In briefLearn Now, Use Next, Trust Later: Step-Level Test-Time Learning for LLM Agents
SummaryThis paper addresses how LLM agents can adapt during deployment by turning new observations into timely guidance rather than waiting for episodes to finish. It introduces StepLearn, a nonparametric framework that uses transition-level hypotheses immediately for the next step, while requiring prospective validation against later observations before cross-episode reuse. Over five rounds on WebArena-Lite and ALFWorld, StepLearn outperforms the strongest baseline, EvoTest, by 2.2-12.7 percentage points across four settings, with all model parameters kept fixed.
In briefAPOLO: Automatic Prompt Optimization for Learning Ontologies from Text
SummaryThis paper presents APOLO, a method that treats ontology learning from text as an explicit prompt optimization problem for large language model modules. A multi-agent system generates text-ontology training pairs from existing expert-curated ontologies, and two learner architectures, greedy and autoregressive, are optimized with GEPA, a DSPy-based evolutionary prompt optimizer. Experiments on the biomedical DOID and plant PO ontologies show improvements after optimization across nearly all model and mode combinations, with autoregressive learners gaining the most.
In briefTSGate: Timestep-aware gated attention to improve diffusion transformer generation
SummaryDiffusion Transformers are a leading architecture for high-fidelity image and video generation. The paper finds that generation quality drops sharply for out-of-domain prompts, such as free-form user descriptions, and links this to attention sinks and weaker early-step image-to-text attention. The authors propose TSGate, which adds a timestep-conditioned bias to the gate signal so gating adapts across denoising steps. Experiments report that TSGate outperforms the baseline and standard gated attention across several benchmarks, with a 9.5% DPG score gain over the baseline on raw prompts.
In briefShallow Queries, Mature Values: Depth-Asynchronous Self-Speculation for Looped Transformers
SummaryLooped Transformers reuse one block across recurrent depths, which makes autoregressive decoding costly. The paper finds that queries and keys reach their final-depth form earlier than values, and proposes Depth-Asynchronous Self-Speculation (DAS), letting shallow drafts read full-depth prefix values, with a DAS-Wave variant. Across four recurrent checkpoints and math and code workloads, DAS-Wave reports 4.00 to 6.96 times mean throughput speedup over full-depth autoregressive decoding in the same inference stack.
In briefTesting whether science AI agents stay reliable when multi-step problems go wrong
SummaryThis paper introduces SciARP, a benchmark that tests LLM-based scientific agents on multi-turn problem solving with scientifically plausible perturbations. It converts 620 scientific problems into 3 to 13 interdependent turns and defines 13 perturbation types. Experiments on eight LLMs from four model families find that current agents remain insufficiently robust, with failures often going undetected, propagating, and resisting recovery.
In briefPapers Without Code: Checking Whether GitHub Repositories Linked in *CL Papers Still Work
SummaryThis squib examines whether GitHub repositories linked in papers from the CL journal and ACL venues over the past decade remain available. Contrary to expectations, repositories linked in recent ACL papers are unavailable at similar rates to older ones, partly due to more empty and placeholder repositories. Similar trends appear at other *CL venues, but not on platforms other than GitHub.
In briefAdapting open-vocabulary segmentation to specialized domains using prompt disagreement as preference signals
SummaryThis paper addresses the drop in performance of open-vocabulary semantic segmentation models in specialized fields such as medical imaging, remote sensing, and industrial inspection. It proposes a preference-guided adaptation framework that replaces dense pixel-level masks with binary preferences, using the disagreement between different prompt templates as a source of supervision. The method mines localized preference queries from uncertain regions and adapts models with Region-Localized Preference Optimization together with consistency regularization. The authors report consistent gains across diverse backbones on the MESS benchmark without pixel-level annotations, and effectiveness under noisy preferences.
In briefRRG-SLAM: Real-time reflection-aware Gaussian SLAM for indoor scenes
SummaryThis paper presents the first real-time Gaussian SLAM system that handles reflections in indoor scenes. It uses a hybrid of a TSDF volume and Gaussians to separate diffuse appearance from reflections on detected planes. Rendering proceeds in three passes, combining a base image with a reflection image. Experiments reported in the paper show gains over existing SLAM systems in reconstruction quality, tracking robustness, and novel-view rendering, with real-time performance preserved.
In briefPairPref: When Memory Should Guide Answers, a Benchmark for Context-Dependent Preference Use
SummaryThis paper introduces PairPref, a benchmark testing whether memory-augmented assistants apply user preferences only when the situation is appropriate. Each pair changes only the situation while keeping the preference, request, and candidate replies fixed. In the selection track models choose the reply that applies the preference appropriately, and in the free-generation track they decide whether to apply it unprompted. Models often apply the preference in both situations, and in free generation both replies were appropriate in only 3.6% to 18.3% of pairs.
In briefSAGE: Aligning Subspaces for Classifier-Free Guidance in MoE Diffusion Models
SummaryThe paper reports a failure mode where, in diffusion Transformers with Mixture-of-Experts routing, the conditional and unconditional Classifier-Free Guidance branches activate different subspaces, so high guidance scales cause collapse. The authors propose SAGE, a training-time regularizer that aligns unconditional MoE activations to the conditional subspace while preserving routing diversity, with zero inference cost. The paper reports that SAGE suppresses extreme drift by 9.2x in toy experiments and gives a 9.3% peak DPG-Bench gain on a 1B-parameter text-to-image model.
In briefEvidence Before Accuracy: MRI-PET Fusion for Alzheimer's Classification
SummaryThis paper proposes a deep network that fuses T1 MRI and FDG PET to classify Alzheimer disease, trained on 554 paired subjects from ADNI. The model reaches an AUC of 0.962, accuracy of 0.909, and F1 of 0.891. The authors test whether predictions rely on AD-relevant brain regions using shortcut controls, region ablations, and attribution methods. They report that hippocampal evidence is carried by MRI and posterior cingulate evidence by PET.
In briefEOPSA: An efficient self-distillation method that makes AI models safer while preserving reasoning
SummaryThis paper examines on-policy self-distillation for AI safety alignment and identifies two inefficiencies: teacher supervision weakens over long generations, and safety-irrelevant stylistic differences dilute the safety signal. The authors propose EOPSA, which adaptively bounds generation length using a Teacher Rescue Rate metric and distills only safety-pivotal tokens. Evaluations on reasoning models up to 32B parameters report about 50% less rollout computation and backpropagation through roughly 2% of tokens, with better safety compliance and reasoning retention than full-token baselines.
In briefXFlow: An Instruction-Guided Workflow for Lesion Segmentation in Chest X-rays
SummaryThis paper presents XFlow, a workflow model for instruction-guided lesion segmentation in chest X-rays. It mirrors how radiologists work from coarse to fine, first localizing lungs and lesion boxes, then refining contours with multi-turn point prompts. Experiments report the best segmentation quality on internal and external evaluation, and better results than ROSALIA when trained on the same annotations.
In briefHALO: Better time series generation using hyperspherical latents and masked autoregression
SummaryMost time series generators first discretize latent representations and then predict them autoregressively, which can lose information and accumulate errors. HALO instead generates in a continuous latent space, using a hyperspherical VAE that keeps latents on a fixed-radius shell to stabilize their values. It then uses a masked autoregressive model to balance parallel decoding with temporal correlation learning. The authors report state-of-the-art generation performance with substantially better inference efficiency than advanced baselines.
In briefLow-confidence remasking limits flexibility in diffusion LLMs, and a fix unlocks diverse outputs.
SummaryThis paper examines why masked diffusion language models lose output diversity when generating in arbitrary order, attributing the loss mainly to the widely used low-confidence remasking (LCR) decoding rule rather than to arbitrary-order generation itself. The authors propose top-probability position selection (TPP) as a replacement, which restores diversity and Pass@k comparable to left-to-right decoding. They further introduce Entropy-Guided Initialization (EGI), which improves rollout diversity and solution coverage beyond left-to-right decoding.
In briefKiT: A Diffusion Transformer Foundation Model for Financial Candlestick Forecasting
SummaryThis paper introduces KiT, a foundation model built on diffusion transformers for forecasting financial candlestick (K-line) data. It reframes future prediction as conditional path generation using flow matching, producing an ensemble of plausible future OHLCV trajectories from a historical window. The model is pre-trained on billions of candlestick bars across three markets and seven timescales. The authors report a mean return RankIC of 0.057 and a mean volatility RankIC of 0.66, outperforming both task-specific financial forecasters and general time-series foundation models.
In briefDoes model uncertainty match human ambiguity? A check using multi-annotator vision benchmarks
SummaryThe paper tests whether AI models find the same image cases hard that humans do, using two vision datasets (FER+ and CIFAR-10H) where each image has several human labels. Eight pretrained models across three architectures were evaluated, and their uncertainty showed only weak alignment with human disagreement. For 50.4% of CIFAR-10H and 33.5% of FER+ images, humans gave multiple valid labels while the models converged on one. The authors conclude that model uncertainty should not be treated as a trustworthy default signal for high-stakes decisions.
In briefTraining-Free Task Routing for Class-Incremental Learning via Feature Distribution Calibration
SummaryThis paper studies how a continual learner can route inputs to the correct task-specific modules when task identity is unknown at inference time, without retraining or adding a separate router. The authors identify feature-, task-, and class-level misalignment as sources of routing error and propose Feature Distribution Calibration (FDC), which combines Task Subspace Filtering, Residual Likelihood Calibration, and Prototype Affinity Calibration. Experiments on eight parameter-efficient methods across five benchmarks report an average final-accuracy gain of 4.39 percentage points.
In briefSubjectAnchor: Keeping Characters Consistent Across Shots in AI-Generated Video
SummarySubjectAnchor is a method for generating multi-shot video stories, where each new shot is created using visual memories extracted from earlier shots. It builds on the Wan2.2-I2V-A14B model and includes subject-related memory construction, subject-aware temporal rotary position encoding, and memory-aware attention partition. According to the paper's experiments, the approach improves identity consistency of subjects across shots compared with representative memory-based and holistic baselines, while keeping visual quality competitive.
In briefCan attack difficulty be estimated before optimization? A study of pre-optimization difficulty in person-vanishing attacks
SummaryThis paper studies adversarial attacks on object detectors and asks whether the difficulty of attacking each input image can be estimated before optimization starts. The authors propose Quad-CLEVER, a geometry-based estimator built from a quadratic approximation of the local person-vanishing margin, and show it correlates with observed optimization cost. They also introduce a difficulty-aware framework that allocates optimization budgets accordingly, improving attack success or reducing iterations on BDD100K and EventPed.
In briefScalable GNN Knowledge Graph Learning via Efficient Message Passing
SummaryGraph neural networks perform strongly on knowledge graph tasks such as link prediction, but their message passing is computationally expensive, especially on large graphs. Existing fixes like subgraph sampling are task- and model-specific and do not reliably guarantee lossless runtime or space savings. The authors extend Relational Sparse Matrix Multiplication to more expressive functions, including 2x2 block-diagonal multiplication, Givens rotation, and circular correlation. The paper reports task-independent runtime and space reductions for current GNNs, with near state-of-the-art performance on challenging tasks at a fraction of the cost.
Huu Tan Mai, Cuong Xuan Chu, Heiko Paulheim et al.
In briefQAMM: Adjoint MeanFlow Matching for Few-Step Offline Reinforcement Learning
SummaryThis paper proposes QAMM, which uses the critic's action gradient as a training signal for MeanFlow's average velocity. The resulting flow policy learns finite-interval transport directly and generates actions with few network calls. On ten HumanoidMaze tasks, QAMM yields effective two-call policies that perform competitively with strong flow-policy baselines.
In briefMASTraceBench: Tracking how proposals in AI agent teams improve or fail to improve collaboration
SummaryThis paper introduces MASTraceBench, a benchmark that diagnoses how collaboration gains arise in LLM-based multi-agent systems by tracking agents' proposal trajectories across six tasks. The authors find that final answers rarely surpass the strongest initial proposal, and strong initial proposals are seldom improved and may regress. They propose CLEARS, which evaluates claims at the claim level instead of exchanging whole proposals, and report that it achieves the highest Collaboration Gain on five of the six tasks.
In briefConCAD: Constraint-Aware Image-to-CAD Generation with Rewards at Two Levels
SummaryConCAD is a framework that generates executable parametric CAD programs from an image of a reference object. It is optimized with GRPO using a code-level constraint reward and an execution-level geometric reward. The authors also introduce G-CSR, a metric that checks geometric constraints extracted from B-rep models to assess design intent. On DeepCAD and Zero2CAD, ConCAD reports the best IoU and Chamfer Distance among compared baselines and also scores higher on G-CSR.
In briefLLN: Learnable Lens Networks for Efficient Long-Range Dynamics Prediction
SummaryThe paper introduces Learnable Lens Networks (LLN), which replace standard residual feature addition with learnable optical transport in an augmented position-angle phase space. Each layer alternates free propagation with a learnable lens field that transforms and focuses trajectories. The authors prove the transport is globally invertible and volume-preserving, with coordinate-wise Gaussian transport also symplectic, while retaining universal approximation. Experiments on diverse dynamical systems report better long-horizon prediction with substantially fewer parameters than same-depth baselines.
In briefPowerBench: A Benchmark for AI Agents Retrieving and Reasoning over Power System Data
SummaryThis paper introduces PowerBench, a benchmark for testing how well large language model agents can autonomously retrieve evidence and reason over interconnected power system data. Because real operational data are confidential, the authors built a generation framework and a synthetic dataset covering 761 devices, 13.35 million hourly telemetry records, and 24,939 documents. They created 300 questions across three task families, and the best evaluated model reached only 74.2% joint accuracy. The authors also analyze agent traces to show performance differences across evidence discovery, retrieval, tool use, reasoning, and answer submission.
In briefM3OS: A Monte Carlo graph search multi-agent LLM system for molecular optimization with traceable evidence
SummaryThe paper presents M3OS, a multi-agent large language model system that separates molecular design reasoning from optimization-state management using Monte Carlo graph search. A persistent graph records evaluated candidates, parent-child transformations, and evaluation evidence, and rewards and visit statistics guide parent selection. On three molecular optimization benchmarks, M3OS achieves higher success rates than baselines.
In briefDecision Value of Perception Compute: Escalating Perception Does Not Always Improve Decisions
SummaryThe paper studies adaptive perception, measuring how escalating an input from a cheap to an expensive perception mode changes downstream decision loss. It introduces DEEP, a benchmark that scores pre-escalation allocators against a budget-constrained oracle under selection, latency, and energy budgets. Experiments on KITTI, nuScenes, and nuPlan show that a substantial share of escalations make downstream loss worse, and perception-level gains often disagree in sign with decision value.
In briefHPMD: A Historical Persian Manuscript Dataset for Searching Words in Handwritten Pages Using Line-Level Labels
SummaryThis paper introduces HPMD, a dataset of 223 pages, 3,678 lines, and 37,631 words from historical Persian books, annotated at region, line, and text levels. The authors propose a baseline trained only on line-level labels that still returns word locations, scoring queries against character posterior matrices. On the test set, posterior-based search raises word spotting F1 from 0.487 to 0.558, outperforming a PHOC baseline given oracle word boundaries.
In briefSmaller Teachers Win With Less Data: Rethinking Teacher Size and Data Selection in Knowledge Distillation
SummaryThis paper studies how the best teacher model size in knowledge distillation depends on the amount of training data, finding that smaller teachers can outperform larger ones when data is limited. The authors split teacher supervision into relational ordering and score geometry to explain this effect. They propose DVA, a method that uses a small teacher to filter samples by difficulty and select diverse relational signals, without needing training dynamics statistics, and it outperforms training-dynamics-free baselines.
In briefFlexLoop: Depth-Elastic Looped Policies for Adaptive Test-Time Computation in RL
SummaryThis paper introduces FlexLoop, a post-training method that turns pretrained fixed-depth looped reinforcement learning policies into depth-elastic ones. The authors observe that such policies make reliable decisions mainly near their full trained depth, so deployment cost stays high even when less computation would do. FlexLoop keeps the original RL objective and uses adjacent-depth policy distillation to move decision quality from deeper to shallower recurrent steps. In experiments on 30 long-horizon goal-conditioned environments, it keeps full-depth performance, cuts average recurrent depth by up to 43%, and reaches up to 1.34x wall-clock speedup in a stress test.
In briefPACER: Availability-aware evidence routing for radiology reports with incomplete clinical context
SummaryThis paper proposes PACER, a framework for generating radiology reports when clinical evidence such as multi-view radiographs and previous reports is only partly available. It works in three stages: refining visual representations, calibrating the language model prefix to the available evidence, and generating structured clinical findings before the report. The authors report state-of-the-art clinical efficacy across all four MIMIC-RG4 settings and strong results on MIMIC-CXR, with competitive language quality.
In briefARS: An Agentic Reward System Using General VLMs to Estimate Robot Task Progress
SummaryThe paper introduces the Agentic Reward System (ARS), which estimates robot task progress from offline trajectories using general-purpose vision-language models without training a separate reward model. A subagent proposes a task-relevant event timeline that a primary agent verifies and revises before per-frame progress estimation. On a semantic-mismatch benchmark and in policy learning experiments in simulation and on a real robot, ARS better suppresses spurious progress than several baselines. The authors suggest structured inference and verification can make general VLMs more useful for robot reward modeling.
In briefCARDAMOM: A Micro-Dialect Arabic Speech Dataset for Speech Recognition
SummaryThe paper introduces CARDAMOM, a dataset of about 40 hours of transcribed YouTube Arabic speech covering 21 micro-dialects from six countries. Each segment carries micro-dialect labels, code-switching information, and perceived speaker gender. Four multilingual speech recognition systems were benchmarked; the best zero-shot system reached 43.47% WER, which fell to 35.21% after adaptation on the dataset.
Bashar Talafha, Samar M. Magdy, Aisha Alansari et al.
In briefCan an Image Determine the Answer? A Benchmark for Visual Answerability in Charts and Scenes
SummaryThis paper introduces a benchmark testing whether visual question answering models answer correctly when evidence is sufficient and abstain when it is not. It spans PlotQA charts, CLEVR rendered scenes, and GQA photographs, with original and edited images for each question. Across 72,000 responses from six model configurations, the highest complete task success rates were 57.0%, 43.5%, and 33.7%. Strong answerability decisions still coexisted with incorrect answers, showing the benchmark exposes failures that decision accuracy alone hides.
In briefSentZero: A sentence-centric vision-language pretraining method for zero-shot, multi-task chest X-ray analysis.
SummarySentZero is a vision-language pretraining framework for analyzing chest X-ray images without task-specific finetuning. It uses large language models to structure and map report sentences, expanding positive pairs, and adds a loss term to reduce false negatives in contrastive learning. It also modulates visual embeddings with sentence-conditioned residuals. The authors report improved zero-shot generalization across diverse downstream tasks and datasets, outperforming prior multi-task zero-shot methods.
In briefLearning new skills while keeping existing answers stable
SummaryThe paper introduces ATLAS, a method that uses retained-domain activation atlases to create an input-dependent rule for task fine-tuning. On Qwen3-8B it reports lower mean retained-output KL divergence than seven published baselines at shared coding requirements. Experiments across five backbones and two retained domains show coding gains with less movement in existing outputs.
In briefCausal Routing for Unlearning: Suppressing Only the Neurons That Express a Concept
SummaryThis paper proposes Causal Routing for Unlearning (CRU), a method for making a large language model forget specific content. It uses one untrained forward pass over the forget set to rank neurons by activation variation, then trains small routing modules that suppress only those neurons while the base model stays frozen. The authors report about 0.01% of the base model's parameters and compare favorably with baselines on the TOFU and RWKU benchmarks.
Bardh Prenkaj, Andrea D'Angelo, Davide Mottin et al.
In briefVL-AcneSeg: A Vision-Language Model That Segments Acne Lesions by Region
SummaryThis paper presents VL-AcneSeg, a framework that uses CLIP and text prompts to segment acne lesions across the whole face, aiming to replace subjective counting-based grading. Under a global prompt that needs no lesion-location information, it reports a Dice score of 0.5082 on an internal clinical dataset, the highest among compared methods. Lesion areas derived from the segmentation correlate with IGA scores (Pearson r = 0.719) at a level comparable to expert annotations (r = 0.658).
In briefPrecise editing and flexible referencing for interactive video worlds
SummaryEditWorld is a video world model that lets users precisely modify generated interactive worlds and use reference images flexibly. It streams editing instructions and reference images during autoregressive generation, using Gated Causal Attention and a Sparse Context mechanism. On the authors' WBench-Editing benchmark it scores 73.8 overall and 80.0 for editing, substantially outperforming existing methods on editing-related metrics.
In briefThe Last Mile Is the File: OfficeEditBench for Preservation-Aware Office Editing
SummaryThis paper introduces OfficeEditBench, a 170-task benchmark for making scoped maintenance edits to spreadsheets, presentations, and documents. Each task requires all necessary updates while leaving protected content unchanged. Across 510 archived outcomes from WorkBuddy, Doubao, and Codex, valid file delivery ranged from 92% to 100%, but no selected output fully satisfied the task contract.
In briefAlignment-Guided Flow Transformer: Improving Efficiency of Vision-Language-Action Robot Policy Learning
SummaryThe paper introduces AGFT, a framework that uses a dedicated alignment loss to enforce alignment among vision, language, and action in robot models. It also adopts a flow-matching objective to require fewer inference steps than diffusion-based policies. The authors report higher success rates and lower inference latency than state-of-the-art baselines on a benchmark, and argue that tri-modal alignment is key to robust manipulation.
In briefPhysioTRACE: Provenance-Aware Stress Tests for Physiological Foundation Models
SummaryThe paper introduces PhysioTRACE, a four-axis behavioral audit for frozen encoders that checks whether a physiological signal model relies on recording conditions (acquisition provenance) as a shortcut for diagnosis. It combines decodability scoring, a stress test that reverses the provenance-target association, removal of a training-localized provenance component, and verification against matched random projections. Experiments on EEG and ECG show the audit can return no reliance, a certified remedy, or a refused remedy.
Ayana Mussabayeva, Anuar Aimoldin, Olivier Oullier et al.
In briefWhen Does Structured Knowledge Help Neural Theorem Proving? A Study of Knowledge-Graph Augmentation for LLM Proofs in Lean 4
SummaryThe paper introduces MathAgent, which builds MathKG, a knowledge graph of 364 Mathlib theorems and definitions linked by 9,434 typed semantic edges, to augment LLM theorem provers. Using a controlled ablation across four augmentation modes and five models on miniF2F, PutnamBench, and MathOlympiadBench, the authors find that Lean specialization matters far more than augmentation. They also report that knowledge-graph context helps small models but hurts large ones, and that different augmentation modes solve different problems, motivating adaptive strategies.
WhyBuilds a knowledge graph to augment Lean provers; controlled ablations across five models show specialization dominates and augmentation effects depend on capability and problem.
In briefEscaping Local Views: Latent Concepts for Interpretable Multi-Agent Reinforcement Learning
SummaryThis paper addresses the difficulty of cooperation in multi-agent reinforcement learning, where each agent observes only part of the environment. It proposes ELV, an interpretable framework that extracts semantic concepts from local observations and encodes them into a contextual latent variable using a variational autoencoder. A dual-path attention mechanism models each agent's decisions, and an intrinsic reward based on next-concept prediction errors encourages exploration. Experiments across multiple environments show that ELV achieves competitive performance while explicitly revealing how agents reason about their decisions.
In briefUnderstanding Generalization Needs Universal Induction: A paper using algorithmic information theory to explain why modern AI generalizes.
SummaryThe paper argues that classical statistical theory relies on handcrafted inductive biases it cannot justify, so it cannot fully explain the success of general-purpose AI. Starting from No Free Lunch theorems, it relativizes Solomonoff induction to an information vantage point that favors short programs with access to existing information. The authors state there is evidence that frontier AI systems roughly approximate this idealized inference.
In briefZonoGPT: A Formal Verification Method for Large GPT Models
SummaryThis paper introduces ZonoGPT, an abstract domain for formally verifying large Transformer models whose space complexity does not grow with network depth. It uses a structured zonotope and a generator reduction mechanism to preserve correlations, with specialized transformations for Attention, LayerNorm, and GELU to maintain precision. The authors state it is the first approach to verify standard architectures, scaling to HuggingFace's GPT-2 Medium (24 blocks, 300M+ parameters) and verifying 1,339 instances across text and vision tasks.
WhyProposes a depth-independent abstract domain with fused transformations to verify GPT-2 Medium, a first at this scale, though evaluation breadth is limited.
In briefRGDT-Bench: A Benchmark for LLM Reasoning in Rule-Based Decisions and Their Justifications
SummaryThis paper studies Rule-Governed Decision Tasks, where models apply external rules to case facts and justify their decisions, as needed in policy, contract, and compliance work. The authors introduce RGDT-Bench, which covers four task tracks and eight task-probe combinations, and attribute failures to rule use, condition, evidence, and aggregation layers. They report that among correct responses, warrant incompleteness averages 40.2% across six evaluated LLMs, and that the best of seventeen existing evaluators reaches only 57.69% task-averaged AUROC. A simple reward model trained with warrant supervision achieves 69.24% task-averaged AUROC among correct answers, outperforming outcome-supervised baselines.
In briefWhen Words Fall Short: A Method Combining Verbalized Reasoning and Hidden Features to Estimate LLM Confidence
SummaryThis paper studies how to estimate how confident large language models are in their answers. The authors find that a dedicated confidence estimator can clearly outperform models' own verbalized confidence, suggesting richer confidence signals in internal representations. They propose IPoET, which alternates policy optimization with estimator updates. Experiments on several datasets and Qwen and Llama models show it beats both baseline types in-domain and performs better or comparably out-of-domain.
WhyProposes alternating policy and estimator training combining hidden features with verbalized reasoning, outperforming baselines in-domain and out-of-domain across datasets and backbones.
In briefMulti-task geospatial reasoning with vision-language models for crowdsourced disaster mapping
SummaryThis paper presents GRDisaster, a framework that uses vision-language models to understand, geolocate, and reason about crowdsourced disaster imagery. It introduces a benchmark of 26,340 images from PhotoMappers covering disasters from 2018 to 2024, linking volunteered, street-view, and satellite images. The framework adds spatial reasoning indicators for validating cross-view matches and assessing damage severity.
In briefSPACE-LoRA: Protecting Activation Subspaces to Prevent Forgetting in Continual Learning
SummaryThis paper studies catastrophic forgetting when LoRA is used to learn tasks sequentially. It notes that existing methods mainly constrain parameter updates and overlook additive interference, where a new residual adapter changes outputs along input directions important to old tasks. The proposed SPACE-LoRA suppresses the new residual branch's responses along those old-task-important activation directions. It sets each module's protection coverage adaptively, based on past-task sensitivity estimated with a Fisher sensitivity-based target.
In briefModeling whole-slide pathology images as evolving tumor microenvironment fields
SummaryThis paper proposes TMEvolve, a framework that treats whole-slide images as latent tumor microenvironment fields over patch graphs. It first forms adaptive soft tissue regions, then evolves them through intra-region diffusion and concept-guided boundary flux between regions. The evolved regions are aggregated for slide-level prediction. The authors report consistent improvements over representative MIL methods on six datasets across survival, gene expression, and subtype classification tasks.
In briefCORTEX: Learning to Share and Specialize in Dense Language Models
SummaryLarge language models are trained on mixed data from different domains, needing both shared and specialized knowledge. This paper proposes CORTEX, a framework inspired by learning dynamics that learns internal modules within dense language models, grouping trainable matrices based on domain-conditioned and cross-domain gradient similarity. Experiments with 160M, Qwen3-8B, and Qwen3-32B models show the highest synthetic-domain exact match and largest average perplexity reduction. It stays competitive on real-domain evaluations and forms identifiable modules.
In briefUnbiased gradient estimation for on-policy distillation of language models, correcting bias from top-k token selection
SummaryOn-policy distillation trains a weaker student language model to match a stronger teacher by minimizing reverse KL divergence on student-generated rollouts. Top-k variants use only the highest-probability tokens, which is cheaper than full-vocabulary estimation but discards the remaining probability mass and introduces bias. The proposed Tail-Corrected Top-k On-Policy Distillation (TT-OPD) adds the sampled token to the top-k tokens to recover the discarded mass in expectation, yielding an unbiased gradient estimator. Experiments reportedly show TT-OPD significantly outperforms the other OPD variants tested.
In brief"Livin' on a Prior": Likelihood Score Approximation for Inverse Problems
SummaryThis paper introduces Likelihood Score Approximation (LSA), a framework that keeps a pretrained unconditional generative model fixed and learns an observation-conditioned model to approximate the likelihood score from paired samples. It can be trained in score or velocity coordinates and supports both deterministic and stochastic sampling, and the prior can be swapped after training. On speech and image inverse problems, LSA works with about 0.01% of the full training data, and on ImageNet-256 it achieves competitive or better restoration quality than strong posterior-sampling baselines with up to several orders of magnitude fewer network evaluations.
Rostislav Makarov, Tal Peer, Danilo de Oliveira et al.
In briefSocial Circuits in multi-agent echo chambers: tracing how messages change a receiving AI's internal activity to select more useful messages
SummaryLanguage-model agents exchange messages to combine evidence, but these exchanges can create echo chambers that reinforce shared errors. The paper introduces Social Circuits, which compares a receiver's answers before and after a message change and restores selected activations to measure how much of the message effect they reproduce. Building on this, it proposes Circuit-Guided Deliberation (CGD), a method that learns to select useful messages using receiver activation changes. Across three models and four datasets, CGD achieves the highest or joint-highest average accuracy in the main comparisons while generating fewer tokens than multi-agent baselines.
In briefMaking LLMs Truly Forget: Deep Unlearning by Cutting Knowledge Paths
SummaryThe paper notes that an unlearned language model can still recover a forgotten fact through multi-hop reasoning over related knowledge. It proposes a deep unlearning framework that works with existing unlearning algorithms, using a graph minimum cut to sever all recovery paths while keeping unrelated knowledge. Experiments reportedly show deeper forgetting than superficial methods while preserving model utility.
In briefWhen Optimizing the Driving Score Stops Proving Real Improvement
SummaryThe paper asks whether gains in autonomous driving benchmark scores remain reliable evidence of better driving once the scores are used as optimization targets. The authors split the scoring process into execution, measurement, subscore mapping, and aggregation, and test it with controlled interventions and closed-loop comparisons. They find that substantial behavior changes receive little score response because distinctions are omitted, thresholded, or attenuated, and that optimization gains can reverse when the execution interface changes. They conclude that metric validity under optimization requires examining both what the scoring process measures and how optimized behavior is executed.
In briefMedCTA: Learning Disease Change Over Time in Medical Vision-Language Pretraining
SummaryCurrent medical vision-language pretraining mainly matches images and reports from each visit, which gives little information about how conditions change over time. The paper proposes MedCTA, which models clinical change both for individual abnormalities and for the overall patient course, using trend labels extracted from longitudinal reports by an offline LLM parser. Experiments on temporal image classification, image-text retrieval, and zero-shot classification report consistent gains over strong medical vision-language baselines.
In briefRemember by Asking: Retrieval-Induced Memory Evolution for LLM Agents
SummaryThis paper introduces RIME, a memory framework for LLM agents that builds long-term memory by using self-generated questions to retrieve focused dialogue evidence. The retrieved evidence and relevant past memories are merged into an evolving memory bank that records time and provenance. When compressed memory is insufficient at inference time, RIME retrieves the original dialogue and its local context. Experiments on LoCoMo with Qwen3-235B-A22B and GPT-5.6 Sol report the best results across three quality metrics, with substantially fewer query-time LLM tokens.
In briefADMIT: A diffusion framework for constrained multimodal treatment trajectories
SummaryThe paper presents ADMIT, a framework combining irregular multimodal representations, treatment-conditioned latent diffusion, and explicit constraints on generated states or actions to estimate clinical trajectories under different treatments. In a simulated dosing-schedule experiment, it predicted most of the tumor-volume change caused by redistributing a fixed total dose. Multimodal inputs improved hidden-state recovery and treatment-contrast error, and exposure projection reduced constraint violations, though enforcement remained incomplete.
In briefCoordinating frequency evolution in flow-matching text-to-speech
SummaryThis paper addresses incoherent frequency evolution during inference in conditional flow matching (CFM) text-to-speech models. It proposes a training-free, frequency-selective boosting strategy that uses the Discrete Wavelet Transform to modulate mel-spectrogram sub-bands during ODE integration, penalizing aggressive low-frequency growth and boosting lagging high-frequency details. Tested on Matcha-TTS, F5-TTS, and IndicF5, the method reduces the required function evaluations from 32 to 26 and improves Fréchet Audio Distance by up to 61%, without hurting mean opinion scores, speaker similarity, or intelligibility.
Isha Pandey, Varad Deshpande, Abhijat Bharadwaj et al.
In briefHUMAN-TCI: A Network for Finding Human Motions from Text Descriptions Using Torso-Centered Interaction
SummaryThis paper proposes HUMAN-TCI, a network that retrieves human motion sequences matching natural language descriptions. It uses a three-stream design that models upper-body, lower-body, and torso movements separately while explicitly capturing how they interact, with a tailored torso attention mechanism. The authors state that the method handles both single-action sentences and long descriptions with sequential or overlapping actions without relying on computationally heavy models.(ja)
In briefAgentHop: A Diagnostic Benchmark for Multi-Step Scientific Question Answering by AI Agents
SummaryThis paper introduces AgentHop, a benchmark of 1,011 multiple-choice questions paired with a seven-tool sandbox that enforces fixed limits on tokens, turns, and tool calls. It breaks a single accuracy score into four axes of agent operation: retrieval, synthesis, tool calls, and resource management. Testing 19 models, the authors find that behavior clusters by model family, and they release the full benchmark and evaluation harness.
In briefLLMs as adaptive meta-solvers: strategy-diverse reinforcement learning for large-scale industrial optimization
SummaryThis paper proposes a framework for training open-source LLMs to solve real-world, industrial-scale optimization problems. The authors show that solver-integrated reasoning, exact combinatorial algorithms, and heuristic search each have complementary strengths across problem structures and scales. They introduce Strategy-Diverse Reinforcement Learning (SDRL), which trains LLMs as adaptive meta-solvers, and report that the framework outperforms existing fine-tuned methods and frontier models, including DeepSeek-V4-Pro and GPT-5.5, on average across benchmarks and on industrial-scale tasks.
In briefQ-learning Penalized Transformer for Safe Offline Reinforcement Learning
SummaryThis paper studies safe offline reinforcement learning, where a policy must satisfy safety constraints while learning only from a fixed offline dataset. The authors propose QPT, a Transformer policy conditioned on trajectory context and target return and cost, trained with a penalty derived from learned reward and cost Q-functions. At inference, the same Q-functions enforce the cost threshold and select the highest-reward feasible action. QPT outperforms strong baselines across 38 DSRL benchmark tasks and adapts zero-shot to different constraint thresholds.
In briefZero-shot topic segmentation of spoken texts using cue phrases
SummaryThis paper introduces Cue-Grounded Segmentation (CGS), a training-free method for dividing spoken documents into coherent sections. It first finds phrases that explicitly signal a new topic and uses their sentence positions as boundaries, falling back to semantic segmentation guided by inferred document structure when cues are insufficient. The authors report that CGS outperforms existing baselines across six benchmarks and six LLM backbones, stays robust on noisy ASR transcripts, and does so at low API cost on proprietary models.
In briefPaper on how interval regret relates to dynamic regret in online learning
SummaryThis paper studies two performance metrics for online learning in changing environments: interval regret and dynamic regret. It shows that an algorithm with optimal interval regret can still fail to achieve optimal dynamic regret, for both convex and curved functions. The authors then propose a reduction that invokes interval regret minimization over an enlarged Euclidean ball with a domain-converted surrogate loss, achieving optimal dynamic regret. As a byproduct, they give the first proper and efficient optimal dynamic regret algorithm for exp-concave functions.
WhyRefutes that optimal interval regret implies optimal dynamic regret, and gives a reduction yielding the first proper efficient optimal dynamic regret algorithm for exp-concave functions.
In briefTraining coding agents to use files as memory for long tasks with reinforcement learning
SummaryLanguage-model agents often face long tasks whose histories exceed their active context window. This paper introduces Coding Agent Memory Gym (CAMG), four long-horizon reinforcement learning environments with shell access and a persistent workspace where agents can create, revise, search, and reuse files as memory. It also presents CAMG-RL, which trains one policy across all four environments using asynchronous PPO, learning file-based memory from task reward. On SWE-bench Verified and MLE-bench Lite, the 4B and 9B CAMG-RL models are competitive with Qwen3.5-35B-A3B and Qwen3.5-122B-A10B, respectively.
In briefMegaGraph: Automated hybrid parallelism for efficiently training large-scale graph Transformers
SummaryGraph Transformers overcome depth and over-smoothing limits of traditional graph neural networks, but scaling them to large graphs causes high memory use from attention and bias matrices, plus workload imbalance from graph embedding layers. This paper presents MegaGraph, the first automated hybrid parallelism framework for graph Transformer training, combining graph-aware context, heterogeneous pipeline, and hybrid data parallelism. An automatic search engine using a Profile-Model-Search workflow selects the best parallelism configuration. Evaluations show it trains large graphs where baselines run out of memory, reducing peak per-device memory by up to 77.8% and achieving up to 4.51x speedup while maintaining accuracy.
In briefBeyond End-to-End Black Box Mapping: An Intentional Agent Framework for Facial Reaction Generation
SummaryThis paper proposes the Intentional Agent, a framework for generating human-like facial reactions that routes generation through an explicit internal state rather than mapping speaker behaviour directly to listener expressions. It uses an internal dynamics model combining emotional drives with an iterative Inner Thought Flow, which can keep updating during conversational silences. On the REACT 2025 dataset it reports an FRDist of 72.39 and FRDiv of 0.5057, and in blinded human ratings the full model did not differ significantly from ground truth while outperforming two baselines.
In briefOSPD: On-policy self-distillation for keeping chatbot personas consistent
SummaryThis paper proposes OSPD, a framework where one model acts as both teacher and student under asymmetric information: the teacher sees a full character profile, while the student sees only a brief summary and generates dialogue from its own policy. The authors add role-aware divergence switching and a progressive trait masking curriculum to internalize character knowledge. Experiments on CharacterBench, CharacterEval, and SocialBench report substantially better persona consistency than supervised fine-tuning and multi-turn RL baselines, without an external teacher or reward model.
In briefPROACT-Agent: A Framework for Real-Time Runtime Oversight and Circuit-Breaking of AI Agents
SummaryThe paper argues that as LLMs become agents, safety risks shift from harmful text to irreversible environmental harm, while existing defenses are mostly retrospective. The authors propose PROACT-Agent, which synthesizes high-fidelity trajectories to train real-time guardrails, and introduce PROACT-Bench, a bilingual benchmark of 155,780 labeled states. Under complete source holdout, the trained guard reports 91.46% unsafe-class F1 and 90.63% exact-boundary detection, and on AgentDojo it reduces non-DoS targeted attack success from 20.82% to 0.40%.
In briefHow human cognitive limits may drive mathematical abstraction
SummaryThis paper argues that human cognitive constraints can drive progress in mathematics rather than only limiting problem-solving. It proposes a theory of mathematical abstraction as a resource-rational search for bottlenecks, where cognitive limits force existing knowledge to be restructured into new formalisms. Historical episodes from mathematics are used to illustrate this process. The authors also discuss implications for machine mathematical discovery.
In briefSalvaging Cached Results When Users Change Their Minds in Tool-Use Agents
SummaryThis paper studies how an agent should handle cached sub-results when a user changes their intent partway through a task, deciding whether to keep, patch, or discard each one. The authors find that salvage quality depends on how the decision role is designed rather than on model capability, with a single revision classification plus deterministic propagation performing best. They prove that policies using only a node's local view cannot be both safe and cost optimal. In three environments, the approach achieved 43% savings over restarting with 100% correctness.
In briefMASCIT: a mask-aware state space classifier for irregular time series
SummaryThis paper proposes MASCIT, a classifier for naturally irregular time series that have asynchronous observations, missing values, unequal lengths, and nonuniform sampling. The model feeds observation masks to its encoder and excludes invalid time steps from gated temporal aggregation. It was evaluated on 34 irregular time series datasets, and factorial ablations compared partial and full selectivity.
In briefDistilling Visual Reasoning into Text: A Method for Vision-Language Models to Reason Without Generating Intermediate Images
SummaryLarge vision-language models often struggle with tasks needing concepts not directly visible in the input image. The paper proposes V2T (Visual-to-Text Chain-of-Thought Distillation), which trains a teacher model with interleaved visual and textual reasoning, then distills it into a student model that does not generate intermediate visual representations at inference. The authors report that V2T outperforms the teacher and existing baselines across multimodal reasoning benchmarks and trains up to 42x faster than state-of-the-art baselines.
In briefFaster few-step diffusion previews that stay faithful to full-step outputs
SummaryDiffusion image generation is slow because users often generate many candidates before keeping one. The paper finds that poor outputs from standard 3-4 step samplers reflect optimizable inputs rather than a lack of reconstruction capacity, since tuning only the initial noise lets frozen samplers closely match full-step results. The authors learn corrections to the initial noise and denoising updates using endpoint supervision, so cheap previews can screen candidates before full-step generation. Experiments report 53-78% lower reconstruction MSE than retrained LD3 on unconditional benchmarks, with better ranking and candidate selection on SD1.5, SDXL, and FLUX.1-dev.
In briefEval4DiRec: A Unified Benchmark Framework for Evaluating Diffusion-Based Recommender Systems
SummaryDiffusion-based recommender systems are a new recommendation approach, but the field lacks a shared benchmark, making results hard to reproduce and compare. This paper introduces Eval4DiRec, the first unified, open-source evaluation framework for these systems. It supports 14 representative models across five recommendation scenarios under consistent settings. Using this framework, the authors benchmark the models and report their potential as well as key factors and challenges affecting performance.
In briefGeoCFM: Positive-Only Conditional Flow Matching for Mineral Deposit Location Sampling
SummaryThis paper addresses mineral exploration, where only known deposits (positive examples) are available and unlabeled areas are not reliable negatives. The authors model mineral targeting as learning a conditional spatial distribution of occurrence locations given geo-images, and propose GeoCFM, a conditional flow-matching model that generates mineral occurrence point sets without pseudo-negatives. On a synthetic benchmark and USGS Earth MRI data with a spatially disjoint tile split, it improves geometric agreement with observed occurrences over score-map and non-conditional baselines, and represents epistemic uncertainty through conditional sampling.
In briefSkillFocus: Evolving Agent Skills Through Capability Decomposition
SummaryThis paper studies how to improve reusable procedural guidance for LLM agents through iterative revision. SkillFocus decomposes recurring task requirements into a capability space that stays fixed during skill evolution, using it to decide what to revise and which evidence to use. Across four heterogeneous benchmarks it achieves the best held-out accuracy on all four, outperforming the strongest competing result by 5.7 points on average while using 24% fewer evolution tokens than the closest iterative baseline.
In briefHyperDAM: A Hyperspectral Video Tracker That Uses Material Cues to Track Objects
SummaryThis paper presents HyperDAM, a hyperspectral video object tracker built on DAM4SAM3 that uses spectral material information to tell apart targets with similar appearance. It adds spectrally consistent memory updates, a causal expander for occluded regions, and a new annotated dataset of 481 HOTC 2026 videos. In the organizer's private evaluation, the system ranked second in HOTC 2026 with 68.0093% AUC and 87.7703% DP@20.
In briefOrg-Agent: a framework for organizational language model agents that coordinate multi-user requests and reuse cross-user knowledge.
SummaryThe paper argues that language model agents serving organizations must coordinate requests from multiple users and use knowledge spread across their interactions. It identifies two capabilities, cross-user interaction and decision-making plus cross-user memory use, governed by organizational constraints. The authors propose Org-Agent, which decomposes tasks into atomic subtasks, builds a dependency graph, schedules them by topological sorting, and executes them with evidence-acquisition and memory tools. Experiments on MUSES-Bench and GroupMemBench show effectiveness on both capabilities, with ablations supporting the value of dependency modeling and tool use.
In briefP2P: A Method for Predicting Single-Cell Perturbation Responses Without Paired Data
SummaryThis paper introduces P2P (Perturbation-to-Perturbation), a method for predicting how a cell population transcriptionally responds to a genetic perturbation. Because Perturb-seq is destructive, control and perturbed cells are never observed as pairs, and cells within a condition are noisy and heterogeneous. P2P uses stochastic cell sets as its supervision unit, learning a reproducible population effect shared by two views drawn from the same condition. Across Adamson, Norman, Replogle K562, and Replogle RPE1 datasets, P2P achieves the lowest expression RMSE and highest Effect Pearson, DEG F1, and DEG average precision compared with GenePert, LinearPert, SLIM, Scouter, and scPILOT. On Replogle K562, Effect Pearson rises from 0.643 to 0.702 and DEG F1 from 0.067 to 0.178 relative to Scouter.
In briefVastMAT: A Large-Scale Multi-Category Benchmark for Tracking Multiple Animals
SummaryThis paper introduces VastMAT, a large benchmark for multi-animal tracking with 2,947 videos, over one million annotated frames, and 337 animal categories. It provides 3.66 million bounding boxes and 22,883 identity trajectories, evaluated under seen and unseen category protocols. Eight tracking methods were tested, with best HOTA scores of 66.37% and 52.90%. The authors also propose Center-Distance-Augmented Association (CDA), a training-free module that improves TrackTrack's HOTA by 1.58 and 1.31 points.
In briefSynIL: Synergy-Based Offline Imitation Learning from Imperfect Demonstrations
SummaryThis paper presents SynIL, a framework that automatically assesses the quality of robot demonstration data without manual labels, for offline reinforcement learning. It draws on neuroscience evidence linking motor synergy to motor skill, and uses self-supervised reward regression to produce dense, per-transition reward signals. On D4RL locomotion and multi-human Robomimic manipulation benchmarks, the synergy-derived rewards correlate strongly with ground-truth rewards, and SynIL outperforms Behavior Cloning. In sparse-reward human teleoperation, it matches or exceeds offline RL trained on true environment rewards.
In briefReciprocal Guidance: a runtime method that balances draft and verify compute so diffusion-AR self-speculative decoding runs faster under varying load.
SummaryThe paper studies speculative decoding in which a diffusion model drafts tokens and an autoregressive model verifies them. It proposes RecGuide, a runtime framework that uses the predictability between drafting and verification to adapt to serving load. At low concurrency it overlaps drafting with verification, and at high concurrency it assigns per-request draft block sizes. Experiments across concurrency levels report consistent throughput gains over vanilla self-speculation, up to 1.8x.
In briefCAR-VLA: Complexity- and Risk-Aware Reasoning for Self-Driving
SummaryThis paper proposes CAR-VLA, a driving vision-language-action model that uses both scene complexity and dynamic risk to decide how reasoning should be performed. It maps four complexity-risk categories to three modes: Fast Intuition, Slow Thinking, and Reflex Response, the last focusing on the most critical hazard. The model is trained with progressive supervised learning followed by reasoning-augmented reinforcement learning. It reports competitive results on NAVSIM v1 (91.1 PDMS), NAVSIM v2 (90.3 EPDMS), and Navhard (35.0 EPDMS).
WhyJointly using scene complexity and dynamic risk to select reasoning modes extends prior adaptive reasoning; competitive results on several NAVSIM benchmarks.
In briefLook Before You Select: Rethinking Vocabulary Sparsification in On-Policy Distillation
SummaryThis paper studies vocabulary sparsification in on-policy distillation (OPD), where a teacher corrects student-generated responses. The proposed SparseOPD method first uses full-vocabulary teacher correction to decide which corrections matter, then backpropagates only through the selected token logits to reduce memory on long sequences. Across six task and scale settings in mathematics, chemistry QA and multimodal reasoning, it reports higher task-average accuracy than Sampled Token and TopK and results comparable to or better than Full Vocabulary. It also reports 99% gradient cosine similarity on 4B mathematics and 70.5% lower backward memory in 8K full-parameter profiling.
In briefA paper on agent memory that builds context at runtime through agentic retrieval instead of in advance.
SummaryThe paper proposes Just-In-Time Agent Memory (JAM), a trainable framework that constructs query-specific context at runtime rather than building memory ahead of time. It pairs a Memorizer, which keeps full raw histories with compact navigational summaries, with a Researcher that iteratively retrieves, inspects, and integrates evidence for each request. The authors introduce Memory-Gym, a data synthesis pipeline, and train the Researcher with supervised fine-tuning followed by Hint-guided Group Relative Policy Optimization. They report stronger task performance than ahead-of-time memory systems and substantially better efficiency than prior trained agentic memory methods.
In briefRoboIRGBench: A Benchmark for Implicit Referential Grounding in Robot Action Models
SummaryVision-language-action models control robots but are usually tested on instructions that state all needed information explicitly. This paper defines implicit referential grounding and builds RoboIRG-Bench, with 40 variants from 11 tasks covering direct, reasoning-mediated, spatial, and contextual references. The tests show that models performing well on explicit instructions degrade sharply when they must recover information from context, especially for reasoning and spatial references. The gap also appears on a real Franka Research 3 robot arm.
Aernaer Akelijiang, Jiannan Li, Zhineng Chen et al.
In briefSurvey of joint and cross-modal video-audio generation and editing methods
SummaryThis survey reviews generative methods that model video and audio together, generate one modality from the other, or edit them jointly. It presents a unified formulation that treats joint generation, cross-modal generation, and joint editing as problems over a single distribution of audio-visual pairs, and compares methods along five design axes. The authors state it is the first overview to systematically classify joint audio-visual editing, organizing it into nine categories spanning 28 edit types. The paper also describes methods, datasets, and metrics for each setting and discusses open problems.
Abhinav Sharma, Sai Karthik Navuluru, Wang Wei et al.
In briefDPS: A Dual-Precision LLM Serving System Using Semi-Unified Memory
SummaryExisting LLM serving systems optimize only KV-cache memory, while model-weight memory stays fixed during execution. DPS serves the full-accuracy model under normal load and switches to a nested lower-precision variant under KV-cache pressure, reusing freed weight memory for KV-cache blocks. It is built on Semi-Unified Memory (SUM) and implemented on vLLM. The paper reports 2.1–3.3× higher sustained throughput and up to +41 pp effective pass@1 over Static FP16, while keeping FP16-class accuracy.
Xuan Truong Nguyen, Tien Son Pham, Tuan Duc Chu et al.
In briefMarathoner: Ultra-Long-Horizon Autonomous AI Agent (AI that keeps working on complex tasks for hours)
SummaryThis paper introduces Marathoner, an autonomous agent model designed to carry out tasks over ultra-long time horizons. It uses a post-training pipeline that synthesizes tasks from large GitHub release PRs, chains tasks together, applies rejection-sampling finetuning, and reinforcement learning with a Later Stage Bonus Reward. The authors report that Marathoner outperforms its base model and a strong proprietary model on five long-horizon benchmarks. They also state it can work for over 10 hours and make more than 1,000 tool calls.
In briefLRC-JEPA: A Compact World Model Separating Dynamics from Residual Visual Context
SummaryThe paper introduces LRC-JEPA, a lightweight end-to-end world model that splits information into a compact predictive latent z, used for dynamics and planning, and a residual-context embedding u that captures persistent visual details. Across four simulated control environments, it improves average planning success by 9 percentage points over a parameter-matched JEPA baseline. On the real-world Bridge-v2 dataset, its 5.5M-parameter encoder outperforms DINO-WM and V-JEPA2 encoders while enabling faster planning.
In briefABC-Align: Prediction-powered alignment that adaptively controls bias when human feedback is scarce
SummaryTraining language models often needs human preference data, which is costly to collect. ABC-Align uses abundant pseudo-label signals to reduce variance and applies a lightweight, adaptive correction based on a small human-labeled subset, with correction strength tuned automatically during training. The authors report that across experiments of increasing scale on LLM alignment with RLHF, DPO, and GRPO, it outperforms prior semi-supervised baselines.
In briefEmergence, Not Bandwidth: Physical Coupling and the Limits of Learned Multi-Agent Communication
SummaryThis paper studies how multi-agent teams with limited communication rates learn protocols and what limits learned communication. The authors test three MuJoCo arenas with each decision fixed at 2 bits, comparing engineered and reinforcement-learned senders. They find that communication value depends mainly on physical coupling between agents, and that under partial coupling an engineered protocol succeeds while a learned one performs like silence. Warm-start experiments indicate that reinforcement learning fails to discover the protocol, and cross-play shows learned protocols are mutually unintelligible.
WhyProves communication limits for rate-limited Dec-POMDPs and uses coupling experiments with 25 seeds to separate engineered from learned protocols, a well-supported contribution.
In briefPersMem: A memory method that builds personality into LLM agents' memory processing
SummaryRole-playing agents usually define personality only in a system prompt, while their memory storage and retrieval ignore it, making it hard to verify behaviour matches the intended personality. The paper proposes PersMem, which maps personality into four memory steps: affective appraisal, retention, passive affect-driven retrieval, and active goal-driven retrieval. Experiments on attachment and Big Five settings show distinguishable personality-related memory-processing patterns.
In briefFrom Static to Dynamic: Distilling Image Expertise into Video Diffusion Models On-Policy
SummaryThis paper proposes MILD, a motion-preserving framework that distills specialized image-model expertise into pretrained video diffusion models along the student's own generation trajectory. A learnable linear connector aligns heterogeneous image and video latent spaces, and image-guided corrections are constrained to retain pretrained video dynamics. An optical-flow-based motion reward is added to improve motion quality and temporal consistency. Across several image experts and video-student backbones, the method reportedly outperforms video-teacher on-policy distillation baselines.
In briefSpexis: Speculative Lookahead Scheduling for LLM Inference
SummarySpexis is a multi-GPU framework for large language model inference that improves pipeline and tensor parallelism through speculative parallelism. It runs speculation alongside normal execution, adding a new parallelism axis without increasing KV-cache memory use. Its lookahead scheduling predicts speculation quality and future memory pressure to reduce wasted speculation, cache eviction, and recomputation. Built on vLLM, it achieves speedups of up to 34% over a baseline using the optimal pipeline and tensor parallelism combination.
