Research agents achieve increasingly strong benchmark scores, yet these scores do not necessarily translate into reliable delivery in ongoing research projects. Many research-agent benchmarks emphasize standalone tasks and relatively self-contained workflows in clean or curated workspaces. However, researchers ask agents to complete one step within an ongoing project, where requests depend on earlier session decisions and workspaces contain unfinished files, outdated versions, and failed attempts. We introduce ARE-bench to evaluate whether research agents advance ongoing projects while respecting prior decisions and existing workspace state, grounded in real session context and in-progress workspaces. ARE-bench includes 75 tasks built on agent sessions and full workspaces from recruited researchers' own projects, spanning seven research stages, 33 requirements, and 150 sub-requirements, with 1,173 binary checklist items in total. Each task is anchored at a real user request, preserves its requirements, and summarizes the earlier requests still in effect as background. The complete workspace is replayed to that boundary without cleanup, preserving artifacts from earlier turns. Each checklist is compiled from an intent-aligned rubric and credits any valid solution. Across seven frontier models, the strongest passes every checklist item on only 66.2% of tasks, and Design and Execution rank among the hardest stages for every model. Three models pass about 69% of tasks when only delivery contracts and general norms are checked, but just 28.4% to 40.0% once historical and material or workspace requirements are also checked. Evaluating delivery without these contextual requirements can therefore overestimate research agents and mask differences between them.
Masked Diffusion Models (MDMs) are well known for fast text generation via parallel decoding. However, independent parallel sampling can commit to inconsistent predictions, and less is known about how to selectively revise these early token commitments. In this work, we introduce $\underline{\mathrm{P}}$lanned $\underline{\mathrm{U}}$nmasking-and-$\underline{\mathrm{R}}$evision Joint $\underline{\mathrm{L}}$earning (PURL) for MDMs, a training-based framework that seamlessly unifies parallel generation with selective token revision. We derive a normalized joint transition whose likelihood factorizes into edit selection, unmask selection, and token prediction. These components are trained jointly on partially masked sequences constructed according to the model's own unmasking policy. During inference, all three decisions are supported by a single shared forward pass, enabling the model to decode masked tokens and revise noisy tokens. Across logic puzzles, mathematical reasoning, and code generation, PURL improves accuracy at matched unmask budgets, with relative gains up to 47.4% accuracy improvement and up to an $8\times$ reduction in scheduled decoding rounds. Beyond these gains, our analysis further reveals general principles of revisable decoding: token edits become more effective as context accumulates; early edits can steer unfinished trajectories, whereas later edits tend to remain localized; and effective edits therefore depend on both calibrating when to intervene and learning where and how to revise. Together, these suggest that effective parallel generation depends not on revising more tokens, but on making the right commitments revisable at the right time.
As electric vehicle (EV) adoption accelerates globally, securing electric vehicle charging station (EVCS) software becomes increasingly critical. The Open Charge Point Protocol (OCPP), widely adopted for EVCS management, places significant responsibility on developers to securely implement specifications that are subject to multiple interpretations. The variety of OCPP implementations across different programming languages provides a unique opportunity to evaluate security tooling in a real-world, multi-language context where implementation variability may introduce vulnerabilities. This study provides the first empirical evaluation of Static Application Security Testing (SAST) tool performance across multiple programming languages using OCPP implementations as a real-world test corpus. We applied both comprehensive multi-language tools (SonarQube and Semgrep) and language-specific analyzers to twelve OCPP 1.6 implementations--each written in a distinct programming language--to examine how detection patterns vary across language ecosystems. Total issues ranged from just two in Rust to 1,689 in Java. Mature programming languages like Java yielded substantially more findings than newer languages like Rust, likely because established languages have more developed SAST tool ecosystems rather than inherently worse code quality. These results suggest that developers should consider available SAST tooling when selecting implementation languages for security critical systems.
The nonnegative rank of a nonnegative matrix $X$ is the smallest number of nonnegative rank-one factors that sum to $X$. Since computing the nonnegative rank is NP-hard, it is common to circumvent this issue by computing lower and upper bounds. In this paper, we propose non-convex formulations and practical implementations for four important lower bounds for the nonnegative rank, namely the fooling set bound (FSB), the rectangle covering bound (RCB), the hyperplane separation bound (HSB), and the self-scaled bound (SSB). In particular, to the best of our knowledge, our algorithm for computing the SSB is the first available in the literature. We compute the SSB for several matrices for which its value had not previously been reported. In some cases, the SSB provides the best known upper bound, while the other three bounds do not. Moreover, on canonical benchmarks, we show that our non-convex approaches are often competitive with existing methods. The paper also provides a consolidated reference for the current state of several classical lower bounds on a large number of benchmark matrices.
Understanding the persistence of learned behaviors in AI safety motivates an investigation of the dynamical counterpart to the equilibrium geometry studied in singular learning theory. We analyze this persistence through transitions between singular wells. We establish an Eyring-Kramers formula for a low temperature Langevin diffusion between two compact connected zero components of a real-analytic potential. The wells may be singular and nonisolated, and the unique minimizing gate is a nondegenerate index-one saddle, with gates understood as inclusion-minimal sets meeting every minimizing path. A boundary-gate construction shows that uniqueness of this inclusion-minimal gate is equivalent to every minimizing path passing through the saddle and one such path meeting the communication level only there. Compact semianalytic topology provides a separating function away from a fixed saddle neighborhood; analytic Morse coordinates give matching capacity bounds by first sending the temperature to zero and then shrinking that neighborhood. An anchored energy estimate and a small-ball potential-theoretic identity give uniformity for deterministic starting points. The leading constant is expressed in terms of the sublevel-volume coefficient, a Gamma factor, and the saddle Hessian.
Time series foundation models (TSFMs) forecast unseen series without task-specific training. Increasingly they are built on recurrent backbones, whose constant per-step cost makes long-context inference cheap, and which, unlike transformers, carry a state. A recurrent backbone computes its prediction from two inputs: the observed context window and the initial recurrent state. In practice only the first is utilised and the initial state is unused by convention. We ask whether the initial state can be utilised in recurrent TSFMs. Using the xLSTM-based zero-shot forecaster TiRex as a case study, we freeze the backbone and train only a small head that writes a context vector into the initial state, so that any change in accuracy is attributable to the injected state rather than to adapted weights. A spectral summary of the input window, learned once and applied unchanged to every series, lowers mean GiftEval MASE by 0.175%, with every seed improving, which neither a parameter-matched LoRA adapter nor a learnable static state reproduces. An embedding of text attached to each series, evaluated on six MoTime datasets, reveals two roles for the state. As a per-dataset adaptation surface, a static state lowers MASE by 0.04-6.96% of the zero-shot baseline (8-70% of the gain of full fine-tuning). As a channel for exogenous information, the text-conditioned state lowers it by a further 0.15-0.41% of the zero-shot baseline on four of six datasets and, combined with a LoRA adapter, exceeds full fine-tuning, which has no access to the text, on one of them. In TiRex, the effect concentrates in the mLSTM blocks and is strongest at short context. The results are preliminary and the improvements small, but they point to the initial state as a direction for further research.
Rapidly evolving LLM architectures and heterogeneous mobile hardware make supporting new models in on-device inference engines a recurring and costly system challenge. Each model may require coordinated changes across conversion, graph construction, operators, and hardware backends. Existing coding agents often miss these cross-layer dependencies and struggle to diagnose failures across host--device execution paths. We present EngineAgent, an agentic system for automating on-device LLM model enablement. Its key insight is to separate reusable engine and hardware knowledge from model-specific implementation. Before enablement, EngineAgent constructs a hierarchy-aware engine index that captures engine abstraction layers and cross-layer dependencies, and pre-validates backend-specific build, execution, and debugging workflows using already supported models. For each new model, a builder uses this knowledge to coordinate cross-layer changes, while an auditor checks implementation completeness and drives iterative debugging through device-side validation. Across 12 tasks spanning three mobile LLMs and four engine--backend pairs on CPUs, GPUs, and NPUs, EngineAgent achieves 75% end-to-end success with either of two underlying LLMs, versus 33.3% for the strongest coding-agent baseline. Success requires compilation, on-device execution, and output-logit agreement with reference implementations. EngineAgent also reduces average per-task time and model-API cost relative to baselines using the same underlying LLMs.
Retrieval-augmented systems often retrieve passages that are accurate and relevant but state the answer at a coarser level of granularity than the question requires. For example, a question may ask which specific disease a drug treats (familial hyperlipidemia), while the retrieved passage mentions the broader disease family (hyperlipidemia). Prior work on context sufficiency treats such coarse passages simply as insufficient. We argue that they form a distinct case, which we call resolvable insufficiency: text alone cannot settle the fine-grained answer, but the broader category it mentions maps to a short list of candidate answers in a domain ontology. When language models read these coarse, family-only passages, they exhibit a phenomenon we call Clade (Category-Level Answering from Deficient Evidence): they answer instead of abstaining or searching further. On biomedical questions from PrimeKG, family-only passages reduce abstention by 13.8 points compared to uninformative passages, and the resulting guesses are wrong about half the time. A counterfactual test confirms the cause: replacing an exact disease name with its broader family name lowers abstention by 17.4 points compared to deleting the sentence. Simple fixes like confidence filtering, scaling to a larger model, and explicitly instructing the model to abstain cannot resolve Clade, and this replicates on ecological species interactions in GloBI. Finally, we test whether an ontology can turn coarse passages into exact answers. Given the correct family, an ontology list includes the true answer in 98.5% of cases, but string matching finds that family on the page for only 19.8%. Letting the model choose the family raises this to 30.0% and end-to-end recovery from 10.0% to 26.8%. Overall, our results show when language models should stop guessing and consult an ontology, and highlight which steps of that process remain unsolved.
Traditional recommendation systems rely on latent (dense) representations, making them difficult to interpret and control. We propose the Controllable and Content-Based Recommendations (CCBR) framework, which builds its recommendations from textual user profile representations. CCBR plugs into collaborative filtering models and introduces controllability via text bottlenecks. We show that CCBR enables text-based and multimodal interventions, allowing users to steer the model towards the directions they prefer. Different from existing controllable recommendation systems, CCBR infers the text summaries directly from item contents (images, audio or video). Across image, audio, and video datasets, we demonstrate that the proposed framework obtains competitive model performance with standard (latent-representation) models while providing controllable model summaries via text. The model also outperforms TEARS, a recent baseline for controllable recommendation systems. Through systematic interventions, we demonstrate the efficacy of the steering mechanism.
Probabilistic linear solvers (PLSs) return probability distributions that quantify uncertainty due to limited computation in the solution of linear systems. The literature has traditionally distinguished between Bayesian PLSs, which condition a prior on information obtained from projections of the linear system, and probabilistic iterative methods (PIMs), which lift classical iterative solvers to probability space. In this work we show this dichotomy to be false: Bayesian PLSs are a special case of non-stationary affine PIMs. In addition, we prove that any realistic affine PIM is calibrated. These results motivate a focus on (non-stationary) affine PIMs, but their practical adoption has been limited by the significant manual effort required to implement them. To address this, we introduce affine tracing, an algorithmic framework that automatically constructs a PIM from a standard implementation of an affine iterative method by passing symbolic tracers through the computation to build an affine computational graph. We show how this graph can be transformed to compute posterior covariances, and how equality saturation can be used to perform algebraic simplifications required for computation under specific prior choices. We demonstrate the framework by automatically generating a probabilistic multigrid solver and evaluate its performance in the context of Gaussian process approximation.
Given an unweighted directed graph, a sparsifier is a nonnegative reweighting of original arcs that preserves every in-degree and out-degree exactly. In this work, we study singular-value sparsifiers, which are generalizations of spectral sparsifiers. Singular-value sparsifiers find a number of applications in theoretical computer science, such as maintaining directed cuts and flows, speeding up directed Laplacian solvers, and sparsifying dense weight matrices in neural networks. Strong singular-value (SV) approximation supports sparse representations of directed random walks with guarantees that remain stable under products, and controls bilinear error in the left and right Schur-complement metrics. We prove that every unweighted directed graph on $n$ vertices admits an $\varepsilon$-SV sparsifier with $O(n\log(2n)/\varepsilon^2)$ arcs, improving the $O(n\log^5 n/\varepsilon^2)$ support bound of Lau, Wang, and Zhou. The sparsifier is a nonnegative reweighting of original arcs and preserves every in-degree and out-degree exactly. We also give a randomized construction with $\tilde{O}(m+n)$ real-arithmetic running time, where $m$ is the input support size, and the output has the same support bound and rational weights of polynomial bit length.
Gradient-boosted decision trees are among the strongest off-the-shelf predictors for tabular regression, but point predictions alone do not quantify uncertainty. Conformal prediction provides distribution-free marginal coverage, yet standard split conformal uses a single global residual quantile and can adapt poorly to heteroscedasticity. We propose LoBoost, a model-native local conformal method that reuses the fitted ensemble's leaf structure to define a multiscale partition of the feature space. Each input is represented by the sequence of leaves it visits along the boosting path, and matching leaf prefixes define nested groups in which residual quantiles are estimated locally. By reusing the predictive structure already learned by the model, LoBoost requires no auxiliary partition or nuisance model, no retraining, and no additional data split beyond standard conformal calibration. Our theory connects the stability of the fitted ensemble and the geometry of its induced cells to local residual-score homogeneity, providing finite-sample coverage-error control and asymptotic pointwise validity as the local calibration size grows and the cells become sufficiently homogeneous. Experiments show a transparent quality-cost trade-off: LoBoost improves over global split conformal on several heterogeneous datasets, but it is not uniformly superior to the adaptive baselines and is not uniformly Pareto-optimal.
Diffusible signals enable communication between immune cells but can also create interference between nearby activation sites. In lymph nodes, effector T cells (Teff) concentrate around antigen-presenting cells (APCs), release interleukin-2 (IL-2) following antigen recognition, and sense IL-2 through cell-surface receptors. Experiments show high-uptake regulatory T cells (Treg) clustered around IL-2-producing cells. However, it is unclear how strongly responses at neighboring activation sites are coupled through diffusible IL-2, and how this coupling is influenced by the spatial organization of Treg acting as IL-2 sinks. We interpret two neighboring sites as a molecular communication system and study the effect of clustering using a three-dimensional IL-2 field coupled to T cell population dynamics. Across a broad parameter sweep, clustering Treg produces a larger mean reduction in neighboring-site interference than adding dispersed Treg. Compared with no Treg, clustered Treg suppress interference while largely preserving the strong local response with corresponding consequences for population expansion and survival. Information-theoretic analysis shows improved discrimination between weak and strong local sources at short separation, whereas at larger separation clustering primarily reduces the remaining estimated neighboring-source information. These findings support a spatial-filtering role for Treg clustering in maintaining locally informative immune responses.
Users often complain that AI assistants such as ChatGPT are getting worse. Some versions do change, but human evaluation leaves a second possibility untested: the people judging may change. We separate the two on Chatbot Arena, a public platform where people compare answers from two anonymous large language models (LLMs) and pick the better one or say both are bad. We study 1,260,967 blind votes cast from May 2023 to August 2024 on pairs of models whose versions are fixed: dated vendor snapshots and models with fixed open weights. For the same pair of unchanged models, the share of votes saying "both are bad" rises by 1.71 percentage points a month on a mean of 17.1%, and by 1.91 points a month in 2024. In fourteen pairs observed often in both 2023 and 2024, it rose in every one, from 15.5% to 23.3%. The ranking does not trend: in 2024, the stronger model's share of decisive votes is flat. The rise holds within the same voter and on the same pair, and across arrival cohorts from their first recorded vote. It holds with prompt controls, in every major language, for dated vendor snapshots, open-weight models and models with stable answer length, and between interface changes. The drift is shared across models: month effects estimated from other models' pairs remove it from each model's own pairs. Platform-wide, the average model would be expected to score 78% against the 2023 average, yet the share of "both are bad" votes ended near where it began. The pattern fits a standard that follows the category's progress, so a product that stands still is judged as declining. Absolute verdicts on AI drifted while relative comparisons held steady, an evaluation drift that matters for LLM leaderboards and for the human feedback used to train models.
Stochastic diffusion samplers often require substantially more neural function evaluations (NFEs) than deterministic samplers to achieve comparable sample quality. Although higher-order discretization accelerates deterministic sampling, its benefits for stochastic sampling remains unclear. We establish matching upper and worst-case lower bounds showing that stochastic samplers with higher-order approximations remain first-order convergent, even with exact score functions. Thus, increasing the approximation order does not generally improve the convergence order. Motivated by this limitation, we introduce forward-value discretization, a training-free framework that evaluates the data/noise prediction model at a lookahead estimate constructed from previous predictions. For deterministic sampling, the resulting Forward DPMSolver (F-DPMSolver), improves sample quality at low NFE budgets through its DDIM-based variant, despite remaining first-order convergent. For stochastic sampling, we develop Forward DDPM (F-DDPM), which evaluates the model at a cleaner lookahead point adjusted for subsequent noise injection. This estimate reuses both the previous model prediction and the previously injected noise. Both variants admit first-order convergence guarantees with one model evaluation per iteration. Across CIFAR-10, ImageNet, FFHQ, and LSUN, F-DDPM consistently achieves lower FID than the evaluated stochastic baselines at matched NFE budgets. With 1.5 times as many NFEs as deterministic baselines, it achieves competitive sample quality, narrowing the practical efficiency gap between stochastic and deterministic sampling. Code is available at https://github.com/Na-Li66/F-DDPM.
RNA sequencing produces rich and diverse datasets of gene expression, offering compelling insights into cellular state and function that have many applications in drug discovery. Modeling such data is challenging due to inherent technical noise and experimental batch effects, as evidenced by many existing transcriptomic foundation models (FMs) underperforming relative to linear baselines. Such results raise the question of whether deep representation learning provides a distinct advantage over the direct use of raw transcript counts. Our work explores this by developing a new self-supervised model, TxFM, with a focus on inductive representation learning evaluations. TxFM employs a masked autoencoding approach tailored to diverse RNA-seq count data, and our ablation study empirically identifies crucial architecture configurations required for strong transfer performance. Additionally, we curate a public training corpus, DiverseRNA-1.4M, and find that TxFM trained on this curated dataset yields high-fidelity gene representations that outperform FMs trained on atlas-scale corpora over 100x larger. Overall, our results indicate that inductive self-supervised learning is a viable modeling approach for transcriptomics representation, provided a careful synthesis of model architecture and training data curation.
LLM agents increasingly operate in long-horizon, verifier-rich environments, but current self-improvement mechanisms often write episode-derived lessons directly into memory, making local success an unreliable signal for reusable and safe knowledge. This paper addresses the core question of when an experience-derived behavioral update should be admitted into persistent agent state. We propose Falsification-Gated Self-Evolution (FGSE), a verifier-grounded framework that represents each candidate lesson as a structured hypothesis with an explicit precondition, behavior change, expected effect, and verifier. FGSE applies the hypothesis only in a temporary state, tests it on target-transfer, falsification, and archived-regression probes, and then commits, refines, or rejects it according to measured gain and risk. Across web, app, tool-use, and long-term memory benchmarks, FGSE achieves strong task performance, including 39.8 WebArena success rate, 78.8 AppWorld average completion, and 0.724/0.487 Pass$^1$ on $τ$-Retail/$τ$-Airline, while reducing falsification failures to 5.0%, regression damage to 1.3%, and harmful committed updates to 3.6%. These results suggest that testing memory updates as hypotheses supports more reliable long-term adaptation.
Unsupervised video object-centric learning aims to decompose dynamic scenes into persistent, object-level representations without supervision. However, existing slot-based methods struggle to maintain stable object identity in challenging settings such as rapid motion and partial occlusion. First, they typically encode both the per-frame appearance of an object and its identity across frames in a single slot vector, creating an objective conflict that leads to slot swapping: reconstruction requires sensitivity to transient visual changes, whereas temporal consistency requires invariance to them. Second, the token renormalization used in Slot Attention can amplify weakly attending slots, allowing them to absorb tokens from other objects and destabilize slot-to-object correspondence. We propose Dual-State Slot Attention (DSSA), a fully self-supervised framework that addresses these limitations by separating appearance from identity and by reducing spurious updates from weakly matching slots. DSSA decomposes each slot into a local state for per-frame appearance and an identity state for temporally stable object information, thereby aligning reconstruction and temporal consistency with separate representations. The identity state is updated through a learned recurrent transition that acts as a temporal filter on the local state, while competition-modulated aggregation (CMA) down-weights updates from weakly matching slots and prevents them from absorbing tokens from other objects. Experiments on MOVi-C, MOVi-D, and YouTube-VIS demonstrate that DSSA consistently improves segmentation quality and temporal consistency over prior methods, while also yielding stronger downstream object recognition and video dynamics prediction. Code is available at https://uark-aicv.github.io/dssa/.
Language-model systems use self-judgment to decide which answers to release, review, or recompute. Because intervention is costly, an assurance score must rank errors well at the available budget, not merely over the full ranking. We compare the judge's continuous verdict-token score with a supervised correctness probe trained on attention-head activations from the same prompt. Exact identities link fixed-budget error recall to residual risk and show that AUROC averages recall over all budgets. Across three open 7-12B models on GSM8K and MATH, the probe finds 1.3-22.5 additional percentage points of errors at 5% review in all six model-dataset settings (median 7.4 points). At 10% routing, probe-guided allocation improves accuracy over equal-cost random allocation in both GSM8K experiments, but beats judge-guided routing only for the weaker judge. Cross-dataset transfer lowers probe AUROC for every model in both directions relative to target-domain training. Because dataset, weight precision, and generation budget change together, this measures a joint protocol shift rather than a dataset-only effect. Internal signals can improve supervised, in-domain budget allocation, but are not domain-general confidence estimators.
External memory enables LLM agents to improve from experience by distilling trajectories of past tasks into concise lessons for new tasks. However, memory can also lead to memory-induced degradation, where an agent fails on tasks it would solve without memory, often because it follows a lesson that is incorrect, or misapplies or misreads a valid one. This is because lessons are inferred from trajectories labeled only with task success or failure, yet are typically provided without the trajectories they came from, leaving the agent no way to check them. To address this, we introduce WikiTrace, an evidence-linked memory system designed to let agents check each lesson against its original trajectory. Specifically, motivated by the finding that a trajectory often still holds the evidence needed to catch an erroneous lesson, WikiTrace stores each lesson as a concise insight linked to a summary of why its trajectory succeeded or failed and to the raw trajectory itself. Then, on a new task, the agent starts from all insights and, whenever one lacks details needed to act, conflicts with another, or comes from a failed task, descends through its links with terminal tools to fill in missing details or correct the insight before acting. We validate WikiTrace on five agentic benchmarks, showing that it consistently outperforms existing memory approaches and a baseline given the same insights alone, while also reducing memory-induced degradation. Our code is available at https://github.com/yukyeongleee/wikitrace
This paper determines the accepting-state spectrum of the right-quotient operation for permutation automata, solving an open problem posed by Rauch and Holzer. They showed that in the unary case in which the languages whose right quotient we consider are languages over an alphabet with but one symbol, the accepting state complexities which could ensue for automata recognizing their quotient is exactly all integers between $1$ and $mn$ inclusive when $m,n \ge 1$; we show that for arbitrary alphabets the accepting-state spectrum is $g^{\operatorname{asc}}_{-1,\mathrm{PFA}}(m,n)=\{0\}$ if $m=0$ or $n=0$, and $g^{\operatorname{asc}}_{-1,\mathrm{PFA}}(m,n)=\mathbb{N}_{>0}$ if $m,n\ge 1$. Our result shows that every positive accepting-state complexity ensues for right quotient as long as both of the languages of which we consider the right quotient are nonempty. $0$ becomes the only unavoidable magic value. Moreover, we show that an alphabet with just two letters can witness any arbitrary $α\ge 1$ when $m,n \ge 1$, in that languages can be defined over a binary alphabet with accepting-state complexity $m$ and $n$ respectively for which their right quotient has accepting-state complexity $α$. What ultimately happens in our proof is that we construct $(A^{\mathrm q}_{m,α},B^{\mathrm q}_{n,α})$ such that the words accepted by $B^{\mathrm q}_{n,α}$ induce a point stabilizer in a symmetric group, and the standard quotient construction giving an automaton recognizing the quotient language then saturates the original final set of $A^{\mathrm q}_{m,α}$ to a full orbit; this gives a minimal quotient automaton with exactly $α$ final states. The complete spectrum which is our result follows from this, solving the open question of what the right quotient's accepting-state complexity for permutation automata is.
Concept vectors aim to enhance model interpretability by linking learned internal representations with human-understandable semantics, but their practical utility is often limited by noisy and inconsistent activations. In this work, we discover the SuperActivator Mechanism: a transformer dynamic that amplifies concept activation gaps, concentrating the most reliable concept evidence into a small set of highly activated tokens, which we call SuperActivators. We theoretically characterize this mechanism, proving that concept-aligned attention produces two competing effects: a tail amplification effect, where gaps among already-high activations grow disproportionately fast, and an equalization effect that eventually dominates as attention concentrates on extreme tokens. We validate the predicted tail amplification empirically, finding that in-concept activation distributions develop well-separated upper tails consistently across models, image and text modalities, and concept types. Moreover, we demonstrate the practical implications of SuperActivators for concept detection and localization. First, using only high-activation tokens as concept indicators improves detection by up to 14 F1 points over competing baselines. Second, restricting detection to the extreme activation tail is not merely sufficient but essential: including activations beyond the top 5-10% of the in-concept distribution introduces noise that degrades performance. Finally, using SuperActivators as instance-specific attribution targets yields more accurate and faithful concept localization heatmaps than using global concept vectors.
The escalating complexity of deep neural networks hinders the deployment of federated learning (FL) on resource-constrained devices, especially for multimedia applications such as on-device image classification and speech recognition. Parallel split learning (PSL) has emerged as a promising solution by offloading a substantial portion of the computational workload to the server through layer-wise model partitioning, making it particularly suitable for resource-limited visual and speech intelligence scenarios. However, the absence of client-side aggregation renders PSL sus ceptible to severe training divergence and potential convergence failure under heterogeneous data distributions. To address this challenge, we propose PSL-RGHA, a gradient-aligned PSL framework that employs reference gradient half-space alignment mechanism to mitigate training divergence among client devices. PSL-RGHA comprises two novel components: reference gradient progressive construction (RGPC) and gradient half-space alignment (GHA). Specifically, RGPC evaluates the directional consistency of client gradients and progressively synthesizes a robust reference gradient via Weiszfeld updates, effectively approximating the global descent direction. Then, GHA constrains client gradients within the reference half-space defined by RGPC to mitigate gradient conflicts and rectify local updates. Extensive experiments demonstrate that our PSL-RGHA framework outperforms existing baselines in various settings.
Adding abstention to a task-adapted multimodal model can discard correct answers, including capabilities acquired during adaptation. We formulate abstention repair as maximizing correct abstention on unanswerable questions under a budget on newly lost correct answers. We propose ReFrain (Regression-Free Abstention), which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with the displayed "none of the above" option while preserving all other responses. "Regression-free" denotes zero loss on the calibration set at the zero-budget target. All abstention methods in our main comparison start from the same answering-SFT checkpoint; we separately track lost correct answers, repaired errors, and retained SFT gains. On Qwen2.5-VL-3B, ReFrain at zero budget retains all 57 SFT-gained answers with no losses and improves correct abstention by 7.30 percentage points. At a 0.5% calibration budget, it reaches 74.05% correct abstention with 8 losses and retains 55 gains; CLIP-UP EmbLoRA reaches 64.65% with 62 losses, 41 repairs, and 40 retained gains. The same probe at its default threshold loses 298 correct answers. With models and thresholds frozen, confirmation on 228 independent question clusters yields 66.05% correct abstention and no losses at the 0.5% point, exceeding EmbLoRA by 32.72 percentage points. Ordinary refusal SFT achieves higher correct abstention at greater regression cost. Code and reproduction materials: https://github.com/ModeLiberaL/ReFrain.
Many open mixture-of-experts (MoE) language models train with Muon instead of AdamW. Most balance the load with a bias on each expert's routing score, adjusted after every step. Their recipes disagree on whether Muon should also train the router. In models with 55M and 108M active parameters, we measure how the optimizer changes the span of these biases, the largest minus the smallest in a layer. The span grows when all tokens favor the same experts, so it shows how hard a layer is to balance. At learning rates tuned on shorter runs, Muon raised the span of the first MoE layer at every seed, roughly doubling it on average, while the other layers changed much less. The rise came from Muon on the rest of the network, whichever optimizer trained the router. In those runs, the inputs to the first router shared a direction, and soon after the warmup its busiest expert averaged four to seven times its fair share. Over the first 3,000 steps, subtracting the batch mean from these inputs removed this overload while barely changing the loss. On the smaller model, halving Muon's learning rate removed most of the rise and slightly improved the loss. Moving only the router to AdamW made the loads shift more during the warmup and slightly raised the loss, so at this scale the results favor keeping the router on Muon.
Adversarial risk analysis (ARA) informs decision-makers when facing intelligent opponents and uncertain outcomes, enabling an analyst to model beliefs about the opponents' preferences, capabilities, beliefs, and the strategic calculations that opponents employ. Major applications cover reliability and risk analysis of critical infrastructures in presence of adversarial threats. In such domains, ARA mitigates typical common knowledge assumptions of standard game-theoretic approaches. However, ARA entails complex modelling and computations. To alleviate these, we explore its combination with fully probabilistic design (FPD), a control theory methodology supporting an agent in choosing the decision distribution that achieves the outcomes with a distribution closest to its ideal. We use FPD to model and forecast the opponent's decisions within an ARA setting and present how to handle security games within this ARA-FPD framework, covering both single and multiple stage problems.
Tushar Tyagi, Aditya Raghavan, Haoran Qiu, Francisco Romero
原文摘要
Vision-language models (VLMs) enable natural language video retrieval, bypassing traditional hand-engineered pipelines. Integrating agentic reasoning frameworks (e.g., ReAct) further advances this paradigm by allowing models to dynamically inspect video content rather than relying on a single visual pass. Applying this pattern naively to video retrieval inherits problems that video systems cannot afford: results vary from run to run under identical conditions, and cost and latency scale with every reasoning step the agent takes rather than with the work the query actually requires. We present VASTRA, an agentic video retrieval system whose execution is dictated by query complexity rather than an open-ended reasoning loop: cheap, reliably answerable queries are ranked directly from pre-ingested summaries, while harder queries trigger targeted verification only where relevance is genuinely uncertain. This keeps control flow external to the model rather than latent in a growing chain of thought, establishing deterministic control flow by construction and end-to-end rank reproducibility under greedy serving. Across 40 benchmark queries, VASTRA exposes explicit presets: VASTRA-Fast runs 12.2x faster (4.2 s vs. 51.2 s) and 10.4x cheaper ($0.0044 vs. $0.0458) than best performing ReAct agent at a 5.0-point mAP reduction, while VASTRA-Max matches peak ReAct agent accuracy (0.653 vs. 0.649 mAP) at comparable latency, trading higher spend ($0.1022 vs. $0.0458) for zero run-to-run variance.
Evaluating candidate parallelization strategies through distributed training runs makes strategy selection prohibitively expensive at scale. Performance simulation offers a lower cost alternative, yet controlled comparisons among existing simulators remain difficult because they differ in the training stages they model and the workloads on which they are evaluated. We introduce LM-SimBench, a unified framework, profiling dataset, and benchmark for evaluating LLM training simulators. The framework decomposes simulation into functional stages from model compilation to schedule level aggregation, connecting execution representations and cost estimators to measurable evaluation targets. Following this framework, we profile 23 model settings from 10 dense and Mixture-of-Experts (MoE) LLM families on an Ascend Neural Processing Unit (NPU) cluster. The resulting dataset contains 348 feasible configurations that jointly vary five parallelism dimensions and activation recomputation, with six views spanning configuration, rank, operator, event, and kernel granularity. Evaluations of representative training simulators demonstrate how LM-SimBench supports comparisons of prediction accuracy, assessment of configuration selection, and diagnosis of prediction errors. Matched GPU-NPU experiments validate the cross-platform applicability of the collection and evaluation protocol and demonstrate NPU-guided GPU configuration selection.
A single validation-tuned threshold can become brittle when dense-score distributions shift. We introduce AutoThreshold, a post-hoc policy that predicts the joint mass of positive and negative pixels over score bins. This target directly parameterizes threshold-indexed confusion counts, avoiding products of independently estimated prevalence and class-conditional rates. Softmax normalization and cumulative decoding yield coherent, monotone confusion profiles; a plug-in decoder then selects a threshold for any $F_β$ objective without retraining the head. Across three independently trained Kvasir pipelines transferred to four polyp datasets, Joint exceeds continuous Otsu by $0.0080 \pm 0.0046$ external macro Dice, with positive gains in 3/3 runs. On the reference pipeline, three head seeds achieve $0.6324 \pm 0.0027$, 0.0085 above Otsu; direct-utility Ridge achieves 0.6342, with the paired difference interval spanning zero. The same Joint checkpoints outperform Otsu for $F_{0.5}$, Dice, and $F_2$. On developmental KITTI Road leave-one-corruption-family-out evaluation, Joint improves over scalar regression by 0.0082 [0.0064, 0.0101], while direct utility is 0.0022 higher. These results establish joint confusion prediction as a useful, metric-retargetable inductive bias, without claiming Bayes-optimal $F_β$ decisions.
Long-horizon embodied tasks require agents to learn, reason, and generalize robustly across complex real-world environments. However, existing approaches often struggle with partial observability, fragmented spatial reasoning, and inefficient integration of heterogeneous memories, limiting their capacity for long-horizon adaptation. To address this, we introduce RoboMemory, a brain-inspired framework that unifies Temporal, Spatial, Semantic, and Episodic memory for efficient long-horizon planning and interactive learning. Its core innovations are a comprehensive memory system designed for embodied tasks, and a dynamic spatial knowledge graph for scalable, consistent memory updates. Extensive experiments on EmbodiedBench show that RoboMemory improves the average success rate of Qwen2.5-VL-72B-Ins by 26.5 percentage points and achieves performance comparable to strong closed-source VLM agents. Real-world trials further suggest RoboMemory's potential for interactive environmental learning by reusing accumulated experience across repeated task attempts. These results demonstrate RoboMemory as a scalable memory-augmented framework for embodied agents, improving long-horizon planning and cumulative learning in complex interactive environments.