In briefPersistent watermarks that keep text-to-image model ownership detectable
SummaryThis paper addresses copyright protection for text-to-image (T2I) generation models, which can be reused without authorization and modified before being served through an API. The authors propose a contrastive-style watermarking objective that explicitly makes the watermarked model behave differently from the original on trigger inputs. Experiments show that trigger data persist much better than prior methods under various downstream modifications and deliberate weakening attempts, with detection rates often approaching 100% TPR@FPR<10^-4.
In briefBeyond a Single Score: What Drives LLMs to Yield to Users Under Pushback
SummaryThis paper studies when large language models abandon correct answers or agree with users who push back, a behavior called sycophancy. Using over 100,000 graded replies from ten configurations across 200 items and 13 pressure conditions, the authors find that how costly a claim is to verify and whether a guardrail covers it matter most. With maximum reasoning enabled, concessions on deep logic puzzles fell to 0% for both models tested.
Guang Yang, Homa Hosseinmardi, Fengchen Liu et al.
In briefTurning Verification Failures into Reusable Guidance for Coding Agents
SummaryThis paper studies how expert diagnoses of formal verification failures can become reusable guidance for coding agents. The approach combines executable language definitions in the K framework with procedures for writing specifications, repairing proofs, and auditing specification adequacy. On HumanEval, a set of 164 Python tasks, a human-guided campaign reached a 164/164 success rate, measured by final AI audit Pass verdicts after two targeted repairs. The authors also built 12 reviewed pairs of clean and defective packages to test whether audits catch problems that passing proofs miss.
In briefFrame Differential Self-Distillation for Video Reasoning: Sparse-Frame Models Learn from Dense Views During Training
SummaryThis paper proposes FD-OPSD, a reinforcement learning method for improving video reasoning in multimodal language models. During training, it compares the policy's token-level preferences for the same sampled response under sparse and dense frame views, and distills the resulting frame differential signal without an external teacher or dense rollouts. Inference is unchanged. On Qwen2.5-VL-7B and Qwen3-VL-4B across six video reasoning benchmarks, it reports higher overall average performance than GRPO, T-GRPO, and Video-KTR baselines at 16, 32, and 64 frames.
In briefOptimal learning rates for spectral Barron functions with deep ReLU networks
SummaryThis paper studies how well deep ReLU neural networks approximate and learn spectral Barron functions. The authors provide new approximation bounds for these networks and establish minimax learning rates for this function class. They show that such functions can be approximated by deep ReLU networks and learned from training samples at a fast rate. This rate is minimax optimal up to logarithmic factors.
In briefSynchronous Multi-view Neural Diffusion: a method that fuses information within and across views simultaneously
SummaryThe paper argues that existing multi-view learning fuses within-view and cross-view information as separate stages, which can distort or compress information along intermediate pathways. The authors propose SynMDiff, which treats the multi-view feature space as a unified dynamical system driven by diffusion and performs intra- and inter-view fusion concurrently. To keep computation tractable, it adds an energy-based topological sampling strategy and an Ego-Net style centralized training architecture, and the authors report that it outperforms baselines by a large margin on real-world datasets.
In briefMaking Code as Policy Great Again: AI Agents That Write and Use Robot Tools
SummaryThis paper introduces URAI (Universal Robot-Agent Interface), in which a programming agent builds reusable robot tools and an execution agent selects and calls them in a feedback loop, with each call running a complete motion locally. Across five RoboDojo tasks, aggregate success rose from 18.0% to 53.0% compared with direct fingertip control. The authors also evaluate URAI on seven real-world AgileX dual-arm tasks.
In briefEvidence First, Arithmetic Second: A System Report and Failure Analysis for DocSem
SummaryThis paper reports on EVICALC, a system entered in the DocSem shared task, which reached 8.61% joint accuracy on 1,730 tasks in the official final test. The system reads a PDF, selects a passage, has a language model write an arithmetic expression, and evaluates it locally. The authors manually analyzed some failures, finding a case where OCR and block grouping merged the relevant passage with unrelated text, and an exploratory study where only 22 of 100 documents returned evidence identifiers. The authors state these findings are descriptive and do not establish the causes of the overall score.
In briefDigital twin framework predicting anatomy for online adaptive proton therapy in head and neck cancer
SummaryThis feasibility study presents an uncertainty-guided digital twin (UGDT) framework that forecasts treatment-day anatomy before treatment. It generates online adaptive proton therapy plans for ten head and neck cancer patients. The plans scored within about 1.7% of physician-approved offline replans, with spinal cord and brainstem doses remaining below tolerance.
In briefRouteRec: Routing Sequential Recommendations by Observed User Behavior
SummaryThis paper proposes RouteRec, a sequential recommender that routes computation using behavior observed in user sessions rather than treating all sessions identically. It summarizes four kinds of behavioral cues (interaction tempo, item-group focus, repetition and carryover, and popularity tendency) to choose expert groups at macro, mid, and micro levels. Across six public datasets and 18 dataset-metric combinations, it ranks first in 12 and second in three, with an average rank of 1.61 versus 4.11 for the next-best baseline.
In briefC-STRIDE: An AI digital twin predicting basin-wide flood depths from sparse stream gauges
SummaryThis paper presents C-STRIDE, an observation-driven AI digital twin that uses short records from a few stream gauges, plus terrain and rainfall, to map water depth across a river basin and extend predictions up to a day ahead. It is trained on simulations from a calibrated two-dimensional hydrodynamic model and needs no separate data-assimilation step. In the Des Plaines River basin near Chicago, it predicts over 4.2 million 30-m grid cells, reducing errors by about 40% compared with gauge records alone, and runs about 150 times faster than the hydrodynamic model.
In briefIntegral Meets Decomposition: A Signal-Level Self-Supervised Paradigm for Fusing Multi-Modal Images
SummaryThis paper addresses the lack of clear supervision in feature decomposition for multi-modal image fusion by proposing a 1D signal-level self-supervised method. The decomposition is reformulated as an integral-driven optimization problem, using the integral area between common and original signals as a constraint. The model follows a two-stage self-supervised framework and reports state-of-the-art results on representative fusion tasks.
In briefThe Invisible Language Tax: French Token Premiums in 2026 LLM Tokenizers and a French-Optimized Prototype
SummaryThe paper measures how many tokens the same content requires across seven 2026 LLM tokenizers. French needs 31% to 58% more tokens than English, and regional languages of France roughly 1.6 to 3.3 times as many. The authors also present Baracoda FR v1.2, a French-optimized tokenizer prototype that uses 11.5% fewer tokens than Tekken on French in their held-out test, though these are segmentation results only and effects on model quality and task cost remain untested.
In briefNot all solutions are equal: how similar network representations can differ in function
SummaryThis paper examines how the internal representations of neural networks relate to their functions. Using analysable two-layer linear networks and numerical simulations of non-linear networks, the authors find that representation and function can be dissociated, so representations may be similar without functional similarity, and the reverse. Robustness to input noise and the level of generalization error do not force representations to match the task. Networks robust to parameter noise, however, have limited representational flexibility and must use task-specific representations.
SummaryReasoning models often keep generating text after their answers have already stabilized. Settle learns when to stop by using answer stability in completed reasoning traces, training only the existing end-of-reasoning token while keeping other predictions close to the base model. At inference it needs only ordinary decoding. On MATH-500 with Qwen3-4B, it cuts token count by 40% for a 0.5-point accuracy drop, and beats supervised fine-tuning on the same traces shortened at their first stable answer by 6.16 points at nearly identical token counts.
In briefWhen Clipping Reverses Correction: Why Clipped Forward-KL Self-Distillation Fails
SummaryThis paper studies how pointwise clipping of the forward KL objective affects on-policy self-distillation (OPSD) for training models on their own responses. In matched runs differing only in clipping, clipped runs produced substantially more repetitions persisting to the end of responses. The authors prove the clipped objective can fail to correct the student toward the teacher and can push token probabilities away from it, which their training runs confirm.
In briefSmaller Frozen Models Generate Better Rejected Responses for Preference Distillation
SummaryThis paper studies where the rejected (negative) responses in preference distillation should come from. Across students from 7B to 72B on code generation and math reasoning, rejects from smaller frozen models at lower inference cost produced stronger students than self-generated rejects. The authors derive a finite-horizon utility bound for DPO in a linearized feature model and test three interventions: mixing rejects from different model sizes, reassigning rejects across prompts, and selecting lower-likelihood candidates under the reference policy.
In briefMeshOctave: A Method That Generates Meshes by Splitting and Rewiring Vertices Across Scales
SummaryMeshOctave is a method for generating compact meshes with explicit topology. It defines scales using dyadic spatial grid resolutions, coarsening meshes by merging vertices in the same voxel cell and restoring detail through split-and-rewire operations. A scale-conditioned masked discrete diffusion model learns these operations, and the authors report better geometric fidelity and topological validity than strong baselines, along with support for adaptive resolution refinement and mesh subdivision.
In briefSparse-WAM: Speeding Up World Action Models with Action-Guided Sparse Imagination
SummaryWorld-action models predict future robot visual states and actions together, which improves control generalization but makes inference expensive. Sparse-WAM is a training-free method that keeps only the future-frame tokens relevant to actions, guided by action-to-future attention, and uses an efficient engine called Pilot to reduce selection overhead. On LIBERO and RoboLab-120 it reaches roughly 2.0x and 1.8x inference speedups over dense inference on an NVIDIA RTX 4090. Task performance is largely preserved.
In briefSimEX: a framework letting coding agents learn in simulation, then control real robots with few physical trials
SummarySimEX is an autoresearch framework in which coding agents first run probe-and-optimize loops in simulation to build a robot toolbox. They then use a few real-robot trials to correct the simulator and repair the toolbox. The authors report that on real tasks such as towel folding, barcode scanning, and plate manipulation, agents acquired skills without demonstrations and with about 10 minutes of real-robot interaction.
Jiaheng Hu, Roberto Martin-Martin, Peter Stone et al.
In briefSCORE-LM: Compact Radar Fault Diagnosis with a Language Interface
SummaryThis paper introduces SCORE-LM, which pairs a small radar encoder called SCORE with an adapted local language model to diagnose radar hardware faults and give maintenance guidance. On eight Rad-R fault recordings excluded from capture-based training, it reports the best results among nine compared models, with 88.39% mean capture recall and 88.20% four-fault macro-F1 at ten frames. Its 39,520 radar inference coefficients are 119.7 times fewer than the strongest competitor, RadrNet-DS-CI, while recall is 15.56 points higher.
In briefControlling Object Hallucination in Vision-Language Models Without Retraining
SummaryThis paper addresses object hallucination, where large vision-language models mention objects absent from the image. It proposes CORAL, a training-free framework that uses uncertainty-aware visual data splitting and mirror statistics computed from symmetrically perturbed inputs. A data-driven threshold controls the expected fraction of false discoveries per image, and experiments on multiple benchmarks report consistent gains over state-of-the-art methods.
In briefPARK: A Training-Free Method for Accurate Block Retrieval in Sparse Attention for Video Diffusion Transformers
SummaryThis paper addresses the high cost of full attention in Diffusion Transformers used for video generation. It identifies two mismatches in existing sparse attention methods that average query and key block representations, which lead to inaccurate block retrieval. The proposed training-free method, PARK, keeps each query separate, normalizes attention per key block, and transforms keys using current queries before clustering. Experiments on HunyuanVideo and Wan report improved retrieval accuracy, preserved generation quality, and faster inference.
In briefScale-split neural operator for memory- and data-efficient 3D turbulence prediction
SummaryThis paper proposes ScaleSplit-NO, which uses two neural operators: a Parent that predicts the global coarse flow field and a Child that predicts full-resolution local patches conditioned on it, so neither works on the full-resolution field. On two high-resolution turbulence benchmarks it outperforms all baselines in accuracy and data efficiency. On JHTDB256 (256³), its normalized mean squared error is 53% lower than the strongest baseline and its training memory is 79% lower than the most memory-efficient baseline. The method is also applied to urban wind prediction in a Montreal district.
In briefFairness Beyond a Single Run: Seed Variability in Speech LLM Adaptation
SummaryThis paper shows that demographic fairness gaps in speech recognition are usually reported from a single training run and may be misleading. The authors fine-tune a speech LLM across five audio compression factors and six random seeds, holding other components fixed, and evaluate on Common Voice and Fair-Speech. On most demographic axes the random seed changes fairness metrics more than compression does, so single-run comparisons can be unreliable.
Srishti Ginjala, Eric Fosler-Lussier, Srinivasan Parthasarathy
In briefRealWorldShop: A Benchmark for Conversational Shopping Assistants in Real E-commerce
SummaryThis paper introduces RealWorldShop, a benchmark built on 3.28M real products, structured shopping sessions, a profile-grounded user simulator, and role-play evaluation. It finds that current systems give locally plausible replies but struggle with tracking state, updating constraints, and converging on grounded products, especially for ambiguous or multi-intent shopping. The authors also propose REALSHOP_AGENT, a session-control framework with explicit state management and runtime guards, which they report consistently outperforms strong baselines on the benchmark.
In briefMaking LLMs Say What They Think: Measuring and Improving Chain-of-Thought Alignment
SummaryThis paper examines whether a language model's chain-of-thought (CoT) reasoning matches the computations it performs internally. The authors propose CoT-Interpretability Alignment (CIA), a metric that compares CoT traces with internal reasoning strategies detected by interpretability tools. Across three tasks and three LLMs, alignment ranged only from 44.8% to 75.9%. Using post-training with task accuracy and parametric faithfulness as rewards, the authors report substantially better faithfulness while keeping or improving accuracy.
In briefBeyond Spatial-Domain Supervision: A Relation-Constrained Space for Multi-Modal Image Fusion
SummaryMulti-modal image fusion lacks ground-truth fused images, so existing methods use source images or gradients as surrogate supervision, which can cause pixel-level compromise or modality bias. This paper proposes a relation-constrained supervision paradigm that moves fusion supervision into a learned relation space, using a learnable adapter to align frozen DINO and CLIP features. The adapter is trained with a self-supervised contrastive ranking objective and alternating optimization with the fusion network. Experiments report significant gains across several mainstream fusion backbones.
In briefWhen Order Matters: First-Speaker Bias in Multi-Agent AI Debate and Personality-Based Mitigation
SummaryThis paper studies sequential multi-agent debate among large language models and finds a first-speaker bias, where agents who speak first disproportionately shape the final answer. As a result, a stronger model placed after weaker ones can lose much of its reasoning advantage. The authors use Big Five personality prompts and report that lowering agreeableness for the stronger agent restores its influence and improves accuracy, while extraversion has less systematic effects.</summary>
In briefTraining-free steering using uncertainty to reduce hallucination in reasoning tasks
SummaryThis paper studies hallucination in large language models, where prior work mainly used confidence estimates to detect or filter wrong outputs. The authors ask whether these signals can instead directly improve answer accuracy. They propose USteer, a training-free method that adjusts layer activations during inference using the gradient of a confidence measure, without changing model parameters or requiring extra supervision. The approach is reported to consistently reduce hallucination across a range of tasks.
In briefScalable multi-agent distribution matching using block-partitioned optimal transport
SummaryThe paper proposes a scalable framework using optimal transport to steer multiple agents so their final positions match a desired spatial distribution. To reduce computation, agents and target samples are split into spatially corresponding blocks and smaller local transport problems are solved. Under a mass-balance condition, the restricted coupling is feasible for the global problem and bounds the Wasserstein cost. Simulations are used to validate the proposed approach.
In briefTargeted Retrieval and Compact Representations: How CoT Improves Long-Context Counting
SummaryThis paper studies how large language models count scattered records in long texts, a needle-in-a-haystack counting task. Across twelve model comparison groups, models with reasoning (Thinking) counted more accurately than those without, especially at larger counts. The authors identify two mechanisms: non-reasoning models attend broadly to many needles, while reasoning models retrieve needles one by one, with more focused attention and more compact internal representations. Causal interventions suggest the reasoning models maintain and update an internal counter as they go.
In briefDetecting AI-generated videos from unseen generators using adversarial training
SummaryThis paper studies why AI-generated video detectors generalize poorly to unseen generators, finding that their representations strongly encode generator-specific cues. The authors propose FORGE, a token-aware adversarial training framework that perturbs generator-linked tokens while preserving real/fake discriminative evidence. The paper reports an average F1 of 92.62 across 29 held-out benchmark entries, outperforming eight baselines by 5.31 points and remaining robust under four post-processing attacks.
In briefRouting probes can improve without new information: an exact-null audit of uncertainty
SummaryThe paper asks whether routing signals in vision transformers, such as expert gates and attention-residual weights, carry information about model errors beyond the model's outputs. The authors redraw correctness labels from an output-only generator to build an exact null where routing is uninformative, and find that a width-matched MLP still reports gains in many evaluations. Selecting checkpoints by validation log loss instead of accuracy removes nearly all these detections, which the authors attribute to accuracy-based checkpoint selection. They also propose a conditional permutation test and argue that fitting a better probe and testing for incremental information require separate validation.
In briefLoop-free inverse RL that recovers rewards by sequential value recovery.
SummaryThis paper proposes Loop-Free Inverse Reinforcement Learning (LFIRL), a method that recovers a reward function from expert demonstrations without alternating reward learning and policy optimization. It uses diffusion policies, whose action-gradient structure encodes the optimal soft Q function, to cast reward learning as sequential value recovery in three stages. The authors report 2-3x training speedups over the fastest baselines on Maze, Franka Kitchen, Adroit Hand Pen, and Push-T, with reward recovery quality matching or surpassing state-of-the-art methods.
In briefDexPolicy: Scheduled Exploration for Guided Dexterous Robot Hand Manipulation
SummaryThis paper studies how exploration noise affects reinforcement learning for dexterous robot-hand manipulation. It proposes DexPolicy, which anneals action noise from broad to narrow over training while keeping other settings fixed. Experiments on YCB objects and a RealMan arm with an Inspire hand report higher Target success rates across three policy-optimization methods. The authors recommend judging schedules by terminal task success under intended execution conditions.
In briefReal-time chatbot detection of romance scams by stage classification
SummaryThis paper presents a chatbot that detects romance scams on social media in real time through the Line Messaging API. Drawing on seven-stage theories from the social sciences, it uses a large language model to identify which stage a scam is in. When a scam is detected, the chatbot explains the stage and offers prevention advice, also generated by a large language model.
In briefDrivingBench: Can Vision-Language Models Drive a Real Toyota Corolla?
SummaryThis paper presents DrivingBench, a benchmark in which general-purpose vision-language models drive a real Toyota Corolla using camera frames and direct steering and velocity commands on a low-speed parking lot cone course. The setup tests observation, action, monitoring, recovery, and long-horizon completion, with inference latency counted as part of the task. Among four models tested, only one finished the course, and the others did not pass 50% of it.
In briefRobust risk-sensitive RL learning from corrupted human feedback
SummaryThe paper studies online reinforcement learning from human feedback that uses conditional value-at-risk, when some preference labels are adversarially flipped. It proposes WSP-CVaR-RLHF, which pairs uncertainty-weighted reward estimation with optimistic CVaR planning. The authors give a regret bound that separates the clean statistical cost from the penalty due to corrupted labels, and extend the analysis to unknown transitions. Experiments under four attacks show lower cumulative regret than an unweighted robust baseline.
In briefSignal-routed temperature scaling: a low-capacity calibrator for small validation sets
SummaryThis paper studies how to recalibrate a classifier after training when only a few thousand held-out examples are available. The authors propose SRTS-BCE, a 10-parameter calibrator that preserves argmax predictions, routes examples into three risk groups using a correctness score built from six logit statistics, and fits one temperature per group. On fine-tuned CIFAR-100 with ViT-B/16 at the full calibration budget it lowers ECE15 from 1.65 to 0.96, and at smaller budgets it outperforms SMART+BCE on three CIFAR-100 backbones. The authors conclude that calibrator capacity is a finite-sample design choice whose preferred level shifts with the amount of calibration data.
Wenhao Liang, Liangwei Nathan Zheng, Lin Yue et al.
In briefEFormer: Time-Aligned Local Correction for Continuous EMG-Based Hand Pose Tracking
SummaryThis paper presents EFormer, a residual correction network that refines a frozen hand-tracking backbone using surface electromyography (sEMG) signals. It combines a high-rate event branch, temporally aligned local cross-attention, and causal rotary position embeddings. On the test set, EFormer reduces mean absolute error to 0.1546634 rad from 0.1745326 rad for the official baseline, an 11.38% reduction. The results indicate that temporally aligned event-feature correction can lower continuous hand-pose tracking error.
In briefAdaptive Self-Consistency: Smarter Sampling Using the Model's Answer Probabilities
SummarySelf-consistency improves LLM reasoning by sampling many reasoning paths and aggregating their final answers. This paper observes that each path's final answer comes from a softmax distribution that can be read from the model's log-probabilities, a grey-box setting. The authors frame inference as sequential mode identification with distribution-valued observations and propose the ASC-D algorithm. On MMLU-Redux, ASC-D uses 46.4% to 95.6% fewer trajectories than answer-only adaptive baselines.
In briefRelaxing the Conditional Independence Assumption in Image Segmentation
SummaryThe paper studies RankSEG, a method that optimizes Dice/IoU scores directly at inference for semantic segmentation, but which assumes conditional independence and ignores label correlations. The authors replace this with a Spatially Localized Dependence structure and use a reciprocal moment approximation with fixed-point optimization to reach O(d log d) complexity. The reported experiments show gains over argmax and the original RankSEG, especially in low-contrast or small-object scenarios.
In briefLearning What to Forget: Distributional Unlearning for LLM Representation Spaces
SummaryThe paper introduces Mamushi, a non-parametric framework for removing the influence of whole data domains, such as toxic language or topical content, from machine learning training data. It ranks forget examples with a probabilistic classifier and shows that thresholding the log-density ratio is the optimal fixed-budget selection rule, with a non-asymptotic transfer guarantee. Experiments on toxic-language and topical-domain removal report a better removal-preservation trade-off than baselines.
In briefWorld-as-Graph: a graph-based world model that makes object relations explicit in latent space
SummaryThis paper proposes World-As-Graph (WAG), an object-centric world model that adds relational inductive bias to JEPA-style predictive representation learning. It builds time-varying latent graphs from object-centric slots and updates object states by combining information from neighboring objects with historical memory, supporting autoregressive future prediction. The authors report superior performance on visual reasoning and robotic manipulation tasks.
In briefPrecipJEPA: A Method for Precipitation Nowcasting That Predicts Future Radar States and Renders Motion and Intensity Change
SummaryThe paper proposes PrecipJEPA, a precipitation nowcasting model that enriches its encoder with an auxiliary path learning from observed radar history. It predicts future radar states and renders them through motion and source-sink fields. On SEVIR and MeteoNet, it improves highest-threshold CSI by 118.6% and 35.1% over the strongest baselines and keeps the highest mean CSI across the 3-hour forecast.
In briefPrototype-guided Bilateral Alignment for Multimodal Federated Learning
SummaryThis paper addresses multimodal federated learning, where clients have different model architectures and imbalanced data modalities, conditions that many existing methods do not handle well. The proposed MFedPBA framework aligns heterogeneous feature spaces using a projection encoder trained with contrastive learning and Gromov-Wasserstein distance, and aggregates aligned logit prototypes using entropy weighting. Experiments reportedly show it significantly outperforms state-of-the-art baselines under model heterogeneity and modality imbalance.
Tianchi Liao Tianchi_Liao, Lele Fu, Sheng Huang et al.
In briefFrom Image Interpretation to Clinical Reasoning: Causal Reinforcement Learning for Physician-Context-Aware Multimodal Prediction of Cardiovascular Events
SummaryThis paper proposes a causal reinforcement learning framework that combines chest X-rays with physician-written clinical histories to opportunistically predict major adverse cardiovascular events (MACE). The framework uses a role-decoupled dual-LLM architecture, a dual-action causal reinforcement learning policy, and causal token pruning. It reports AUROCs of 0.720, 0.760, and 0.845 on an internal cohort, an emergency department cohort, and the external MIMIC dataset, outperforming unimodal baselines and state-of-the-art medical vision-language models, with higher GREEN scores and expert preference for reasoning quality.
In briefGraphForge: Training Work Agents on Real File Workspaces Using Evidence Graphs
SummaryThe paper introduces GraphForge, a framework that builds workspaces of real files from occupation-based seeds and links them into an evidence graph. Both the task statement and its grading rubrics are derived from this graph, so requirements and verification are tied to specific files. Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories improves results on GDPVal, Workspace-Bench-Lite, and SpreadsheetBench II, and rejection fine-tuning selected by these rubrics yields further gains.
In briefLearning Where to Steer: Noise-Space Geometry for Efficient Offline Multi-Objective Optimization with Generative Models
SummaryThis paper addresses offline multi-objective optimization, seeking better objective trade-offs from a fixed dataset without querying the objectives. Instead of guiding every sampling step of a diffusion model, the method steers the initial noise, estimating a few sensitive directions once per task with a Recursive Feature Machine. The authors prove the noise displacement increases the learned scalarized objective in expectation, and report the best average hypervolume rank among generative methods across 47 Off-MOO-Bench tasks.
In briefFlow Matching with Noisy Latent Structure: Beyond Exact Low-Dimensional Support
SummaryThis paper studies Flow Matching under a model where a low-dimensional latent map is perturbed by nondegenerate Gaussian noise, making the target distribution full-dimensional. The authors construct a spatially regular ReLU velocity class and derive non-asymptotic high-probability approximation and estimation bounds whose sample-size exponent depends on the latent dimension rather than the ambient dimension. They also propagate velocity error through the transport ODE to obtain Wasserstein convergence guarantees, showing that exact low-dimensional support is not necessary for latent-dimensional statistical behavior.
In briefHow far can a classifier's reliability drift when its confidence distribution stays fixed?
SummaryThe paper studies how a classifier's conditional accuracy can change even when its confidence score distribution stays exactly the same. Under covariate shifts that preserve this distribution, with reweighting limited by a chi-squared budget, the authors derive the worst-case change in reliability as a fragility profile. They estimate this profile and a lower confidence bound on ImageNet, finding the bound positive for several classifiers in both splits.
WhyProves worst-case drift equals sqrt(budget) times within-level variance on a computable interval, validated with ImageNet experiments and split-sample lower bounds.
In briefWarm-starting PDE Solvers Across Dimensions Using Any-Dimensional Machine Learning
SummaryThe paper gives mathematical conditions under which a learning-based PDE solver trained in low dimensions can be applied zero-shot to higher-dimensional problems, based on symmetries of the equation and initial data. When only the equation has these symmetries, which is common in physics, the authors propose a principled way to warm-start higher-dimensional solvers. The method is applied to the heat equation, Burgers' equation, and compressible Navier-Stokes equations. The authors report that a surrogate trained on 2D Navier-Stokes data outperforms a 3D-trained baseline on 3D test data while using 12% of the FLOPs and 20% of the data.
Wilson G. Gregory, George A. Kevrekidis, Ben Blum-Smith et al.
In briefRisk-Aware Adaptive Evaluation: Finding High-Impact Failures on a Limited Budget
SummaryThe paper proposes a risk-aware contextual Thompson Sampling policy that decides which interactive agent scenarios to run and rerun under a fixed trial budget, weighting them by pre-execution context and impact score. Tested by offline replay on 70 airline scenarios and 824 recorded trials, it recovers 86% of impact-weighted failures with only 50 trials, versus 25% for uniform allocation. The authors note that the advantage shrinks as the budget approaches the corpus size.
In briefFLOW: A Feature-Level Optimal Warping Method for Generalized Remote Physiological Measurement
SummaryThis paper proposes FLOW, an optimal transport-based framework that aims to make remote photoplethysmography (rPPG) more robust to domain shifts caused by lighting, motion, and sensor differences. It combines a temporal refinement module with a prototype-based cross-temporal optimal transport module to achieve domain-invariant feature alignment while preserving physiological rhythms. Experiments on four rPPG benchmarks reportedly show state-of-the-art cross-domain performance.
In briefAssembling Task-Specific Agent Harnesses at Test Time from Reusable Primitives
SummaryThis paper introduces STITCH, a framework that mines reusable harness primitives from failed task trajectories. At test time, it selects suitable primitives based on task information and compiles them into a task-specific agent harness without generating or repairing mechanism code. Experiments report that STITCH raises task success rates by up to 12 points over fixed harness baselines, with a test-time composition overhead of 2.7%.
In briefGeneralized Residual Closure: A Unified Learning Framework Balancing Stability and Plasticity
SummaryThis paper introduces Generalized Residual Closure (GRC), a framework that views learning as recursively closing discrepancies relevant to future prediction. It distinguishes two modes of internal learning, transformation within a representation and revision of the representation itself, and gives a criterion for when revision is necessary. In an affine model, the authors derive a necessary-and-sufficient condition for stability-plasticity compatibility and a unique solution to a constrained quadratic update problem.
In briefRemoving Context-Triggered Deception from LLMs via Contrastive Forget Sets
SummaryThe paper observes that LLMs can know the truth yet give false answers when context rewards deception, arguing this is a context-conditioned behavior rather than missing knowledge. It proposes PACT, which builds a contrastive forget set from the model's own deceptions and trains toward pressure-aware honest responses while keeping benign uses of the triggering context. On two 32B reasoning models, held-out deception drops from over 50% to under 3%, while system-prompt adherence, secret-keeping and reasoning traces remain at base-model levels.
WhyProposes contrastive unlearning targeting deceptive behavior rather than knowledge, exposing context blindness from target distillation; two 32B models show deception falling from over 50% to under 3%
In briefCellMSA: Context Modeling for Single-Cell Gene Expression Representation Learning
SummarySingle-cell transcriptomics data are high-dimensional, sparse, and affected by batch effects, while existing models usually encode cells independently or denoise only within a batch. The paper proposes CellMSA, which borrows the multiple sequence alignment idea from protein modeling to retrieve related cells from different batches and cell types as context for each target cell. The cross-cell patterns are summarized into a context-dependent gene-pair representation that guides the target-cell encoder. Pretrained on about 109 million cell observations, the authors report that CellMSA consistently outperforms existing methods across multiple benchmarks.
WhyAdapts MSA-inspired cross-cell context retrieval to single-cell models, pretrained on about 109 million cells and outperforming baselines on multiple benchmarks.
In briefAnalyzing Question Patterns and Student Engagement in Live Classroom Polls
SummaryThe study analyzes 604 real-time classroom poll questions and 340,668 responses from 2,807 learners across 47 live sessions, reading each poll against its lecture transcript. Most questions are lower-order and fall into seven instructional functions, with responses following a poll's function rather than its wording. The class majority answered correctly 88.5% of the time, but high-consensus incorrect answers could not be detected by agreement alone.
Rohit Sharma, Pavani Ayinampudi, Aditya B.M.V. et al.
In briefDAMPER: Return-Priority Gradient Control for Smoother Robot Policies
SummaryThis paper addresses highly oscillatory actions produced by actor-critic policies in continuous control. It proposes DAMPER, which combines the native actor gradient with a temporal-consistency gradient using conflict-conditioned projection and adaptive magnitude control. Experiments with TD3 and SAC on six continuous-control tasks show reduced action oscillation relative to the native agents in all 12 task-backbone pairs, with return trade-offs that depend on the task.
In briefAmortized neural posterior estimation for borrowing external data while accounting for exchangeability
SummaryThis paper studies how to supplement small current trials with external or historical cohort data, a common need in drug development. The authors pretrain one neural network on simulated pairs of current and external datasets, so it outputs an approximate posterior for the current-study target in a single forward pass. In simulations with outcome drift and joint mismatch, the method reduces absolute bias by up to about five-fold versus the best classical baseline while keeping Type I error near nominal, and takes about 8 ms per dataset, roughly 10^3 times faster than MCMC-based borrowing. In an analysis of ADNI data, where mild cognitive impairment outcomes differ across cohorts, it recovers the later cohort's risk level.
In briefCertified approximation for interpretable representer landmarks (CAIRN)
SummaryThis paper studies representer explanations, which rank training landmarks by their influence on a self-supervised representation. The authors propose CAIRN, a framework that propagates the combined error of four stacked approximations of the empirical neural tangent kernel into the ranking and yields a high-probability top-K certificate. They also use Stochastic Lanczos Quadrature to estimate effective dimension and guide the landmark budget.
Jayanta Mukherjee, Shourya Verma, Mengbo Wang et al.
In briefMEMO: A Multi-Level Entity-Aware Memory for Understanding Streaming Video
SummaryStreaming video understanding requires models to keep rich visual detail over long time spans, but existing memory methods often lose fine-grained information. MEMO splits streaming video into semantically coherent chunks and organizes each into structured memory, using lightweight global and entity-level representations as retrieval indices while storing high-resolution visual content separately. At inference, it retrieves query-relevant visual evidence for downstream reasoning. The method is training-free and plug-and-play with existing multimodal large language models, and it reports consistent gains and state-of-the-art results on StreamingBench and OVO-Bench.
In briefK2P: A method that turns a stronger model's knowledge into prompts without labeled answers
SummaryThis paper introduces K2P, a method that transfers reasoning from a stronger teacher model to a frozen student model through reusable prompts, without requiring ground-truth answers. K2P synthesizes instructions from teacher solutions, refines them using paired teacher and student responses, and uses answer agreement to guide search and selection. Across reasoning tasks and students, the authors report that K2P outperforms other label-free alternatives and remains competitive with supervised prompt optimization.
In briefFFASR: Benchmarking Far-Field Speech Recognition with Simulated Room Impulse Responses
SummaryThis paper introduces FFASR, a held-out corpus of 15,637 utterances and an open leaderboard covering nine conditions, each varying one acoustic factor. Dry speech from 15 talkers is convolved with room impulse responses from 14 furnished rooms to create far-field speech. Mean word error rate rises from 4.4% in near-field speech to 41.3% in the static low-SNR condition. On an office-lab pair, measured and simulated error rates agree closely on average.
In briefUnmerge: A Fast Way to Make AI Models Forget Specific Data
SummaryThis paper proposes Unmerge, a method for removing the influence of a chosen "forget" dataset from a trained model without full retraining. It treats unlearning as the inverse of task arithmetic, factorizing the forget component in a low-rank basis and subtracting it from the merged task vector. On class-level unlearning with ResNet-50 on CIFAR-100 and Tiny ImageNet, it outperforms baselines, and further studies report application to ViT-S/16 and Llama-3.2-3B.
In briefLearning a continuous neural representation of stochastic hybrid systems
SummaryStochastic hybrid systems combine continuous random dynamics with random resets triggered at boundary surfaces, and their probability evolution is described by a hybrid Fokker-Planck equation. The paper shows that such a system can be approximated by a stochastic differential equation with continuous sample paths in a higher-dimensional latent space. Auxiliary variables turn stochastic resets into deterministic ones, and an embedding theorem then removes explicit reset terms from the equation. A loss function matching state distributions lets a single latent SDE reproduce the system's probability evolution without mode labeling, trajectory segmentation, or event-based simulation.
In briefConsistent Plan-Act: Aligning Planner and Actor State Beliefs for Long-Horizon Agent Tasks
SummaryThe paper studies coordination failures between a planner and an actor in long-horizon agentic tasks, finding that the two agents often make contradictory claims about the same task state, which the authors call planner-actor state mismatch. They propose Consistent Plan-Act (ConPAct), which feeds detected contradictions back to both agents and fine-tunes them on curated consistent interactions. According to the paper, ConPAct improves performance across environments and model settings, raising MiniGrid success rate from 38.6% to 54.4%.
In briefVERA: A Framework for Multi-Agent Control That Checks Whether Actions Are Feasible and Credits Them Counterfactually
SummaryVERA is a centralized-training, decentralized-execution framework that separates estimating whether an action is feasible from assigning credit to actions. Each agent predicts a five-dimensional verifiable feasibility representation, supervised during training by exact action-conditioned margins. In a dynamic space-air-ground integrated network, VERA reports 55.33% success and 0.45% coverage violation, improving success over the strongest baseline by 8.74 percentage points.
In briefVOSSA: Voiceprint optimization for real-time voice conversion models
SummaryReal-time voice conversion systems usually use speaker embeddings pretrained for speaker verification, which may not suit frame-level speech generation under streaming constraints. VOSSA extracts speaker information from intermediate content encoder layers, aggregates it with attentive statistics pooling, and trains it jointly with voice conversion objectives, removing the need for a separate speaker encoder. Across six datasets, it improves F0 dynamics and vowel-discriminative acoustic cues while keeping NISQA-MOS, WER, and speaker similarity comparable. Perceptual tests indicate gains in naturalness, speaker similarity, intelligibility, and vibrancy.
Mu-Ruei Tseng, Waris Quamer, Ghady Nasrallah et al.
In briefCorrect but Costly: The Efficiency Gap in LLM Optimization Modeling
SummaryThis paper examines whether large language models can turn natural-language problem descriptions into mathematical models and solver code that are both correct and efficient. The authors build OptTips, a knowledge base of 50 expert modeling techniques, and OptDachshund, a multi-agent framework that produces EfficientOpt, a benchmark of 561 expert-reviewed tasks with paired reference implementations. Evaluating 11 LLMs, they find that for most models, generated programs that solve correctly take longer to run than expert counterparts, so they argue evaluation should cover both correctness and computational efficiency.
In briefSTRATA: A Role-Based Multi-Agent System That Learns From Past Games in Strategy Titles
SummaryThis paper presents STRATA, a hierarchical multi-agent system for the real-time strategy game Red Alert. Strategic, logistics, and tactical decisions are split among three role-specific agents, while a review agent turns game traces into validated experience cards that guide future play. In a fixed scenario, using the learned experience cards raised the observed win rate from 30% to 100%.
In briefNear-minimax-optimal convergence rates for asynchronous TD learning from Markovian data
SummaryThe paper studies the last iterate of importance-weighted temporal-difference (TD) learning run on a single trajectory from a finite Markov decision process. It proves a high-probability sup-norm error bound after a number of transitions that scales with the effective horizon and the minimum stationary probability. A three-state, two-action construction provides matching minimax lower bounds up to logarithmic factors.
WhyProves near-minimax-optimal last-iterate rates for asynchronous importance-weighted TD with matching lower bounds, substantially improving known bounds.
In briefDoes Learning Protein Folding Transfer to Broader Reasoning Skills?
SummaryThe paper builds FoldingCorpus, a question-answer dataset derived from protein structures, and Fold2Reason, a post-training method that uses discrete structural answers and continuous 3D geometry. On FoldBench, Fold2Reason reaches 2.7 to 3.5 times the structure prediction scores of Qwen3.5-9B. It also raises macro-average accuracy across 10 spatial, graph, scientific, and general reasoning benchmarks from 45.09% to 48.33%. Matched controls built from random, synthetic, and shuffled structures gave much smaller or negative gains.
In briefPredicting audio token attention before the language model runs to prune audio tokens
SummaryLarge audio language models convert a minute of speech into 750 to 1,500 tokens, all of which must be prefilled. The paper finds that the attention an audio token receives across the language model can be linearly predicted from its encoder output before the model runs. Based on this, the authors propose Triage, which cuts audio tokens before the language model and, on multiple choice tasks, corrects the prediction using attention observed at layer 2.
WhyShows audio token attention is linearly predictable from encoder outputs before the LM runs, enabling pruning; validated across many models and tasks, with impact mainly in efficient audio LLMs.
In briefTask-utility IQA: judging image quality by how well it serves embodied tasks
SummaryThis paper argues that image quality assessment for embodied imaging should judge whether an image serves a task goal, rather than regress scores that match generic human ratings. The authors propose an MLLM-based embodied IQA agent that reasons about task requirements, grounds quality evidence, and produces decision-facing assessments. A compact study on VisDrone finds that conventional perceptual metrics and generic MLLM-based IQA scores weakly predict downstream VLM task success, while a task-conditioned utility score better predicts whether the visual input supports the target task.
In briefEstimating Nonlinear Scalar Dynamics Parameters from Video: Invariants, Calibration, and Identifiability
SummaryThis paper studies whether parameters of nonlinear scalar second-order ODEs can be recovered from pixel observations in video. It shows that identifiability depends on the equation family: some parameters are uniquely identifiable, others only through invariant combinations, and some require external physical calibration. For laws at most linear in velocity, the authors derive explicit parameter relations and calibration conditions. Experiments on synthetic systems and real pendulum and free-fall videos support these predictions.
In briefSoft-mask diffusion editing for adjustable, pixel-level redrawing of images and videos
SummaryThis paper presents SoftPaint, a zero-shot sampling method that uses soft masks to set a different edit strength for each pixel, ranging from fully preserving the original content to fully re-synthesizing it. The authors design a Langevin-iteration-based sampler that respects per-pixel mask strengths and applies to both image and video diffusion models, enabling tasks such as video editing. The method is gradient-free and memory-efficient, and the paper reports smooth, pixel-level edits across multiple image and video backbones.
In briefExternalizing Reasoning to Make LLM Stock Prediction Narratives More Faithful
SummaryThis paper proposes a framework that uses large language models to explain cross-sectional stock return predictions, combining six-month temporal changes in XGBoost SHAP importance with historical regime analogs. The authors progressively supply raw SHAP sequences, deterministic temporal descriptors, and feature relations, and verify each generated claim against provenance-linked evidence. For Qwen3-32B-Instruct, evidence faithfulness rose from 0.696 to 0.996. Historical analogs did not improve automatic faithfulness scores but received higher human-rated usefulness.
In briefTalk2Agent: A Benchmark for Voice Interfaces to Text Agents
SummaryThe paper introduces Talk2Agent, a benchmark that tests how well voice interfaces pass spoken task instructions to LLM-based computer-use agents. It builds spoken versions of tasks from WildClawBench and OSWorld and compares ASR models, audio-capable LLMs, contextual biasing, and LLM-based ontology repair. The authors also propose an execution-free evaluation that measures how much task-relevant information survives the voice interface, reporting a 0.246 gain in Pearson correlation over WER/CER on 32 hours of real speech.
In briefWhen Context Changes: Why LLMs Answer With Outdated Values
SummaryThis paper studies how LLMs sometimes answer with an old value after a preference or fact has been updated, a failure called stale binding. The authors introduce CICM, a benchmark for tracking updated information in conversations and agent logs. They attribute the errors to attention drift, where old values receive more attention than the current one, and show that redirecting attention without further training corrects most of these errors.
In briefAdaOcc: Adaptive 3D Occupancy Prediction for Embodied Robots and Agents
SummaryThis paper proposes AdaOcc, a point-based method for adaptive 3D occupancy prediction in embodied scenarios. It accepts varying numbers of RGB views together with estimated depth maps or LiDAR scans, and represents occupied regions with sparse semantic points trained by progressive query learning. The computational budget can be adjusted through the number of query points and decoder layers. The authors report new state-of-the-art results on Occ-ScanNet.
In briefGeoNest: Learning to pick failure-aware neighborhoods for packing irregular shapes into a circle
SummaryThis paper addresses the two-dimensional irregular knapsack problem, which aims to maximize material use when packing polygons into a fixed circular container. The authors propose GeoNest, a large neighborhood search framework guided by a reinforcement-learning graph policy that pairs unplaced polygons with leftover pockets and identifies blocking placed polygons. On the new CircleNest-Bench benchmark of 2,391 instances, under the same total time budget, GeoNest improves mean utilization over a state-of-the-art standalone packing solver by about 0.9% on three main test sets and about 0.6% on a held-out industrial set.
In briefEfficient multi-modal driving trajectory planning with reward-guided fine-tuning (research paper)
SummaryThis paper proposes EMPlan, a method for planning trajectories in autonomous driving. It pairs sparse anchors with an offset refinement module to predict multiple trajectories efficiently. Safety is improved through reward-guided fine-tuning that uses rule-based rewards and unpaired preference supervision. The authors evaluate it on the non-reactive NAVSIM benchmark and report a favorable balance between accuracy and efficiency.
In briefTRACE: Target-aware retrieval and verifiable extraction for literature QA
SummaryThis paper presents TRACE, a system for answering questions about academic papers with verifiable answers and precise evidence. It combines target-grouped retrieval, independent evidence localization, table extraction, and fail-closed validation across 27,487 indexed papers. On the official 71-question test set it scores 0.760613 overall, with 0.5423 table-row F1 and 0.9728 paper F1. Remaining errors mainly involve locators, observation units, and row keys.
In briefOptimal Design for Active Preference Learning with Biased LLM Judges: a method for choosing comparisons that better match human preferences despite LLM judge bias
SummaryThis paper studies active preference learning for LLM alignment, where biased LLM judges supplement costly human annotations. The authors propose Nuisance-Adjusted Optimal Design (NAOD), which adjusts for judge deviations when selecting comparisons, prioritizing those informative about the target human population, and optimizes it with the Frank-Wolfe algorithm. Experiments on Chatbot Arena data across 17 judges, 15 budget configurations and 15 random cluster-level splits show NAOD reduces proxy policy mean regret by 29.1% relative to a matched target-information design and improves held-out human-preference prediction.
In briefSeparating spherical reasoning from dense prediction for 360° depth estimation
SummaryThis paper addresses the spatial distortion of equirectangular panoramas by reasoning over quasi-uniform Fibonacci nodes on the sphere. The proposed Fibonacci Spherical Graph captures local and long-range dependencies, and a Spherical Context Conditioning module guides pixel-aligned depth prediction. The authors report superior depth accuracy over existing approaches on three benchmarks.
In briefOnline evolution strategy that refines flow-matching robot policies from their own trajectories
SummaryThe paper proposes Online-ES, an online adaptation framework for flow-matching vision-language-action (VLA) robot policies. It applies evolution strategy in the action trajectory space, perturbing sampled trajectories, evaluating their execution outcomes, and transferring the update direction to model parameters via a self-supervised MSE objective. According to the paper, experiments in simulation and the real world show policy improvement comparable to reinforcement fine-tuning, without learning a value model or computing advantages.
In briefReducing Unnecessary Response Length in RL Post-Training via Online Distillation
SummaryThis paper studies how reinforcement-learning post-training can make responses to already-solved problems unnecessarily long, which it defines as the length-scaling tax (LST). The authors propose Length Self-Distillation (LSD), which applies on-policy distillation to solved prompts and keeps the original RL objective for unsolved ones, using an exponential moving average of the online policy as teacher. According to the paper, LSD reduces LST from 19.0% to -3.7% on single-turn reasoning and from 31.4% to 13.7% on multi-turn agentic tasks.
In briefVisualizing How Well Diffusion Models Cover Possible Outputs
SummaryDiffusion distillation speeds up image and video generation with few-step models, which are often assumed to match their slower teachers. This paper uses pass@k, the chance that at least one of k samples meets a quality standard, to test whether distilled models keep the teacher's coverage beyond single-sample scores. It finds that the training objective used in distillation determines whether this coverage is preserved, across both image and video generation.
In briefDiagnosing where and why multimodal LLMs fail via causal task decomposition
SummaryThis paper proposes a causal decomposition framework to tell whether an MLLM's failure on a compositional task stems from a deficit in the targeted capability or from errors in upstream prerequisites. The authors introduce CADET, a diagnostic benchmark of 10 composite tasks, 46 unit tasks, and over 33,000 human-annotated questions. Supplying correct prerequisites eliminates 54% of errors on cognitive tasks, and a single most important prerequisite captures 84% of the total gain.
In briefOpenJev-RLCD: A Working Implementation of RLCD
SummaryThis paper presents a working implementation of reinforcement learning for calibrated decisions (RLCD) for reasoning models. The model samples a rationale, then is scored on the answer probabilities it gives afterward using a strictly proper scoring rule, aiming to make its probabilities well calibrated. With Qwen3-1.7B on two reasoning tasks, RLCD matches or beats SFT, RFT/STaR and GRPO in accuracy and outperforms them in selective prediction. Code and results are publicly available.
In briefHigher Scores, Worse Answers: Protocol-Level Rubrics to Curb Reward Hacking in Rubric-Based RL
SummaryThis paper shows that in rubric-based reinforcement learning, where no verifier exists, summing weighted criterion verdicts lets a model offset a missed key decision with unrequested advice, raising scores while worsening answers. On clinical consultation, rubric coverage rose but appropriateness fell below the untrained model. The authors propose Protocol-level Rubrics (ProRubric), which groups criteria into a few dimensions that count only when all their criteria hold and no failure clause fires, improving appropriateness by 10.8 points and achieving the best seven-benchmark average at both scales.
WhyIdentifies additive rubric aggregation as the cause of reward hacking and proposes grouped protocol-level dimensions, validated on clinical and seven benchmarks.
In briefLearn-Then-Differentiate Gradient Estimation: a unified framework for estimating gradients by fitting a model to simulation outputs and differentiating it.
SummaryThe paper develops a unified framework explaining what learn-then-differentiate (LTD) differentiates and how accurately it estimates gradients. For models with a weighted representation, LTD differentiates a learned representation of the underlying probability measure. The authors show how accuracy guarantees for fitted models carry over to gradients and higher-order derivatives, with rates approaching the standard Monte Carlo rate under suitable smoothness conditions. The framework recovers established results for kernel, local polynomial, and kernel ridge regression, and gives further guarantees for multiple kernel learning and smooth neural networks.
In briefscTrilemma: Balancing Identity, Invariance, and Fidelity in Single-Cell Representation Learning
SummaryThis paper addresses representation learning for single-cell RNA-seq, arguing that one representation must preserve biological identity and state, stay robust to nuisance context, and keep gene-level variation for expression analysis. The authors introduce scTrilemma, a latent-bottleneck VAE that routes expression-derived variation to the embedding, decoder, or prior, trained with a single reconstruction objective and no target annotations. In zero-shot evaluation across successive CZ CELLxGENE Census releases, it leads on all three demands. Code is publicly available.
In briefFrameMorrow: Choosing Past Video Frames Using Predicted Future Needs
SummaryThis paper addresses how long video generators should pick which past frames to reuse as generation grows longer. It proposes FrameMorrow, which predicts a small set of prospective tokens representing what future frames will need, then uses them to select relevant historical frames. Because it selects explicit frames, it can be plugged into diverse generators, including closed-source ones, with little extra inference cost. The authors report consistent gains in long-range consistency, visual quality, and action alignment across five benchmarks and 11 models.
In briefOptimal VC dimension of margin-based contrastive learning
SummaryThis paper studies the generalization of contrastive learning, which learns geometric representations from anchor-positive-negative triplets. For any margin α in (0,1), the authors prove the VC dimension is O(n/α²), improving the previous O(n log(n)/α²) bound. They also give a matching lower bound Ω(n/α²) for α ≥ max(n^{-1/2}, d^{-1/2}), showing the bound is optimal up to constant factors.
Dionysis Arvanitakis, Vaggos Chatziafratis, Yiyuan Luo et al.
In briefReSCENE: Server-side replay to structurally reduce forgetting in federated continual learning
SummaryThis paper proposes ReSCENE, in which each client uploads a small condensed surrogate of its local data, and the server keeps surrogates of past tasks and trains the global model on them alongside current-task surrogates. A temporal herding step compresses the accumulated surrogates into a buffer, and the authors give a theoretical analysis showing this buffer represents the original task data more closely than full accumulation. On CIFAR-10, CIFAR-100, and TinyImageNet, ReSCENE outperforms seven baselines in average accuracy by up to 31.1 points, while using as little as 0.11× client computation and up to 179× less upload.
In briefNAMOH: Sparse attention that scales parameters and context together
SummaryThe paper introduces NAMOH, a native sparse attention mechanism in which each token activates K of H attention heads. Each head keeps only its assigned tokens and performs causal attention within that subsequence. The authors report that NAMOH can outperform fully activated models with the same total parameters and enables more efficient long-context inference than smaller dense models with matched active parameters. It is said to be compatible with GQA and existing sparse attention methods.
