendering quality.
Dimensionality reduction has proven powerful for identifying neural manifolds, which are low-dimensional structures underlying high-dimensional neural activity. These low-dimensional representations have improved the interpretability of population-level coding. Yet whether such low-dimensional representations are biologically relevant and confer functional advantages in learning systems, or merely reflect neuron-level activity, remains contested in neuroscience. We show that an explicit information bottleneck forcing a recurrent neural network to learn a low-dimensional representation is necessary for rotational and out-of-distribution generalisation in a time-series prediction task. Using information-theoretic measures of causal emergence, we characterise the dynamics of this representation across the memorisation-to-generalisation transition, finding a non-monotonic trajectory which shows an initial decrease, a minimum, and a subsequent rise to a maximum, even as prediction loss falls monotonically. This trajectory scales with task complexity, and the magnitude of emergent structure reliably predicts generalisation performance. Analysis of hippocampal CA1 and medial prefrontal activity in rats learning an alternating maze task reveals analogous non-monotonic emergence dynamics that track behavioural performance. Together, these findings indicate that the ability of neural networks to learn compact, distributed and emergent representations confers a functional advantage for generalisation, supporting a causal role for learned representations in cognition.
Knowledge of the hidden interior structure of an object, including the presence and location of cavities, is essential for predictive simulation, robotic manipulation, and nondestructive inspection. However, such cavities leave no direct signature on the static visible surface, making collision-induced deformation one of the few cues used to infer their presence. Structure from Collision (SfC) addresses this problem by recovering the hidden interior structure of an object from multiview observations of a collision event. SfC is challenging for two reasons: hidden cavities are only indirectly observable through surface motion, and voids correspond to the absence of material, making direct optimization using standard density representations difficult. The density-based SfC baseline inherits both difficulties. Surface-matching gradients do not directly indicate the cavity location, while density gradients can be weak or uninformative inside or near void regions. Therefore, we reformulate SfC as hypothesis-space reasoning, representing hidden cavities as explicit structural hypotheses whose variables encode available structural assumptions rather than relying on implicit density updates. A global stage performs gradient-free optimization over this hypothesis space using forward-only material point method (MPM) simulation. A local stage then refines the selected hypothesis as a Lagrangian particle system through gradient-based optimization using differentiable MPM simulation. On the SfC benchmark, our method improves hidden-structure reconstruction and future prediction over the density-based SfC baseline. Interior-topology stress tests and robustness analyses characterize its scope and limitations. The project page is available at https://www.kecl.ntt.co.jp/people/kaneko.takuhiro/projects/sfc-hypothesis/.
Conversational agents can retrieve information directly or delegate retrieval to a subagent. This design choice affects how retrieval history is maintained, what context each model call receives, and what information must be re-retrieved in follow-up questions. We compare three designs: skill-based retrieval in the main agent, a non-persistent retrieval subagent that generates a findings response, and our persistent retrieval subagent, which maintains retrieval history across turns and returns retrieved evidence directly to the main agent. Our approach also allows the main agent to request corrections from the same subagent. We evaluate the three designs on enterprise questions posed independently, within shared conversations, and as dependent follow-ups. Across five multi-turn settings, our approach uses 43.3% fewer recorded agent-model tokens and incurs 45.0% lower estimated API cost than skill-based retrieval, while reducing average response time by 9.1%. Compared with non-persistent retrieval, it reduces response time by 15.6 seconds per question, while using 65.1% more tokens and incurring 29.1% higher estimated API cost. Our approach shifts the retrieval subagent from producing synthesized responses to maintaining retrieval state and supplying evidence that the main agent can directly evaluate and use for answer generation.
LLM agents are expected to act over multiple turns, using search, browsing interfaces, and tools to complete user goals. Yet not every goal is well specified or achievable in an available environment. In such cases, a reliable agent should recognize that further interaction is unlikely to help and abstain from additional tool calls. We define Agentic Abstention, the problem of deciding when an agent should stop acting under uncertainty. Unlike standard LLM abstention, which is usually evaluated as a single-turn answer-or-abstain decision, agentic abstention is a sequential decision problem: an agent can answer, abstain, or gather more information at each turn, and the need to abstain may only become clear after interacting with the environment. We study this problem across web shopping, terminal environments, and question answering, evaluating 13 LLM-as-agent systems and 2 agent scaffolds on more than 28,000 task instructions. Our results show that the main challenge is not only whether agents can abstain, but also when they abstain. Some agents never abstain when they should, while others do so only after many unnecessary interactions. This gap is especially large on tasks where the instruction appears feasible until the environment reveals otherwise (e.g., no valid result matches the instruction). We further find that model scale, reasoning, and agent scaffolding affect abstention in different ways, where larger or more capable models sometimes perform worse at timely abstention. Finally, we introduce CONVOLVE, a context engineering method for improving agentic abstention that distills interaction trajectories into reusable stopping rules. On held-out WebShop tasks, CONVOLVE substantially improves timely abstention without updating model parameters, raising Llama-3.3-70B's timely recall rate from 26.7 to 57.4. Our dataset and code are available at https://agentic-abstention.github.io.
Jincheng Li, Yutian Zhang, Jianyu Zhang, Xinshun Wang, Mengyuan Liu
Affordance grounding asks which visible part of an object performs a requested operation, a basic capability for embodied manipulation and human--robot interaction. Most existing methods follow a System Two paradigm: a multimodal language model first writes a chain of thought and then turns its conclusion into a box or points for a segmenter. This incurs considerable inference cost, gives the semantic decision little direct control over the extent of the final region, and makes mistakes hard to correct. Yet locating an affordance is essentially a perceptual judgment that need not be argued out in words. We therefore propose, in which frozen segmenters propose candidate regions and a System One model decides directly which of them perform the operation, generating no text and reading each decision from a single forward pass. Our contributions are twofold. (1) An inference framework that confirms candidate regions level by level, from object to part to sub-part, with evidence passed between the levels in both directions so that the levels correct one another. (2) A System One model trained for this framework: existing System One models are built for general-purpose decisions, judge one candidate per request and are ill-suited to the framework's region questions, so ours learns directly from the requests the framework itself produces and judges all candidates of a level at once. Trained only on InstructPart and evaluated with definition-aligned operation words, outperforms chain-of-thought methods in domain and in zero-shot transfer to UMD and IIT-AFF, is better overall than existing System One models run in the same framework, and needs fewer judge requests. To our knowledge, is the first approach to bring System One models to affordance grounding.
Jaehyuk Lee, Hanyoung Kim, Yanggee Kim, Donghun Lee
Vision Transformers (ViTs) face severe computational bottlenecks due to the quadratic complexity of self-attention at high resolutions. Existing token reduction methods rely on local metrics -- such as single-layer attention scores -- that are inherently vulnerable to the attention sink phenomenon, where uninformative tokens are paradoxically preserved over salient foreground objects. We propose ASAP (Attention Sink Anchored Pruning), a training-free framework that recasts this sink as a feature. Modeling ViT information flow as a Lazy Random Walk, ASAP identifies the sink as a dominant accumulator of probability mass. By computing the diffusion distance to the sink within the cumulative transition matrix, ASAP partitions tokens via Radial Diffusion Clustering and compresses background redundancy through Transition Weight Pooling in a single shot. Extensive experiments across image, video, and vision-language tasks demonstrate ASAP outperforms state-of-the-art methods, accelerating throughput by up to 48% while maintaining or even exceeding baseline accuracy.