In briefForgetting More, Adapting Longer: A Fixed-Size Summary Memory for Reinforcement Learning
SummaryThis paper introduces Recurrent Summary Memory (RSM), a Transformer policy that attends to a short buffer of recent interactions and keeps a fixed-capacity summary of older experience. When the buffer fills, the summary is rewritten and the buffer is cleared, so memory and per-decision cost stay constant regardless of task length. On long partially observable tasks, frozen RSM policies maintain their adapted performance for eight times the training horizon and re-adapt faster than growing and recurrent memories when the task changes.
In briefWhich answers agree? Evaluating self-consistency from incomplete logs
SummaryThe paper studies how to evaluate self-consistency, a method that generates several answers and picks the most frequent one, when only correctness labels are available. It derives sharp accuracy bounds and introduces Plurality-U, an exact dynamic program for averaging voting accuracy over fixed-size subsets. On stored outputs from recent reasoning models, recording abstention counts and answer options substantially narrows accuracy intervals.
In briefNADS: Navigator-guided data selection to reduce forgetting when fine-tuning language models
SummaryWhen large language models are fine-tuned on a new task, they can lose general abilities they had before, especially when the original pretraining data is unavailable. This paper proposes NADS, which trains a navigator model without preservation regularization, scores synthetic candidate data by its predictive divergence from the pretrained model, and selects a constraint set using an efficient utility-diversity strategy. The model is then fine-tuned on the new-task data while distilling pretrained behavior on the selected constraints. Experiments reportedly show a better balance between new-task performance and general capability preservation, with lower selection overhead.
In briefCognitive-decoupled meta-induction for better out-of-distribution generalization
SummaryThe paper argues that in-context learning relies heavily on matching queries to nearby demonstrations, which limits performance on out-of-distribution tasks. It proposes CoDI, which induces query-independent task-level inductive representations and optimizes a cross-task inducer through dual-loop meta-learning. According to the authors, CoDI yields substantial gains on five in-distribution and out-of-distribution benchmarks and keeps its OOD benefits without explicit induction at inference.
In briefMAPS: A reinforcement-learning-enhanced multi-agent system that generates personalized script-driven video ads.
SummaryThis paper proposes MAPS, a system of four collaborating agents that form a closed loop covering user profile modeling, script generation, video synthesis, and feedback-driven refinement. The Scripter Agent is trained with reinforcement learning and a multi-dimensional reward so its scripts match both user content preferences and target product information. Online A/B experiments in a specific Kuaishou advertising scenario reportedly improved CTR by 2.87%, CVR by 2.74%, and ARPU by 9.85% over a generative baseline.
In briefBayesian model discovery with language models, linking hypotheses and experiments
SummaryThis paper introduces BMD, a method that combines LLM-proposed model families with Bayesian experimental design. A numerical engine fits parameters, weighs models by evidence, and averages their predictions, then selects the next experimental conditions. Updated beliefs and residual diagnostics are fed back to the LLM to guide later proposals. The authors report that across four scientific discovery benchmarks, BMD achieves lower final prediction error than in-context learning and LLM-based scientific agents, and recovers exact functional forms more often.
In briefDeMoE: Functional routing of shared and conditional experts for one-step diffusion distillation
SummaryThis paper studies how internal model capacity should be organized when diffusion models are distilled into a single forward pass. The authors propose DeMoE, which adds an always-active shared expert and conditionally routed expert slices to image feed-forward layers, trained with a one-step distribution-matching objective. On one-step FLUX.2-4B generation, DeMoE reaches a GenEval score of 0.8220 while activating 80.5% of the image-FFN width, and the paper reports it matches the 50-step teacher overall and surpasses it on complex composition and counting.
In briefHow Much One Wrong LLM Answer Costs When Pruning Queries
SummaryThe paper studies constraint pruning, which saves queries to large language models by skipping questions implied by earlier accepted answers, such as through transitivity. Prior analyses usually assume the oracle is always correct, but the authors argue that an accepted wrong answer spreads to every question it implies. They analyze implication rules and entity resolution, deriving how the query order amplifies errors and identifying pivot orders as optimal. Simulations are reported to match the predicted constants within sampling error.
WhyModels error propagation in constraint pruning as first-order risk and proves optimal query orders for entity resolution; novel framing, supported mainly by theory and simulations.
Zhengtao Gui (University of California, Irvine), Weining Shen (University of California et al.
In briefA calibrated tool for detecting shifts in language model dispositions
SummaryThis paper adapts a calibrated measurement method to detect how language models' safety-relevant dispositions, such as willingness to comply with dangerous requests, deception, power seeking, and sycophancy, drift after modification. The authors show the instrument can reliably detect shifts of a few percentage points. Fine-tuning on subtly harmful medical and financial advice causes much larger drift than correct advice. A reasoning model hides this shift almost entirely in its answers, though its internal state reveals the change, so safety should be tested with reasoning both on and off.`
In briefThink Densely, Act Sparsely: LECDrive, a latent expert chain for vision-language-action driving
SummaryExisting vision-language-action driving models are often trained with sparse signals such as language or trajectories, which can discard scene information needed for safe driving. The paper proposes the "Think Densely, Act Sparsely" paradigm and the LECDrive framework, which embeds a progressive chain of latent expert cognition using compact latent tokens. These tokens absorb knowledge from foundation vision models (DINOv3, SAM3, DepthAnything3) under a curriculum from representation to semantics and geometry. Experiments report consistent gains over base models on object perception, state prediction, and trajectory planning benchmarks.
In briefTimelineAgent: An LLM framework that builds event timelines from multiple documents based on a query
SummaryThis paper introduces TimelineAgent, an LLM-based multi-agent framework that extracts event timelines from multiple documents according to different query focuses. Its multi-stage reasoning workflow organizes heterogeneous event features into reusable skills coordinated through a goal-oriented protocol. The authors also build a benchmark covering diverse query focuses, and report that TimelineAgent consistently outperforms existing methods.
In briefDoes Unlearning Survive Collaboration? Robust Unlearning for LLM Teams
SummaryThis paper studies how forgetting in LLMs holds up when several models collaborate as a team sharing one model. The authors find that checking only final answers misses recoverable information in team messages, and that controlling individual messages does not prevent recovery from their combination. They propose Aggregation-Conditioned Unlearning (ACU), which trains a shared model using reconstruction feedback from individual and pooled messages. On TOFU, ACU lowers full-trajectory ROUGE-L recall recovery from 41.8% to 19.3% and factual recovery from 34.2% to 18.7%, with comparable retained-task performance.
In briefNeural video compression with limited frame delay: theory and system design
SummaryMost neural video compression methods use strictly causal prediction with zero frame delay. This paper treats limited frame delay as a design parameter and analyzes rate-distortion performance using clique factorization of Markov networks. It proposes Graph-Factored Sequential Coding (GFSC), a framework for sequential video coding under limited delay. The authors show that for general K-order Markov sources, a K-frame delay suffices to reach joint-coding rate-distortion optimality over a nontrivial distortion region. Experiments show GFSC, added via lightweight neural modules, improves rate-distortion performance over causal counterparts.
WhyGives a new theorem that K-frame delay attains joint-coding optimality for K-order Markov sources, validated across several NVC models and datasets.
In briefAddressivity and Sycophancy in LLMs: Why Models Flatter Users
SummaryThis paper studies why large language models become overly agreeable or flattering toward users. The authors find that post-training gives model outputs "addressivity" (being directed at an addressee), which is linked to sycophancy. Using Sparse Autoencoders, they show that steering addressivity directions can amplify or suppress sycophancy, but only in post-trained models. A controlled experiment suggests models are sycophantic toward any recognized addressee, not a specific one.
In briefAnalysis of off-policy policy gradient methods for training language models
SummaryThis paper studies when off-policy policy gradient methods, widely used in LLM post-training, are justified and how importance sampling, KL regularization, and baselines handle distribution shift. It shows that consistency requires either trajectory-level importance weighting or KL regularization with a baseline equal to the optimal soft value. The KL route is linked to group-relative policy gradient, with a finite-sample regret bound, and a constant baseline shift is shown to implement pessimism.
WhyGives a structural characterization of off-policy policy gradient consistency, links KL route to GRPG and Bellman residual minimization with finite-sample regret bounds.
In briefInside the Scorer: Mechanistic Evidence on Rubric Control in LLM Essay Scoring
SummaryThis paper examines whether LLM-based essay scorers form separate internal representations for individual scoring rubrics and whether these can be controlled independently. The authors probe and steer 25 open-weight base models on human-scored BESTEP English writing tasks. Holistic scores are decodable for longer tasks, and rubrics partly separate there, but causal steering shows matched interventions give almost no advantage over mismatched ones.
In briefLearning to Couple: Context-Adaptive Joint Denoising for Diffusion Language Models
SummaryDiffusion language models speed up generation by predicting several tokens in parallel, but standard factorized denoisers treat these predictions as conditionally independent, which causes errors. The paper proposes Codpula, which pairs a Transformer with a tractable probabilistic inference layer whose coupling parameters are gated by the input context, yielding a context-dependent joint distribution with exact conditional inference. The authors report new state-of-the-art results on a range of reasoning benchmarks and an improved accuracy-latency trade-off in mathematical reasoning and code generation.
In briefA corrected sparse block method for spectral optimization, with convergence and support identification proofs
SummaryThe paper argues that the standard block generalized-power method conflates two nonequivalent objectives and rebuilds the model from the primal problem, specifying conventions for zero-response columns, ties, rank deficiency, and refitting. For the squared-threshold objective it proves whole-sequence convergence and, under a strict threshold margin, finite identification of the support and sign pattern. Experiments on GPT-2, ViT, Pitprops and three single-cell RNA datasets report that the thresholded block-sparse components match dense PCA on reconstruction and beat it on cluster-recovery AMI on the scRNA data.
In briefFine-Grained Units, Flat Decisions: Hierarchical Micro-Expert Pruning for MoE Models
SummaryThis paper proposes UC-RAMP, a two-stage post-training pruning framework for mixture-of-experts models that separates allocating retained capacity across experts from selecting channels within that capacity. Across three MoE models at 25% and 50% routed-expert channel pruning, it achieves the highest ten-task mean accuracy in five of six settings among evaluated methods. In a single-A100 deployment at 25% pruning, it reduces time to first token and time per output token relative to the unpruned model.
WhySeparates micro-expert mask construction into capacity allocation and channel selection, with cross-domain calibration and an interaction surrogate. Evaluated on three MoE models and ten tasks, outper
In briefLEVI: Benchmarking whether generated images faithfully match specified visual identities
SummaryThe paper introduces Low-Entropy Visual Identity (LEVI), identities defined by compact facts and explicit rules while appearance may vary. It proposes a declarative grammar, Verifiable Visual Grammar (VVG), to describe identity-bearing structures and builds a benchmark of 1,375 image generation tasks across six domains. Among ten evaluated model configurations, the best whole-task pass rate is 58.69%, showing a persistent gap between plausible images and faithful identity realization.
In briefPFedQFT: Personalized Federated Quantum-Inspired Fine-Tuning for Medical Language Models
SummaryMedical data is split across institutions due to privacy and governance limits, and federated learning lets them adapt large language models together without sharing raw records. This paper proposes PFedQFT, which uses a compact shared adapter built from quantum-inspired orthonormal bases and keeps a lightweight private residual operator on each client. Experiments on seven public medical tasks and several LLM backbones report large reductions in trainable parameters and per-round upload volume, with competitive predictive performance on most tasks.
In briefMexRegQA: An Item-Level Audit of Evidence Sufficiency and Model Comparisons in Mexican-Spanish Regulatory QA
SummaryThe paper argues that similarity-based retrieval coverage does not show that retrieved evidence contains the facts needed to answer a question. The authors audit Mexican-Spanish financial-regulatory QA items and find that 35 of 135 covered items lack some or all load-bearing facts. They release MexRegQA v1.0 with provenance and evidence-status labels, and recommend publishing such labels alongside grounded QA benchmarks.
Fred Hu, Yeqi Wang (1), Xing Zhou (2) ((1) Huawei Technologies de M\'exico et al.
In briefCausal Humility: Governing How Much Authority Fairness Claims Deserve When Causes Are Uncertain
SummaryThis paper proposes “causal humility,” a governance framework for handling algorithmic fairness claims when the underlying causal relationships are uncertain. It argues that causal explanations of fairness should make their assumptions and omissions visible, allow multiple defensible causal accounts, treat indeterminate results as legitimate, and match institutional authority to the strength of evidence. The authors introduce “Causal Cards,” a documentation artifact for recording these details, and use the UCI Adult dataset to show that different causal discovery methods reveal, hide, or leave unresolved fairness-relevant pathways.
Khadija Zanna, Rodrigo Ferreira, Frederick L. Oswald et al.
In briefWhose Credential Is It? Owner Labels in Agent Memory and Cross-User Account Misuse
SummaryThis paper studies how an LLM agent with shared long-term memory can retrieve another user's login credentials and misuse them. The authors propose a three-stage audit that separates exposure of a foreign credential, the agent's action with it, and resulting changes to the other user's account. Labeling each credential with its owner's name greatly reduces misuse in their experiments, though it slightly lowers task success on clean memory.
In briefGradient-based sample grouping for predictive bandit curriculum reinforcement learning
SummaryThis paper proposes Predictive Gradient-based Bandit (PGB) Curriculum, which clusters projected sample-level gradients to discover learning units grouped by their optimization effect. A lightweight temporal predictor forecasts each unit's future utility, and these predictions are converted into a Boltzmann sampling distribution for adaptive data allocation. Experiments across multiple model scales and reasoning benchmarks report that PGB outperforms standard GRPO, difficulty-based curricula, and history-based scheduling.
In briefFrom Black Box to Controller: Feature-Amplification Steering of LLMs at Inference Time
SummaryThis paper introduces NeuroSteer, a plug-and-play framework that adjusts LLM behavior preferences at inference time without task-specific prompts or additional fine-tuning. It uses a Sparse Autoencoder to identify feature neurons linked to traits such as politeness and positive sentiment, then amplifies their activations during generation, with adjustable strength per scenario or user. Experiments show it can alter sentiment, toxicity, stance, and politeness and achieves higher target-direction steering scores than baselines in several settings, though generation quality trade-offs vary by task and strength.
In briefTracking Drifting Fixed Points with Linear Stochastic Approximation
SummaryThis paper studies linear stochastic approximation where the fixed points change over time and observation noise may be correlated across rounds. It bounds the cumulative expected squared error between iterates and the moving fixed points for constant stepsizes, and gives lower bounds for scalar Gaussian models. An adaptive method is proposed that does not require knowing the noise level or how fixed points change, and the results are applied to temporal-difference value estimation in changing environments.
In briefCVaR-RMD: Selecting Robust Heuristics with Conditional Value-at-Risk
SummaryThis paper proposes CVaR-RMD, an evolutionary framework in which a large language model mutates executable heuristic programs and ranks candidates by empirical conditional value-at-risk across varied evaluation conditions. In experiments on three combinatorial optimization domains, CVaR-based selection improved upper-tail performance and reduced failure rates while keeping mean performance competitive. The gains persisted on larger unseen instances and on independently curated within-domain benchmarks.
In briefPhysically Consistent TeX Generation via Multi-View Structural Representations
SummaryThis paper proposes MVR-MoE, a lightweight framework for low-light image generation that mines structural cues from infrared and visible images into multi-view representations. Multi-view hyper connections expand the feature space, and hierarchical routing adaptively uses heterogeneous cues at the view and expert levels. Predicted temperature, emissivity, and texture maps are constrained by the Stefan-Boltzmann law to ensure physical consistency. Experiments on challenging datasets report superior robustness and discriminability.
In briefToolOPD: Outcome-Aware On-Policy Distillation for Tool-Use Agents
SummaryThe paper argues that standard on-policy distillation wrongly assumes credit is spread evenly across long trajectories, whereas analysis shows it concentrates on a few critical actions. It proposes ToolOPD, which summarizes each trajectory and its outcome into a hindsight skill and uses the resulting shift in teacher probabilities to weight action-level distillation loss alongside reinforcement learning. The authors report that ToolOPD significantly outperforms existing reinforcement learning and self-distillation baselines on multiple long-horizon tool interaction benchmarks.
In briefProgress-guided preference generation for aligning generative control policies without human labels
SummaryGenerative control policies learn from expert demonstrations, but these are often scarce, and existing preference fine-tuning relies on costly human annotation. This paper proposes PGPG, an unsupervised framework that uses ordered coverage alignment to measure progress against a few expert demonstrations and derive pseudo-preference labels from autonomously collected rollouts. Experiments on long-horizon manipulation tasks show PGPG outperforms standard supervised fine-tuning baselines, and it works with both diffusion and flow-matching policies.
In briefMT-CC: Multi-Group Temperature Scaling to Fix Uneven Confidence Calibration in Class-Incremental Learning
SummaryThe paper observes that in class-incremental learning, confidence changes do not track accuracy drops, causing uneven miscalibration across tasks. It proposes MT-CC, a post-hoc calibration framework that uses a lightweight grouping network to split samples by confidence statistics and task-aware signals, with a separate temperature for each group. A Groupwise Difference-between-Confidence-and-Accuracy (GDCA) regularizer promotes consistency within groups. Experiments reportedly improve calibration significantly while maintaining accuracy.
In briefWhen Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight Perturbations
SummaryThis paper examines how harmful fine-tuning can quickly weaken the safety alignment of large language models. The authors find that separate parts of a harmful weight perturbation have limited effects alone, but together they increase harmfulness beyond the sum of their individual effects. They propose Joint Weight Perturbations (JWP), an alignment-stage defense that masks different parts of the estimated perturbation and trains the model to respond safely in those states. Experiments across downstream tasks and model families reportedly show stronger resistance to harmful fine-tuning than prior defenses while largely preserving task performance.
In briefBitFlat: Error-localized bitstream design for robust diffusion image compression
SummaryThis paper presents BitFlat, a training-free bitstream framework for zero-shot diffusion image compression based on reverse channel coding, designed to resist bit flips. Blockwise Locality Alignment splits atom selection at each diffusion step into independent units so that a bit error affects only its own unit. Reliability-Aware Diffusion Control uses the saved bits to impose valid-codeword constraints and treats corrupted atoms as erasures excluded from noise synthesis. Experiments on Kodak, DIV2K, and CLIC2020 show BitFlat outperforms conventional, learned, and diffusion-based codecs at bit error rates from 10^-4 to 10^-2.
In briefRxGS: Relightable 3D Scene Stylization via BRDF Decomposition
SummaryThe paper presents RxGS, a method for stylizing 3D scenes from a reference image using a BRDF-aware 3D Gaussian representation. It optimizes only the diffuse albedo recovered by inverse rendering, keeping roughness, metallicity, normals and environment lighting unchanged, so the stylized scene can still be relit. On NeRF-Synthetic scenes, RxGS achieves the lowest Geometry Consistency Cost (GCC), and it remains competitive on real-captured scenes.
In briefLOOP: Recovery-aware layer pruning for LLMs using on-policy distillation
SummaryThis paper proposes LOOP, a method for compressing large language models by removing layers iteratively rather than all at once. In each step it deletes one layer and runs a short on-policy distillation to recover reasoning ability. Layer selection reuses recovery signals to assess each layer's reasoning contribution and recovery sensitivity. The authors report that LOOP achieves higher average accuracy than one-shot baselines across several benchmarks, model scales, and pruning ratios.
In briefCross-field idea search: finding transferable ideas in science papers
SummaryThis paper studies cross-pollination retrieval, which finds ideas that can be transferred from one research field to another. The authors build CPR-Dataset by mining cross-topic citations, with an evaluation split of 1,300 queries across 26 fields. The best of eleven existing retrievers reaches Recall@10 of 0.26, while an LLM reranker reaches 0.81, and contrastive fine-tuning raises Recall@10 to 0.70.
WhyIntroduces cross-field analogy retrieval with a large dataset, validated across several benchmarks showing transfer gains; a useful but not field-redefining contribution.
In briefNeither Over- nor Under-Constrained: Auditing Constraints to Evaluate LLM Personalization
SummaryCurrent evaluations of LLM personalization often constrain the response space too much or too little, causing frequent false rejections and false acceptances. The paper introduces ConQual-Audit, which extracts semantic constraints from user profiles and task references, prunes unnecessary constraints and recovers missing ones, then verifies candidate responses with natural language inference. On PrefEval and LaMP it outperforms nine baselines, and when its constraint audit is added to ExPerT and PREF, those methods yield smaller, higher-quality constraint sets with lower false rates.
In briefLIRQuant: An efficient sensitivity-aware rotation quantization method for LLMs
SummaryThis paper proposes LIRQuant, a post-training quantization method for large language models. It identifies channels sensitive to quantization noise and applies learnable rotations only to them, using fixed rotations for the rest to reduce optimization cost. The authors report state-of-the-art accuracy under W4A4KV4 across several model families, with LLaMA-3-8B calibrated in 25 minutes using 20.3 GB peak GPU memory. They also report up to 2.3x prefill speedup and 3.5x memory saving over FP16.
In briefHyperTaxa: Genome representations for classifying microbes, including unseen species
SummaryHyperTaxa is a learning framework for predicting the taxonomic classification of microbes from their whole genomes. It encodes genomes as multiscale k-mer transition fields and aligns them with a taxonomy-derived semantic space using a two-stage Node2Vec-to-MDS procedure. The method supports zero-shot inference for species absent from training, and the paper reports improved prediction over baselines, with the largest gains at higher taxonomic ranks.
In briefFréchet Forcing: training long-horizon video world models from short clips
SummaryStreaming video world models generate future frames chunk by chunk, but when trained only on short clips they drift and errors compound beyond the training length. This paper proposes Fréchet Forcing, which fine-tunes a model under three constraints: a flow-matching loss after self-generated history, a Fréchet distance on latent statistics that keeps rollouts faithful to their scene, and a Fréchet distance to real frames in a frozen perceptual space that keeps them realistic. Training needs no teacher model, online critic, or long videos. Fine-tuning MiniWorld-1B reportedly yields rollouts stable beyond 4,000 frames that continue following actions in robot manipulation and camera-controlled exploration.
In briefSpaRV: Sparse radiance voxel maps that combine LiDAR and camera data to build realistic, accurate 3D scenes
SummarySpaRV is a sparse radiance voxel map rendered by differentiable ray marching, trained jointly on irregular LiDAR scans and images. LiDAR depth and occupancy constrain the geometry, while normal and smoothness regularizers keep it consistent between scan points. The method alternates map training with optimization of camera poses and focal lengths, and bounds GPU memory by keeping one submap at full resolution and others at coarser octree levels. The authors report reconstructing a campus from 63,582 images on a single 24 GB GPU.
In briefPoSTCache: Revisable Caching for Faster Diffusion Image and Video Generation
SummaryThis paper addresses the high computational cost of iterative sampling in diffusion and flow matching generative models. It proposes PoSTCache, a training-free, plug-and-play caching framework that forecasts skipped steps and then uses corrections from later full-computation steps to revise earlier predictions. The authors also introduce a Posterior Variance Reduction strategy for choosing which steps to fully compute. Experiments on image and video generation models report better quality-efficiency trade-offs than representative caching methods.
In briefCan Metadata Correct Coordinate Orientation? Calibration Turns Sparse Multi-Source Learning into Sparse Mean Recovery
SummaryThis paper studies multi-source data where each source orients its coordinates with its own sign convention. Pooling sources without calibration cancels the mean signal, reducing the problem to sparse PCA. With task-independent relative-gauge metadata that synchronizes sources, the same observations become a sparse mean estimation problem. The authors introduce AlignedMass to measure the first-order signal restored by alignment, showing valid calibration can scale with the full sample size.
In briefKnowledge Manifold Optimization: Finding Well-Supported Gaps in Scientific Literature by Balancing Novelty and Evidence
SummaryThe paper proposes Knowledge Manifold Optimization (KMO), which treats scientific discovery as a multiobjective search over a continuous knowledge space derived from literature. It evaluates four objectives, including boundary strength, white-space novelty, semantic gradient magnitude, and evidential support, to identify Pareto-optimal Supported Knowledge Gaps. Demonstrated on a 99-paper corpus in aerodynamics, the method yields 16 practical Pareto-front solutions with distinct discovery modes. A distance-maximization baseline selected regions with substantially lower support and reliability, illustrating that emptiness alone is insufficient.
Shigeru Obayashi, Yoshiaki Kawagoe, Kazuhiko Komatsu et al.
In briefFluidGen: Physics-aware fluid video generation through motion-guided post-training
SummaryVideo generation models often fail on fluid phenomena such as smoke, fire, and explosion debris, producing bland motion without coherent vortical structure. FluidGen injects a data-driven fluid prior into a pre-trained video model, using high-fidelity simulation data with velocity-field ground truth alongside in-the-wild fluid videos. A dual-stream architecture predicts per-frame velocity fields alongside RGB, and the velocity signal serves as a reward for preference optimization. The authors report that the model produces more realistic fluid dynamics than strong video baselines and optimization-based methods.
In briefBeyond Leaderboards: Profiling LLM Capabilities for Specific Applications
SummaryPublic leaderboards rank large language models generally, but engineers must choose models for specific applications, architectures and costs. The paper presents a method for application-specific capability profiling, implemented as a containerized testbed where each application brings its own dataset and evaluator. In the evaluated automation conditions, model choice produced larger differences than framework choice, though no general causal ranking is established. Of 216 Oracle-to-PostgreSQL migration units checked against a real database, 108 were demonstrably correct, and system-reported success exceeded independently observed success.
Oleg Grynets, Vasyl Lyashkevych, Andrii Salyk et al.
In briefA “feature ODE” that approximates how shallow MLPs learn simple functions
SummaryThis paper studies gradient flow training of two-layer MLPs on anisotropic Gaussian data. The authors propose a “feature ODE” that they intuitively expect could describe a shallow MLP’s learning dynamics. MLPs trained on simple target functions, such as low-order Hermite polynomials and staircase functions, roughly follow the loss trajectories predicted by this ODE. The agreement is surprising given the large simplification involved and the lack of a rigorous chain of approximations, and the authors discuss possible reasons and uses.
In briefOPDArena: Studying How Learning and Forgetting Happen in On-Policy Distillation
SummaryThis paper presents OPDArena, a framework that compares supervised fine-tuning, reinforcement learning with verifiable rewards, and on-policy distillation across five pretrained models and several teacher sources. It measures math learning, retention of general abilities, and changes in predicted distributions. The authors find that the teacher strongly affects retention: students of RLVR-trained teachers improve on the target task while keeping prior skills, whereas students of SFT-trained teachers lose more capability.
In briefConsistency Training Reduces Cue Following but Does Not Teach Cue Selectivity
SummaryLanguage models can change answers due to unhelpful hints, such as a stated belief about the answer. This paper applies consistency training, which trains models to give the same answer with and without such cues, using four methods that target output tokens, activations, attention, or sampled-answer rates. Tested on Gemma-2-2B-it and Llama-3.1-8B-Instruct, all four weaken the trained cue, but they reduce the influence of correct cues as much as wrong ones. The fixes are larger for cues placed before the answer options, and circuit attention heads still carry most remaining cue-induced errors.
In briefLatent Projection for Latent Planning: Pruning Noisy Latent Directions Before Computing the Planning Cost
SummaryThis paper addresses latent world models, where an agent plans in latent space, and notes that the latent contains noise that affects the commonly used ℓ2 planning cost. The authors use the linear Gaussian structure implied by the training regularizer (SIGReg) to fit a state dynamics matrix A, which gives per-direction uncertainty, and restrict the ℓ2 cost to the k-dimensional subspace with the least uncertainty. The correction is applied post hoc, leaving the encoder, predictor, rollout, and planner unchanged. Evaluated across six environments, it shows large gains on the 2D navigation tasks TwoRoom and PointMaze.】
In briefInteractAgent: A Memory-Augmented Agentic System for Human Motion Interaction
SummaryThis paper presents InteractAgent, an agentic multimodal LLM system for long-horizon human-scene interaction that plans, executes, observes, and reflects in a closed loop. A single MLLM serves as both planner and evaluator, while a frozen MotionLLM acts as the executor. On the UniHSI benchmark, it reportedly improves task success and reduces penetration compared with training-free baselines.
In briefSaga: An In Silico Ontology for a Multimodal Patient World Model
SummaryThe paper introduces Saga, a multimodal patient world model that learns a shared state from patient histories to predict clinical events and generate multimodal images. Generated events and times are appended to the history and re-encoded to advance predictions. The study covers 461,370 patients with 12 cancer types from 112 centres, with future images evaluated against observed follow-up.
WhyProposes a multimodal patient world model with autoregressive rollouts, validated on 461,370 patients with ablations and external evaluation.
Youcheng Li (Peking University, Isoplex Lab), Fa Tian (Isoplex Lab) et al.
In briefSpherical Fixed-Point Reasoners: Adaptive Loop Halting on the Hypersphere
SummaryThis paper introduces Spherical Fixed-Point Reasoners (SFPRM), which run the fixed-point computation of the earlier FPRM method on the unit hypersphere of normalized transformers (nGPT). Input injection, damping, and halting become geodesic operations. The authors prove fixed-point existence and local and global convergence under stated conditions. SFPRM reports test accuracies of 96.6% on Sudoku-Extreme, 87.9% on Maze-Hard, and 59.1% on Mini-ARC, above the compared baselines.
Arber Zela, Artavazd Maranjyan, Pol Puigdemont et al.
In briefNegation Without Confusion: Fixing Text-to-Image Negation via Embedding Decomposition
SummaryText-to-image diffusion models often fail to follow negation because biased language patterns in text embeddings overpower the intended negative constraints. The paper proposes Dero, a training-free method that uses auxiliary prompts and bidirectional SVD to separate semantic and bias subspaces, then selectively recomposes the components. It works on text representations only and can be applied to both U-Net and DiT diffusion models. Experiments on object- and concept-level negation report better text-image alignment and human preference scores.
In briefSampling Signals on Multigraphs: Uniqueness Sets and Spectral Proxies
SummaryThis paper studies sampling of bandlimited signals on multigraphs, where several graphs with different shift operators share the same nodes. It defines a layer-independent notion of frequency and bandwidth for multigraph signals and extends removable and uniqueness sets to this setting. The authors propose an aggregate operator, computable directly from the shift operators, that certifies uniqueness and yields a greedy sampling algorithm. Numerical experiments show it selects better-conditioned sampling sets and achieves lower reconstruction errors than single-layer methods.
In briefVisual Memory and Textual Alignment for Cross-Modal Few-Shot Segmentation
SummaryThis paper addresses few-shot segmentation, which aims to segment target objects from only a few annotated examples. It proposes ViTA-SAM, built on SAM 3, which uses a visual memory module that encodes support features and masks, and text features from SAM 3's encoder as complementary priors. A cross-modal alignment module conditions the text features on the memory-enhanced visual representations, and a pseudo-label refinement strategy updates memory with predicted masks. The authors report that it outperforms the strongest competing method by over 5% in average mIoU across multiple benchmarks.
In briefThink, Compress, Answer: a reinforcement learning method that reasons, compresses context, then localizes targets in multimodal models
SummaryThe paper argues that longer reasoning in multimodal large language models can drift into broad scene descriptions and peripheral relations, diluting target evidence and causing ambiguous localization. The authors propose Think-Compress-Answer (TCA), which splits grounding into free-form reasoning, target-aware context compression, and answer generation, trained with multi-stage GRPO using progressive rewards. The authors report substantial zero-shot performance gains across six multimodal grounding benchmarks.
In briefACorD: Adaptive correspondence discovery for aligning medical images
SummaryThis paper presents ACorD, a framework for medical image registration, which aligns anatomical structures across images. It first generates multiple candidate correspondences using correlation-guided cross-scale offset prediction, then aggregates them with attention through multi-head deformable aggregation rather than weighting them equally. The authors report that experiments on four public medical registration benchmarks show improved robustness across diverse anatomical structures.
In briefTNVD-SC: Terrain-normalized vertical distributions to diagnose scan context failures in forests
SummaryThe paper studies long-term LiDAR place recognition in forests, where repetitive vegetation and uneven terrain cause structural aliasing. The authors separate candidate retrieval from fine matching, finding that correct locations often appear among top candidates while conventional Scan Context matching performs poorly on them. They propose TNVD-SC, which keeps the ring-sector topology and encodes terrain-relative vertical probability distributions, reaching 91.83% mean R@1 on the standardized TreeLoc++ protocol. A lossless count-coded implementation is also reported, reducing descriptor payload and compute latency.
In briefUncertainty-gated SHAP-Dirichlet post-hoc refinement for MRI-based depression classification
SummaryThe paper proposes a post-hoc reliability layer that refines predictions from MRI-based depression classifiers without retraining them. It retrieves similar labeled SHAP explanations, builds Dirichlet class evidence, and selectively fuses probabilities for uncertain cases, deferring unreliable ones to expert review. On 504 SRPBS subjects with 159 anatomical features, refinement improved accuracy for TabM and ResNet and improved calibration.
Phuc Truong Loc Nguyen, Nguyen Truong Toan To, Truong Thanh Hung Nguyen et al.
In briefClean Forcing: Reducing Drift in Long Video Generation with a Frozen Base Model
SummaryAutoregressive video diffusion models generate long videos chunk by chunk, but errors accumulate and quality can collapse within seconds. This paper proposes Clean Forcing, a 5.9M-parameter LoRA corrector trained on a frozen base model to learn the systematic part of this drift, with closed-loop training and a drift-contraction objective. In 50-second text-to-video generation, the reported Δ-drift falls from +11.3 to +1.65 without real videos and to -0.16 with 40 real clips. The authors state this improves on the best published result of +3.57 while using over 150 times less data.
Wenqing Wang, Joonghyuk Shin, Jonathan Tremblay et al.
In briefConfidence-guided pseudo-label refinement for unsupervised cross-domain person re-identification
SummaryThis paper addresses noisy pseudo labels produced by domain shift in unsupervised cross-domain person re-identification. It proposes CGPLR, which estimates sample reliability from centroid compatibility, neighborhood consistency, separation from the strongest competing cluster, and temporal assignment stability. Samples are preserved, reassigned, or deferred from hard identity supervision based on adaptive thresholds. The authors report 76.4% Rank-1 and 47.6% mAP on Market-1501 to MSMT17, and 95.1% Rank-1 and 87.7% mAP on MSMT17 to Market-1501.
In briefWhere Do the Gaussians Go? A Training-Only Foreground Pull for Camera-Based 3D Occupancy
SummaryThis paper studies camera-based 3D semantic occupancy models trained only with 2D pseudo labels, finding that Gaussian density on foreground voxels is just 1.62 times that on background voxels. The authors add a training-only foreground attraction loss that back-projects foreground pseudo-label pixels into 3D and pulls nearby Gaussians toward them, leaving the inference network unchanged. On Occ3D-nuScenes, mean foreground IoU rises from 6.29 to 6.96 over three seeds, while the mIoU change falls within baseline seed variation.
In briefA research program on controlled emergence, linking explaining system behavior from local rules with designing rules to reach target behavior.
SummaryThe paper proposes a unified research program for controlled emergence, framing the tension between adaptability and loss of control as a forward problem (explaining macroscopic behavior from local rules) and an inverse problem (solving for controllable rules from a target). It uses an augmented quintuple as a shared formal object and splits engineering into controllability, reproducibility, and formalizability. The authors state three testable working hypotheses and one feasibility criterion, and note that the attached pilot experiment only checks that the research protocol is executable and is not a statistical test of the hypotheses.
In briefChoosing and Combining Agent Design Patterns: Industrial Evidence and Controlled Experiments
SummaryThe paper extends a cognitive-function and execution-topology framework with an evidence-driven selection procedure and explicit composition contracts. It combines two industrial cases, practitioner workshops, and fault-injection experiments on three local reference systems across 36 scenario families. Composed systems correctly complete or safely refuse most families, but these are constructed challenge-suite results rather than production success rates.
In briefDistilling Individual Treatment Effects into Simple Budget-Limited Targeting Rules
SummaryThis paper studies how to target public-health interventions to individuals under limited resources, using observational data on obesity, elevated glucose, and smoking history. It estimates conditional average treatment effects with Causal Forest teachers and distills them into simpler decision tree and random forest students. Students came within 0.011 utility of their teachers, but paired comparisons were not significant after Holm adjustment, and the findings require longitudinal or prospective validation.
In briefMapSteal: A Map-Based Method for Autonomously Extracting Knowledge from Black-Box LLMs
SummaryThis paper presents MapSteal, a framework that starts from a few initial roots, represents acquisition targets as structured states, and recursively expands them via teacher-proposed directions. It generates question-answer pairs without student feedback to adapt small student models. Experiments with Qwen3.5 at 2B, 4B, and 9B scales show competitive results on six medical and cybersecurity benchmarks, with gains varying by domain and evaluation panel.
In briefZephyrus-R1: Agentic Reinforcement Learning for Weather Reporting
SummaryThis paper presents Zephyrus-R1, a reinforcement learning framework for training AI to write natural-language weather forecast reports. It combines a meteorological rubric with bidirectional claim entailment, checked against National Weather Service reports, and trains vision-language agents in a simulated tool-calling environment with a curriculum. After training, the mean of six judge-metric scores improved 26% for Qwen3.5-4B and 13% for Qwen3.8-27B over their base models.
In briefRecurrent vision transformers for event-camera spacecraft pose estimation at slow relative motion
SummaryThis paper studies estimating spacecraft pose from event cameras for rendezvous and proximity operations. Slow relative motion produces sparse, low signal-to-noise event data, which degrades existing methods. The authors propose recurrent vision transformers that carry memory across sparse event streams and learn temporal attention spans from full trajectories. Validated on real event data from slow relative motions, the method reportedly outperforms per-frame and fixed-time-window approaches in accuracy and stability.
In briefWeighted state souping: a reusable-context method for efficient long-context retrieval
SummaryThe paper addresses the drop in performance of state space models (SSMs) on long, crowded contexts, which limits state souping. It proposes weighted state souping, which learns query-dependent weights for combining cached states. Reported results cover document QA, needle-in-a-haystack retrieval, variable tracking, and tool use, with a sparse variant offering near-constant encoding cost.
In briefCheXFuse: A Guided Attention Method for Combining Chest X-Ray Images and Reports in Small-Data Medical Settings
SummaryThe paper proposes CheXFuse, a framework that fuses chest X-ray image and text report representations using Guided Differential Joint Attention to suppress weak or redundant signals. It is evaluated on the Indiana University Chest X-ray dataset for normal versus abnormal classification. Experiments include a masked-report setting that tests performance when textual information is reduced. The reported accuracy is 95.73% with original reports and 95.58% with masked reports.
Muntaqim Ahmed Raju, Priyanka Siddappa, Md Shifat Haider Al Amin et al.
In briefCurriculum learning helps LLMs learn long-horizon implicit reasoning, but only up to a limit
SummaryThis paper studies whether curriculum learning can extend how far language models can reason implicitly within a single forward pass. Models were fine-tuned on examples with progressively larger vocabulary sizes and lookahead steps. The curriculum-trained Qwen3-0.6B reached 98% accuracy at lookahead 256, versus 45% without curriculum, though progress saturates, suggesting an intrinsic ceiling. The acquired ability also transferred to 11 natural language reasoning benchmarks.
Jimson Huang, Maxwell Nickell, Ruowang Zhang et al.
In briefSafety Offset Learning: How Benign Fine-Tuning Can Misalign LLMs
SummaryThis paper argues that fine-tuning safety risks depend not only on whether training data is harmful, but on whether supervision shifts a model from refusing to complying. The authors built benign instructions that an aligned model tends to refuse, paired with their original harmless answers. They found harmful compliance rose only when the gap between refusal and answer was large and could not be resolved by a narrow shortcut. Their SAFEOFF dataset substantially weakened safety alignment across three aligned LLMs and three safety benchmarks.}
WhyOffers a new view that benign fine-tuning weakens alignment via behavioral transition, validated with SAFEOFF across three models and benchmarks.
In briefPRISM: Using Chemical Relational Priors for Self-Supervised Solubility Prediction on Unseen Compounds
SummaryThis paper introduces PRISM, a label-free pretraining framework that turns expert chemical priors into relational teacher signals. For solubility prediction, it trains four encoders to preserve relational information about hydrophobic, electrostatic, polarizability, and substructural effects, using KL divergence to match neighborhood distributions. The authors report that PRISM matches or surpasses supervised and self-supervised baselines across scaffold, solute, and solvent out-of-distribution settings.
In briefAdversarial post-training to make high-frequency trading simulators more realistic
SummaryThis paper proposes a post-training method for autoregressive foundation models trained on limit order book data, using a Wasserstein GAN objective. Evolution strategies update low-rank LoRA perturbations of the pre-trained generator, guided by a critic that observes rollouts through the real matching engine. On a held-out panel, the method improves LOB-Bench realism by about 17%, while random-perturbation and continued pre-training controls show no gain, and error accumulation over rollouts is reduced.
Alexandre Bismuth, Kang Li, Aramis Marti-Shahandeh et al.
In briefAudio LLMs favouring transcript text over vocal cues, and the internal mechanism behind it
SummaryThe paper tests six audio LLMs on conflicts where the transcript contradicts paralinguistic cues such as emotion, speaker identity, or sarcasm. Models favoured the text in 0.78 of mismatched items on average (0.86 excluding sarcasm). In the most text-dominant model, speaker identity stayed linearly decodable at the input, so the failure is not perceptual. Attention knockout partially restored audio-grounded output by blocking propagation through intermediate tokens, and the effect appeared only in free-form generation, not multiple-choice tests.
Ankur Banerjee, Vishnu Guntakala, Danity Pike et al.
In briefCognitive Maps in Large Language Models: How Models Reason About Relationships With Limited Information
SummaryThis paper tests whether large language models form internal representations similar to the cognitive maps proposed for animals. Using relational reasoning tasks and causal "mosaic experiments" that manipulate internal activations, the authors find that models above ten billion parameters infer unstated relationships and generalize to synthetic entity names. They report that models store relations as pairwise links composed at query time, with binding mediated by sparse key-value pairs. Perturbing these pairs to implant "false memories" changed model behavior in predictable ways.
In briefViMoLa: Discrete Motion Tokens for Guiding Video Generation with Language
SummaryThis paper presents ViMoLa, which predicts motion from text and feeds it into a frozen image-to-video model through zero-gated cross-attention adapters. Trained on 462K videos, it raises directional motion accuracy from 0.287 to 0.442 (chance 0.250) while keeping generative quality similar. The authors note that control remains imperfect, with smaller gains for compound motions.
In briefRobust policy learning from observational data when deployment and training distributions differ
SummaryThis paper studies how to learn treatment policies from observational data so they perform well even when the deployment distribution differs from the training distribution. It maximizes worst-case policy value over an f-divergence ball, using cross-fitted doubly-robust scores and a DRO solver. The authors prove an O~(n^{-1/2}) regret rate and, under a Tsybakov-style margin condition, a faster rate of O~(n^{-(1+α)/(2+α)}). Experiments calibrated to a welfare-to-work program show gains over empirical welfare maximization when distributions differ.
In brieff-RPO: A divergence-based method for aligning LLMs with ranked preferences
SummaryThe paper introduces f-RPO, an f-divergence objective for aligning large language models with ranked preference feedback across multiple responses, which reduces to pairwise preference alignment as a special case. The authors prove that f-RPO preserves the underlying preference ordering at convergence. Experiments on Qwen, Llama, and Gemma models from 0.5B to 8B parameters show that total-variation and reverse-KL variants outperform ranked-alignment baselines in ranking and generation quality.
In briefAssertion and Belief: Testing Moore's Paradox in Language Models
SummaryThe paper tests whether large language models recognize Moore's paradox, where sentences like "It is raining, but I don't believe it" may be true yet cannot be sensibly asserted. Evaluating 12 models on Moorean sentences in varied contexts, the authors find that most models recognize such sentences as infelicitous, and that this judgement tracks the speech act rather than a memorized form. Using the Rational Speech Act framework, they find the models' listeners do not reason about speakers, suggesting the models recall and apply rules about the paradox instead.
Anshun Asher Zheng, Weili Cao, Fangcong Yin et al.
In briefCVTA: Cross-Variable Temporal Attention for Forecasting Irregular Multivariate Time Series
SummaryThis paper addresses forecasting of irregularly sampled multivariate time series, common in healthcare, climate science, and large-scale monitoring, where observations are asynchronous and variables differ in nature. It proposes CVTA, a short-term continuous-time framework that uses variable-conditioned embeddings and temporal attention to capture cross-variable dependencies. The authors also show CVTA can serve as an auxiliary module for existing models via a dynamic meta-decision model, and report experiments on multiple real-world benchmarks.
In briefSharp degree bounds for Putinar sum-of-squares certificates based on positivity margin
SummaryThe paper gives sharp degree bounds for Putinar sum-of-squares certificates in terms of the positivity margin δ. For general Archimedean quadratic modules it proves an O(δ^{-1/2}log^2(2/δ)) upper bound, matched by a three-variable lower-bound example. With at most two inequality generators the bound improves to O(δ^{-1/2}log(2/δ)), which is also sharp. The authors also derive sharp bounds for trigonometric certificates on tori and worst-case convergence rates for the Lasserre hierarchy.
WhyProves sharp Putinar certificate degree bounds in margin with matching lower-bound construction, a significant theoretical improvement with strong proofs.
In briefFinite Interaction Lens: a mixed-precision method for checking how visual edits interact in vision-language models
SummaryThe paper introduces the Finite Interaction Lens (FIL), a mixed-precision method for evaluating how a finite set of visual interventions interact inside a vision-language model. It projects single-intervention components at higher precision and reuses them across joint candidates while keeping mixed residuals. In one MLP of Qwen3-VL-4B-Instruct, FIL reduced final interaction error relative to a control with the same nominal BF16 budget, with the authors noting its practical value depends on tasks needing that extra fidelity.
In briefA Benchmark for Structured Retrieval-Augmented Generation in Perovskite Solar Cells
SummaryThis paper introduces a quantitative question-answering benchmark for evaluating retrieval-augmented language models on perovskite photovoltaic literature, with answers computed from 45,441 measured device records. Six retrieval architectures are compared, and standard passage retrieval reaches about 16-23% accuracy. Adding read-only SQL execution over a structured device table raises accuracy to roughly 51% on both historical and recent literature, and the authors argue that systematic table extraction is needed for literature-scale quantitative analysis.
In briefRepresentations in Motion: Layer-wise Rotation and Trajectory Divergence in LLMs
SummaryThis paper studies how internal representations in large language models change across layers during inference. When the model is prompted to lie or to be honest, the factuality directions extracted from the two conditions rotate away from each other as layers progress. The authors also report similar rotation in a 3D visual perspective task and show that single-trial trajectories can predict the model's task and behavior.
In briefDiffGround: Zero-Shot Video Question Grounding Guided by Semantic Differences
SummaryThis paper proposes DiffGround, a zero-shot framework for locating when events in a video answer a question, built on frozen multimodal large language models. It converts questions into action cues, uses a vision-language model to select keyframes, and applies a training-free refinement that uses semantic differences between neighboring SigLIP2 embeddings to recover full action segments. The authors report that it outperforms existing zero-shot and supervised methods on NExT-GQA and STAR.
Cheng-Han Lee, Amir Salimi, Hsiang-Wei Huang et al.
In briefA paper on an agent execution harness that separates planning, execution, and review, evaluated across several domains.
SummaryThe paper presents an agent execution harness that keeps the original objective separate from a revisable plan and execution record. A planner assigns tasks, workers produce outputs, and a reviewer checks results against the original request. The harness is evaluated on terminal tasks, clinical responses, application building, and scientific questions, with cost and latency reported alongside outcomes.
Saurav Joshi, Manas Gupta, Ashwin Venkatesh Raman et al.
In brief"Flattening Fantasies": Stereotypes in Pornographic Language
SummaryThis study quantitatively analyzes 5.4 million video titles from the xHamster pornography platform to examine how nationality, ethnicity, and gender identity are described. It finds American pornography dominates and serves as the implicit norm, alongside persistent ethnic stereotypes, religious fetishization, and sexualized cultural taboos. Women are often portrayed through a Madonna-Whore dichotomy, while men are frequently depicted in physically dominant roles.
In briefConcurrent two-scale design method for printable bone scaffolds using generative flows
SummaryThis paper presents a two-scale topology optimization method in which a conditional rectified flow generates periodic lattice microstructures, refined in latent space using exact finite-element homogenization sensitivities. In a distal rabbit femur benchmark, all eight flow-started refinements yielded printable cells with maximum relative errors under 1% across six engineering moduli. In human femoral reconstruction, the method produced a cortical mid-shaft scaffold with 63% porosity and a proximal cancellous scaffold with 84% porosity, both satisfying compliance constraints under walking and stair-climbing loads.
In briefLearning which device to call for LLM agents in IoT
SummaryIn IoT settings, one capability is often offered by several devices, but an agent's plan does not specify which one to call. The paper formulates device selection as a contextual bandit problem, combining adaptive context partitions with hierarchical endpoint trees to balance task success against service cost while respecting capability, authorization, battery, and availability constraints. In a simulated loading-dock workload, the proposed router achieves higher step success and better cost-adjusted service than nearest-device, advertised-quality, UCB1, and LinUCB baselines, with lower simulated energy and latency.
In briefFTFM: A Factorized Tabular Foundation Model That Keeps Cell Detail While Cutting Computation
SummaryThis paper introduces Factorized Tabular Foundation Models (FTFM), which adapt to new tabular datasets in context for classification. Existing models either keep cell-level detail at high computational cost or run faster by compressing each row, losing fine-grained information. FTFM uses one contextual latent per row and one per column, and updates each cell through their interaction, reducing repeated attention. The authors report competitive predictive performance on benchmarks such as TabArena and BeyondArena.
In briefEvading Synthetic Image Detectors with Amortized Generative Attacks
SummaryThis paper systematically evaluates three transferable attacks against synthetic-image detectors: the iterative methods MIG and MuMoDIG, and the amortized generative attack Tricker. The authors introduce a two-axis protocol that independently varies the surrogate detector and the generator supplying input images, measuring effectiveness, cross-detector transferability, image quality, and runtime. They find that transferability depends on both the detector and the input-generator distribution, with MIG and MuMoDIG highly sensitive to the detector-generator pair while Tricker is more uniform.
In briefNarrow Decision Models: A System-1 Architecture for Calibrated, Typed Decisions
SummaryThis paper proposes Narrow Decision Models (NDMs), small models trained for one organization's decisions that return calibrated probabilities instead of generated text, covering choosing among options, scoring on an ordered scale, and judging yes/no statements. Their architecture is open, can run on the organization's own infrastructure, and has invariances that can be tested. The authors use court communication routing as their running case and report minimal validation on public and fictitious data, plus a first instance trained on real legal documents that agrees with a production LLM classifier on 91% of held-out documents.
Carlos Andr\'e Antunes, Aldry Albuquerque, Gabriel de Melo Schmith et al.
In briefSymbolic Constraint Agents: Automatically Learned Rules That Stop LLM Agents' Tool-Call Errors
SummaryThis paper presents Symbolic Constraint Agents (SCA), a framework that mines programmatic rules for tool calls from API documentation and refines them using execution traces. At runtime, an SMT solver checks the agent's tool calls against these constraints, issuing a warning or automatically repairing the call when the valid replacement is uniquely determined. Experiments on AppWorld and τ²-bench report higher error-detection F1, improved task success, and fewer failed calls across several models.
In briefStreamAhead: Real-time streaming video Q&A using adaptive precomputation
SummaryStreamAhead is a system for answering questions about streaming video in real time, where questions can arrive at any moment. It adaptively decides how much computation to finish for each video segment before a query arrives, completing key-value caches only for segments retrieved to answer a question. The authors report accuracy competitive with state-of-the-art methods, up to 10.4x lower time-to-first-token, and sustained processing at 30 FPS with P95 TTFT under 0.5 s.
In briefEditing as Language: Teaching LLM Agents a Native Grammar for Editing Their Own Context
SummaryThis paper proposes Native Context Editing (NCE), a framework that teaches language model agents a grammar of special tokens to edit their own context through language. The method is trained on iteratively built trajectory segments so that context editing becomes an internal part of generation. On four long-horizon agentic benchmarks, NCE raises average task success rate over the base models and reduces average token usage.
In briefEmotiQ: A quantum residual specialist framework for text emotion classification
SummaryThe paper proposes EmotiQ, a framework in which a frozen transformer classifier makes primary emotion decisions, while specialist modules, including quantum encoding models, intervene only on difficult emotion boundaries identified from validation behavior. Specialists are admitted only if they meet validation-based criteria. Across six text emotion datasets, the admitted policy raises mean macro-F1 over the baseline and improves five of six datasets. ZZ interaction feature maps were the most consistently effective quantum component.
In briefBeyond the Record: Evaluating Probabilistic AI Weather Forecasts in Extreme Tails
SummaryThis paper compares four advanced AI weather models with the physics-based IFS-ENS forecast over land in 2022, at 1.5-degree resolution and 2 to 10 day lead times. It uses 1979–2021 ERA5 monthly records as historical baselines and measures how far 2022 records exceed local variability. The AI models outperform IFS-ENS for heat and cold records at nearly every lead and depth, with heat advantages growing for deeper records, while wind advantages shrink with depth and lead time.
Bastiaan Van Velthoven, Cas Decancq, Diego G. Miralles et al.
