In briefMetaPersona: Building Synthetic Populations from Empirical Social Science Evidence
SummaryPersona-based LLM social simulations often lack a principled way to choose attributes and assign their values, which can misrepresent real population structure. This paper introduces MetaPersona-DB, a dataset of over 11,000 empirical human-subjects studies, and MetaPersona, a framework that samples synthetic populations from empirical priors linking demographics, latent attributes, and outcomes. Results across three case studies, baselines, and frontier models vary by task, performing strongly on misinformation belief and AI-tool sentiment but mixed on income redistribution.
In briefMSU Team's Solution for GenText-Forensics 2026: Multilingual Document Forgery Detection with Forensic Reports
SummaryThis paper presents Team MSU's entry to the ACM MM 2026 GenText-Forensics challenge, which asks systems to detect forged multilingual document images, locate manipulations, identify attack types, and write forensic reports. The approach combines a document tampering detector with two LoRA-adapted Qwen3-VL-32B models in a decomposed chain-of-thought pipeline, trained by distilling reasoning traces from a larger teacher model. The team reports that this method placed third in the challenge.
Kirill Koltsov, Aleksandr Gushchin, Dmitriy Vatolin et al.
In briefDecode-Latency Feedback Prefill: A Model-Free Controller and Its Limits
SummaryThe paper proposes DLFP, a model-free controller that resizes prefill chunks overlapping active decodes, using observed decode latency as feedback, to reduce interference in concurrent LLM inference. On Qwen3-0.6B on one A100, P99 inter-token latency fell about 27.7% on average, while mean P99 time to first token rose 34.8%. The gains did not generalize to Qwen3-8B, Qwen3-32B, or a two-GPU tensor-parallel setup, which the authors attribute to scheduler-call intervals being a poor proxy for completed GPU iteration time.
In briefFine-tuning diffusion language models by choosing which tokens to reveal and which to predict
SummaryThis paper studies supervised fine-tuning of discrete diffusion language models, where masking determines both visible context and prediction targets. The authors propose GoldiMask, which selects revealed context tokens by approximately maximizing a submodular objective and weights the remaining targets by their benefit from that context and learning potential. Across three backbones and three datasets, it achieves the highest average accuracy in most settings, with gains on reasoning and code generation. Ablations show both components contribute, and it reduces decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding.
Loay Mualem, Llu\'is Pastor-P\'erez, Vinh Tong et al.
In briefLeanPolish: Using Verified Edits to Shorten Lean Proofs
SummaryThe paper introduces LeanPolish, a symbolic Lean 4 pipeline that releases 33,402 accepted local edits and 65,596 failed attempts from the same states. The authors show that simple goal-independent rules can achieve perfect ranking on first-success search, so evaluation must avoid such shortcuts. A trained ranker chooses the best candidate on 70.1% of held-out states, compared with 36.9% for the strongest frozen baseline. Iterating the symbolic pass also raises miniF2F compression savings from 19.7% to 27.5%.
In briefLearning to Plan from Random Exploration: a paper on long-range planning using only experience gathered by random wandering, without reward-driven training.
SummaryThe paper asks whether experience from random exploration alone can support long-range planning without policy-improvement training. It analyzes random walks to show what temporal relations encode, and learns them with a conditional energy-based model from observation pairs, without action or reward labels. At test time a local dynamics model and the temporal model select actions that improve progress toward the goal, with experiments on maze planning from states and images.
In briefAligned training data can still cause misalignment through context confusion in language models.
SummaryThe paper identifies "context confusion," where post-training on aligned data causes language models to behave inappropriately in different contexts. It demonstrates this across gender equality, privacy, and physical safety, finding the misalignment is narrow rather than emergent. The authors report that general alignment data does not reduce it effectively, while targeted domain data or in-context examples substantially do.
Yavuz Bakman, Duygu Nur Yaldiz, Baris Askin et al.
In briefPhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos
SummaryThis paper introduces PhyProbe, an evaluator that scores how physically consistent generated videos are. It extracts features from a frozen pretrained spatio-temporal encoder and maps them to a scalar violation score using a lightweight scoring head. Training combines pairwise ranking, regression on noisy scalar labels, and anchor-based calibration. Experiments report that it outperforms prior methods on most pairwise benchmarks and correlates strongly with human judgments.
In briefLower bounds for online learning with linear oracles
SummaryThe paper proves that methods making a constant number of linear minimizations per round cannot beat the T^{3/4} regret rate of online Frank-Wolfe on general convex sets, confirming a conjecture by Weibel et al. The authors construct adversarial instances that give regret lower bounds as a function of the horizon T and the number of oracle calls b. They also provide a second construction for one-call fixed schedules.
WhyProves a conjectured T^{3/4} regret lower bound for fixed-coefficient linear-oracle learners, extending it to all deterministic learners with explicit constructions.
In briefDoc2LoRA: Decodable Representations of Scientific Ideas
SummaryThe paper represents each scientific paper as a LoRA adapter generated by a Doc-to-LoRA hypernetwork, so every point in the space corresponds to a language model that can answer natural-language questions. On American Physical Society papers, labels generated at subfield averages were closer to official field names than five baselines, and abstracts written at points between two papers shift smoothly with the mixing weight. A small invertible transform also makes the embeddings competitive for search, close to SPECTER2 and EmbeddingGemma.
Chand Sahil Mansuri, Joel Zachariah, Sadamori Kojaku
In briefA self-evolving agent harness that improves itself across multiple tasks
SummaryThe paper proposes a framework where the same frozen language model both solves tasks and then, acting as proposer, reads its run records and edits the harness code around it. Evolution draws tasks from five benchmarks across different domains, and generalization is tested on five separate out-of-distribution benchmarks never used during evolution. Starting from a 49-line seed harness, the first-stage harness improves the average score by 4.48 points in distribution and 12.64 points out of distribution.
WhyProposes one frozen model acting as both solver and optimizer, validated on in- and out-of-distribution benchmarks, a substantive advance over prior self-evolving harness methods.
In briefTALK-Dem: Benchmark for Robot Task Planning Under Dementia-Related Speech
SummaryThis paper introduces TALK-Dem, a benchmark of 4,800 instructions that tests how large language model based robot planners handle speech patterns associated with dementia. It covers five communication patterns at three intensity levels. Experiments on six open-weight models show substantial performance drops compared with ideal instructions, and the proposed CARE retrieval method improves average task success over standard prompting.
In briefCan an AI Agent Rediscover a Blaschke-Curve Invariant?
SummaryThis paper studies whether an AI agent can rediscover a mathematical invariant of generalized Blaschke curves for one fixed degree-four Blaschke product, without being told the target theorem. The agent was given numerical coordinates of six pair-lines for 80 boundary configurations and found a homogeneous cubic that predicts 480 lines from unseen parameter values with a very small recorded residual. The authors note that a deterministic polynomial degree-search baseline also recovers the cubic, so the study does not show an advantage over polynomial fitting.
In briefScene Fragmentation Hurts Vision-Language Models More Than Question Splitting
SummaryThis paper introduces Layered-VQA, a benchmark of 93 scenes and 300 questions in which each image is split into ordered RGBA layers that recompose exactly into the original. The authors evaluate eleven open-weight and two proprietary vision-language models and find that fragmenting the scene sharply reduces accuracy, while recomposing the same layers largely restores it. They also report that even oracle-selected evidence can underperform the complete scene, and that grounding degrades faster than answer accuracy as more evidence is required.
L. D. M. S. Sai Teja, Ufaq Khan, N. Siva Gopala Krishna et al.
In briefFaster diffusion language models via a discrete average generator
SummaryThis paper introduces the Discrete Average Generator, an extension of MeanFlow to continuous-time Markov chains, aimed at making few-step generation in discrete diffusion language models more efficient. The authors show that the average generator satisfies a self-consistency identity, which they use as the basis for their training objective. On Potts model simulations, the objective reduces the total variation distance of the K-step sampler by up to 67%. On OpenWebText, the method reports the lowest generative perplexity among compared methods for 8 to 64 sampling steps with a 16× speedup, and comparable performance to existing methods on ImageNet.
In briefSimulator-Refined Diffusion for Radio-Frequency Inverse Design
SummaryThis paper proposes Simulator-Refined Diffusion (SRD), a method for generating printed circuit board layouts that match target S-parameters. It pairs a low-fidelity differentiable surrogate, whose gradient suggests a perturbation direction, with a high-fidelity non-differentiable simulator that searches along that direction. According to the paper, SRD outperforms current state-of-the-art methods, with simulated S-parameters up to 21.2% closer to targets for in-distribution cases and up to 19.8% for out-of-distribution cases.
Jinhao Liang, Jacob K. Christopher, Michael Frei et al.
In briefTraining-free method helping vision-language models classify specialized images from few examples
SummaryThe paper observes that generative vision-language models such as Qwen-VL and LLaVA perform poorly on specialized domains (called distant out-of-distribution) because the needed discriminative features were never learned. The authors propose Inductive Visual Logic (IVL), a training-free framework that uses the models' remaining descriptive ability to extract visual traits from few-shot support images and organize them into per-class dictionaries. At inference, hierarchical filtering classifies images using spatially grounded trait evidence. According to the abstract, IVL achieves the highest aggregate accuracy across several distant-OOD benchmarks with two backbones, with traceable predictions.
In briefOn-Policy Distillation's Off-Policy Teacher Problem
SummaryOn-policy distillation lets a student learn from trajectories its own policy generates, with a teacher providing dense supervision. The paper points out that the teacher must continue student-generated prefixes rather than its own, and its performance degrades as these prefixes grow longer. The authors propose SCOUT, which periodically trains the teacher with reinforcement learning on verifiable rewards to continue student prefixes. Experiments show improved teacher continuation and more effective distillation across several configurations.
In briefBeyond Mode Collapse: Diverse Synthetic Expert Conversations Generated with Generative Flow Networks
SummaryThis paper proposes using Generative Flow Networks (GFlowNets) to generate diverse, high-quality synthetic conversation data for post-training large language models. The method models latent conversation structure with a Gaussian mixture over key interaction features, so expert strategies are sampled in proportion to their prevalence in the training data. In tutoring and emotional support dialogues, it reportedly balances fidelity, mode coverage, and authenticity better than reinforcement-learning and end-to-end LLM baselines, without copying training data.
Sumit Asthana, Michael Ion, Kevyn Collins Thompson
In briefTesting whether language models stay safe during long adversarial chats, not just single harmful prompts
SummaryMost safety tests check models with one harmful prompt, but real systems face long, adaptive conversations. This study has a second language model act as a persistent adversarial user across multiple turns, with a safety classifier labeling each response. Across three open-weight models and two harmful prompts, the rate of safe responses fell sharply as conversations lengthened. The authors conclude that strong single-turn safety does not necessarily persist under sustained adversarial interaction.
In briefContinual learning of dynamical systems in recurrent networks via recyclable unit gating
SummaryThis paper studies how recurrent neural networks can learn several dynamical systems in sequence without forgetting earlier ones. The authors benchmark standard continual learning methods on an Almost-Linear RNN and propose CRUG, which uses differentiable gates with an L0 penalty to assign units per task and recycles unused units for later tasks. The authors report that CRUG gives the best reconstruction-capacity trade-off with zero forgetting, and that forward transfer helps more when tasks share similar dynamics.
In briefHalluScoring 2026: first shared task on detecting LLM hallucinations and verifying answers
SummaryThis paper presents HalluScoring 2026, a shared task testing whether systems can detect hallucinations and verify facts in Arabic question answering. It has two tasks with four subtasks, using the HalluScore and HalluTruthQA Arabic datasets. Thirteen teams took part, and ten submitted system description papers. Results show hallucination detection remains difficult, with top AUC-ROC scores of 0.7717 and 0.7670 on Task 1, and top combined scores of 0.8824 and 0.8565 on the Task 2 subtasks under assisted evaluation.
Aisha Alansari, Abdessalam Bouchekif, Ahmed Hasanaath et al.
In briefTAGGRAPH: Tag-Augmented Graphs for Retrieving Persistent Agent Histories
SummaryThis paper studies long-term memory for LLM agents, which must recall past conversations, and proposes a controlled framework for comparing memory systems built on shared conversational memories. On LongMemEval-S, the best graph configuration, AdaptiveGraph, reaches 0.844 MRR, below BM25 (0.867) and OpenClaw (0.880). On ATANT Core, localized graph traversal beats diffusion and BM25. The authors conclude that retrieval strategies should be evaluated together with the memory setting and against strong lexical baselines.
In briefMILO: Automatically Discovering Better AI Agent Harnesses Through Multi-Agent Evolution
SummaryThis paper introduces MILO, a framework that co-evolves AI agent harnesses (the structures controlling model execution and environment interaction) along with the search strategy used to find them. On Terminal-Bench 2.1, PaperBench, and DeepSWE, MILO-discovered harnesses outperform eight state-of-the-art harnesses and six search methods with both frontier and open-weight models. The authors also report improved best-known bounds on several EinsteinArena open problems.
In briefFunction-Space Transformer with Adaptive Anchors: A New Framework for Learning from Continuous Functions
SummaryThe paper introduces the Function-Space Transformer (FST), which learns from functions using a spatially adaptive continuous latent representation. Features are stored at anchors whose locations are predicted from the input and refined recursively through function-space interactions. On PDEBench Burgers and Darcy flow PDE tasks, FST substantially outperforms Perceiver IO and is highly competitive with the Fourier Neural Operator. On ImageNet-1K, it achieves higher accuracy than the ViT baseline with fewer parameters.
In briefTrackFish3D: Self-supervised 3D tracking of schooling fish from multi-view video
SummaryThis paper presents TrackFish3D, a self-supervised method that tracks dense groups of schooling fish in 3D using multiple calibrated cameras, without identity labels or 3D ground truth. It is trained once on unlabeled footage and applied directly to unseen test videos. On the authors' benchmark, 3D MOTA rises from 87.7% for the strongest baseline to 95.8%, and it reaches 81.1% on the 3D-ZeF zebrafish benchmark versus 77.4% for the best geometric baseline.
In briefHow AI Tutors Handle Students Stuck on Problems
SummaryThis study analyzes 1,260 real sessions with a guided LLM chemistry tutor, covering 20,462 student turns and 6,630 impasse turns involving conceptual errors, uncertainty, or help-seeking. Three tutoring conditions were simulated to compare how the tutor guides students through impasses. Each additional impasse turn reduced the odds of next-turn recovery by 12.7%. The authors suggest impasse depth and type can serve as real-time signals for graduated assistance.
Bakhtawar Ahtisham, Kirk Vanacore, Alessandra Napoli et al.
In briefOpenCollab: A programmable, controllable multi-agent coding framework for studying agent collaboration
SummaryThis paper introduces OpenCollab, a framework for multi-agent software engineering that lets researchers program agent collaboration structures and run them on a controlled, shared runtime. The authors define Adherence, a metric for whether agents actually follow the declared organization, which varies from 47.2% to 97.2% depending on configuration. On agentic coding benchmarks, the authors report that the Duo configuration reaches new state-of-the-art results against several mainstream harnesses, while the single-agent configuration uses the fewest tokens.
In briefSemantic Inpainting for Animatable Gaussian Head Avatars (SInGA)
SummaryThis paper presents SInGA, a method that learns animatable Gaussian head avatars from a single image. It uses a semantic inpainting framework in UV space, leveraging facial symmetry cues, to complete unobserved facial regions before regressing Gaussian attributes. Multiple Gaussians are stacked per location to enhance detail. The authors report that the approach generalizes across identities without per-identity optimization and yields improved completeness, identity preservation, and consistent rendering from unobserved views.
In briefActivation-Conditioned Self-Distillation: A Reasoning Training Method Without Reference Solutions
SummaryThis paper introduces Activation-Conditioned Self-Distillation (ACSD), which extracts a steering vector by contrasting activations of self-generated trajectories that reach verified correct answers with those of the remaining trajectories. A frozen base model applies this vector at each prediction position, and the student learns from its next-token distributions on its own prefixes. Distillation needs no problem-specific reference text or teacher parameter updates, and only the student is used at inference. The authors report the highest mean accuracy over four math benchmarks on each of five models.
Zhexi Lu, Subhajit Chaudhury, Tejaswini Pedapati et al.
In briefCARAT: Do Materials LLMs Reason or Recite?
SummaryThis paper presents CARAT, a framework for testing whether materials-science LLMs reason from crystal structure or copy answers already present in their input. It holds questions and gold answers fixed across eight input views and adds fine-tuning, masking, and shortcut checks. The authors find that the frozen model often repeats a structural relation without using it, and that matched supervision and removing the relation change this behavior.
WhyMatched views, answer masking, and shortcut ablations separate reasoning from recitation, exposing confounds with baselines; methodologically novel with fairly solid evidence.
In briefE2E-SWE: Testing whether AI coding agents can build working software projects from scratch
SummaryThis paper introduces E2E-SWE, a benchmark that tests whether coding agents can build complete, installable software repositories end to end. It contains 186 whole-repository generation tasks across 11 programming languages, where agents receive only a natural-language specification and an empty workspace and must pass a hidden test suite. Evaluating 13 frontier models, pass@1 ranges from 11.7% to 67.7%, showing clear differences between models.
In briefEVOKE: Drawing Out World Knowledge in AI Agents for Transferable Decisions
SummaryLarge language model agents often perform poorly when moved to unfamiliar environments. The paper proposes EVOKE, a post-training method that holds the state and history fixed and ranks the same candidate actions under different goals, so agents must rely on world knowledge rather than superficial habits. The authors evaluate it on three backbone models across diverse tasks. They report gains in task performance, generalization to unseen environments, and data efficiency.
In briefHermes: teaching models to manage context so extra test-time compute helps
SummaryThis paper studies how a model can use extra compute at inference time across multiple context windows, specifically by deciding when to open fresh contexts and what information to carry forward. The authors introduce Hermes, a family of configurable harnesses that gradually give the model more control over context allocation, and Hermes-Learn, a two-stage training framework. They report that capable models scale with added inference compute, while smaller open-source models struggle initially, and that Hermes-Learn closes this gap. The gains are said to generalize across benchmarks and models and to transfer to other test-time scaling methods.
In briefExploreNet: Learning Where to Explore in Diffusion GRPO
SummaryThis paper studies how image generators trained with group-relative reinforcement learning, such as Flow-GRPO, explore by adding uniform Gaussian noise to every latent element at each denoising step. The authors propose ExploreNet, a policy that predicts a per-element noise scale from the latent, step, and prompt, and is discarded after training. On Stable Diffusion 3.5 Medium, it improves held-out GenEval2 by 14% over Flow-GRPO while leaving inference unchanged.
Shuyue Stella Li, Xiaochuang Han, Yulia Tsvetkov et al.
In briefAbsorbing State Phase Transitions in Multi-Agent Search
SummaryThis paper uses the absorbing state phase transition formalism from statistical mechanics to predict whether LLM-based multi-agent systems succeed at search tasks. The authors sort search tasks into four types and derive a critical communication degree d_c, the minimum number of agents each agent must communicate with to keep incorrect hypotheses from spreading. Tested on real search, software debugging, and physical mechanism discovery tasks, agreement with the theory was mixed.
In briefModal Kinetic Typography: Animating Letters Through Their Natural Vibration Modes
SummaryThe paper introduces modal kinetic typography, which animates vector glyphs using modes derived from a finite-element eigenproblem on each letter's outline and its parts. A frozen video diffusion model supervises only the amplitudes and phases of these modes, keeping motion smooth and the letter legible. Compared with Dynamic Typography and AniClipart, the authors report more articulated and smoother motion with less outline tearing, and human raters preferred the results.
In briefMulti-agent LLM discussion helps less when agents withhold dissent
SummaryThis paper examines conflicting reports on whether discussion among LLM agents improves accuracy or produces incorrect consensus. It introduces a simple model built on four behaviors: withholding dissent, internalizing a stated answer, reconsidering after seeing dissent, and correcting toward the right answer. The model predicts discussion can overturn an incorrect majority only when the withholding rate is below a critical value, and the reported experiments show gains shrink as withholding rises. Instructing agents not to withhold dissent, or turning reasoning off, increases the gain from discussion.
In briefGestalt: A Meta-Foundation Model Built from Existing Models for Astronomy
SummaryThe paper combines 22 frozen foundation models from eight families by concatenating their whitened image embeddings and compressing them with a randomised SVD into a 1024-dimensional representation. On galaxy imagery from HSC, JWST and DESI Legacy Survey, this combined embedding outperforms every individual model on 19 of 21 tested metrics for physical property and morphology estimation. The authors report that performance improves with basket size and architectural diversity, and that it transfers across surveys without additional training.
In briefUsing large Lagrangian models to find experimental signatures beyond the Standard Model
SummaryThis paper presents hAIthem, a framework combining reinforcement learning with large language models to search theory parameter spaces for regions not yet excluded by experimental constraints. In a proof-of-concept study on single dark scalar multiplet models, the reinforcement learning search found more viable regions with greater physical diversity than an evolutionary-algorithm baseline. The authors also report that the framework proposes novel combinations of previously studied observables, such as applying a halo-independent kinematic ratio to paleo-detectors.
Ibrahim Elsharkawy, Victoria Knapp-Perez, Wahid Bhimji et al.
In briefPoint2Part: A unified prompt-based method for splitting 3D shapes into parts
SummaryExisting 3D part decomposition methods may produce parts that overlap or leave gaps, which hinders later part-level uses. This paper proposes a promptable model where users specify each desired part with a 3D point, and the model outputs non-overlapping parts that together cover the whole shape. Built on a pretrained 3D generation model, it handles image-to-part generation, mesh-to-part generation, and part segmentation in one shared latent space. The authors report better part-quality results than prior methods on all three tasks.
In briefSkill-space shooting: a method that lets robots improve their own control policies after failures by reusing short learned skills.
SummaryThe paper introduces skill-space shooting, which uses foundation model guidance to explore corrective actions built from reusable short skills and converts successful trials into policy improvements. Real-world experiments show that autonomously acting policies improve repeatedly. The authors also report that skills can be shared across tasks, reducing the teaching needed to improve on new tasks.
In briefImagine3D-LLM: Teaching multimodal models to picture a 3D scene before answering
SummaryThis paper addresses the difficulty multimodal large language models have in combining multi-view images into a coherent 3D understanding. It proposes Imagine3D-LLM, which appends learnable summary tokens after image tokens, decodes them into a compact 3D Gaussian Splatting representation supervised by a photometric reconstruction loss, and trains this jointly with next-token prediction. The authors report that the model consistently outperforms prior approaches on multiple spatial reasoning and 3D understanding benchmarks.
In briefLocal Denoising Breakdown and Semantic Speciation in Generative Models
SummaryThis paper studies two time windows in generative model dynamics: a speciation window where a sample commits to a semantic class, and a nonlocality window where local context becomes insufficient for generation. Under a common-cause hypothesis, the authors prove the nonlocality window lies within the speciation window. They identify conditions under which both windows converge to a single time as system size grows, a phase transition, and verify this analytically for Gaussian mixtures.
In briefSynthetic counterfactual videos help humanoid robots learn to carry objects
SummaryThe paper introduces PRISM, a framework that turns a few real human-object interaction videos into a large, diverse set by generating counterfactual videos with video-to-video generation. A contact-anchored real-to-sim pipeline reconstructs human and object motion and converts it into physically plausible trajectories for training. A single policy trained this way is deployed on a real humanoid without real-world fine-tuning, using onboard depth, and picks up, carries, and drops boxes, barrels, bins, and balls.
In briefAdversarial Training for Pixel Diffusion: Using Adversarial Learning to Correct Missing High-Frequency Detail
SummaryThis paper studies why pixel diffusion models underrepresent fine-scale natural-image statistics. The authors add an adversarial loss to predicted outputs at non-high-noise timesteps of a pretrained model, keeping the original objective, architecture, and sampling unchanged. Across two pixel backbones, the method improves distribution fidelity, coverage, prompt alignment, and perceptual quality, and restores missing high-frequency spectral power. Under the tested latent diffusion configurations, the same procedure did not produce comparable joint gains.
In briefSTEPQuant: Where and When Quantization Errors Matter in Recurrent States
SummaryLinear attention replaces growing KV caches with fixed-size recurrent states, which can still become a major memory bottleneck when serving many requests concurrently. The paper proposes STEPQuant, a spatial-temporal post-training quantization framework for Delta-rule recurrent states that allocates precision by error magnitude and memory lifetime and fits key-row and value-column scales. Experiments on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct show accuracy close to FP32 states at a nominal 6-bit budget. Integrated into SGLang, the 6-bit version reports over 5x recurrent-state compression and up to 68.7% lower total serving memory.
In briefLeapQuant: Accurate Low-Bit Quantization for Linear Attention Recurrent States
SummaryThis paper targets hybrid LLMs that use linear attention, such as Gated DeltaNet and Kimi Delta Attention, where repeatedly reading and updating the recurrent state is a major inference bottleneck. It proposes LeapQuant, a training-free method that quantizes the state once per window of tokens and keeps the largest outliers as high-precision Compensator Tokens, with residual smoothing before quantization. Experiments on Qwen, Kimi, and GLM models report accuracy comparable to FP32 under 8-bit state quantization, with speedups on NVIDIA GPUs.
In briefCropland PAtteRNS: Parallel attention network for crop mapping from satellite image time series
SummaryThis paper introduces PAtteRNS, a hybrid transformer-convolutional model that applies self-attention separately to the temporal, spectral, and spatial dimensions of Sentinel-2 satellite image time series. A new parallel transformer design reduces the computational cost of this triple-factorised attention. Tests on multiple tile sizes of the PASTIS and MTLCC datasets, including an ablation study, report better crop segmentation and parcel boundary quality (Boundary IoU) than compared state-of-the-art models. The authors also note that flawed class groupings and incomparable results across tile sizes affect fair evaluation, and they call for standardised dataset construction practices.
Joseph Metcalfe, Sara Sharifzadeh, Fabio Caraffini
In briefRethinking Representations for World-Action Modeling: the ReWAM paper
SummaryThis paper studies how to design the representation space in world-action models, which jointly learn robot policies and predict future observations. It finds that neither reconstruction fidelity nor pre-trained perceptual features alone ensure effective policy learning. The authors propose ReWAM, a representation-centric model built on pre-trained DINO features, reporting 93.6% success on RoboTwin 2.0 without generative video pre-training.
In briefSpectral Bounds on Distortion in LLM Graph Reconstruction
SummaryThis paper examines how language models reconstruct graphs, showing that the common single-number evaluation (Wasserstein distance between Laplacian spectra) is sandwiched between two edge-count measures, each scaled by 2/n. The bounds are proven sharp, and the authors show the gap can certify mixed editing where edges are both added and removed. Experiments on 135 reconstructions from three open-weight models over 45 synthetic graphs characterize these regimes.|
In briefEmoRES-TTS: Residual-Enhanced Vector Steering for Emotional Speech
SummaryThis paper addresses the unreliable emotional expression of emotion-conditioned text-to-speech models by proposing EmoRES, a training-free vector steering method. It finds that an emotion vector splits into a shared component that moves speech away from neutral and a residual component that points toward the requested emotion, and controls the two separately. On IEMOCAP with IndexTTS-2 and CosyVoice2, EmoRES outperforms the CoCoEmo baseline on all four objective emotion metrics, and human evaluation reports gains in emotion identification and naturalness.
In briefDMA²: Pixel-space diffusion distillation with adversarial and anchor losses
SummaryThis paper proposes DMA², a method for distilling text-to-image diffusion models into few-step generators working directly in RGB pixel space rather than latent space. It combines a fixed high-noise teacher-matching band, DINO-based adversarial patch guidance, and AF-Loss, a parameter-free semantic distribution objective in the DINOv2 space. The authors report that the four-step student outperforms the 25-step teacher and other few-step distillers on DPG-Bench, GenEval, VQAScore, and COCO30K.
In briefBeyond the Timeline: Long-video memory using grounded entity biographies
SummaryThis paper introduces Grounded Entity Biographies (GEB), a memory framework for answering questions about long videos. It groups visually grounded observations of the same physical object across clips into retrievable biographies while keeping the context of each moment. Across four benchmarks, including day-long and week-long recordings, GEB outperforms prior memory frameworks in multiple-choice and open-ended question answering, reaching 72.0% accuracy on EgoLifeQA.
In briefLongLive-Plug: One-time distillation of reusable video-generation add-ons
SummaryThis paper presents LongLive-Plug, a once-for-all distillation framework that trains reusable LoRA adapters on a base video diffusion model. The adapters provide single-pass classifier-free guidance, few-step sampling, and long-context error correction. They can be plugged into compatible downstream models without further training, and the authors verify this on 54 downstream models across three backbone families and eight task categories.
In briefPowerSim: Differentiable Physics Simulation and Rendering with Power Diagrams
SummaryThis paper presents PowerSim, a method that couples a pre-trained PowerFoam scene, which uses a power diagram-based 3D representation, with the Material Point Method (MPM). Simulated motion directly drives the scene's geometry and appearance without an intermediate representation. The authors demonstrate applications including interactive simulation of static scenes, recovery of spatially varying material fields, compositing primitives from separate scenes, and ray-traced reflections that update as objects deform.
In briefFracGen: Video Generation That Learns How Objects Stretch and Break
SummaryFracGen is a video generation model that creates controllable fracture dynamics from a single image of an intact object, guided by physics signals. The authors built FracSim, a simulation framework that adds a damage model to material point method simulation to produce paired fracture videos and physical field maps. According to the paper, FracGen outperforms existing video generation baselines in both physical and visual fidelity.
In briefTraining Transformers to feed deep states back to shallow layers via teacher supervision
SummaryThis paper observes that standard Transformer language models are feed-forward, so deep representations never feed back to shallower layers and information across generation steps passes only through decoded tokens. The authors propose LIFT (Latent Information Feedback Transformer), which uses next-token distributions from an off-the-shelf pretrained model as teacher-supervised states, so the model predicts both the next token and the next state during parallel pretraining. Experiments on models from 135M to 1B parameters report that LIFT outperforms standard Transformers and baselines on language modeling, reasoning, and procedural tasks under a token-matched budget.
In briefThinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning
SummaryThis paper introduces agentic meta-reasoning, an inference-time framework that treats an agent's run-time control choices, such as reusing earlier work, starting fresh, or stopping, as an explicit structured reasoning process. Worker components carry out the task while a controller consolidates results, evaluates options under the remaining budget, and dispatches work. The authors report gains over direct control and several baselines across benchmarks, including ProgramBench, with improvement continuing at budgets where direct control plateaus.
In briefLIFT: Controlling Future Scene Layout in Video Generation Under Large Viewpoint Changes
SummaryThis paper presents LIFT, an image-to-video framework that adds Layout-In-FuTure control alongside camera control, so users can specify what appears in a future view and where, especially in the final frame. The authors use on-policy self-distillation (OPSD) to transfer the ability of a teacher model using dense per-frame layouts to a student model that only receives last-frame layout guidance. They also release LIFT-Vista, a dataset with large viewpoint changes and temporally consistent camera and layout annotations. Experiments reportedly show gains in video quality, future-layout controllability, and camera controllability over other methods.
In briefLearning meta-skills for building agent execution environments at test time
SummaryThe paper studies whether a Builder model can learn to construct better execution environments for a Target agent while both models' weights stay fixed. It introduces Meta-Skills, principles that specify when support is needed and what resources to provide, learned from the Target's execution feedback on a development set. On Harness-Bench and NewtonBench, full-bank meta-skills improve macro-average performance by 8.95 percentage points over no-skill construction.
In briefAdviSD: Training a Small Advisor to Guide Large Language Models via Targeted Self-Distillation
SummaryThe paper studies a small trainable advisor that steers a frozen language-model executor using natural-language advice. It proposes AdviSD, which combines outcome-based reinforcement learning with selective self-distillation from a feedback-conditioned copy of the advisor, and argues theoretically that some corrections can limit learning. Experiments with Qwen3-8B advisors on BFCL-v3 and EnvScaler report gains over advisor-GRPO, and the trained advisors are said to generalize to out-of-domain tasks and different executors.
In briefBreaking the Uniformity Trap: Scaling Video Diffusion Models with SplitMoE
SummaryThe paper argues that conventional visual Mixture-of-Experts models scatter coherent video patches across experts because of uniform usage regularization, causing structural distortion. It proposes SplitMoE, which divides the expert pool into semantic experts and generic experts and uses prototype-guided routing with pull-push regularization so tokens cluster by semantic attributes. Under equal activated-parameter budgets, the authors report gains in convergence speed, routing coherence, and video generation quality over load-balanced MoE baselines.
In briefLongHarness Bench: a stress-test benchmark for long-context language model harnesses
SummaryThis paper introduces a benchmark for judging both the effectiveness and efficiency of harnesses that let language models work over long contexts. Its tasks require lexical and semantic retrieval plus strategic reasoning over global and local context. The best model-harness combination reaches 68% macro-average accuracy across four evaluation suites, so the benchmark remains challenging. The authors also find that the same model is markedly more or less efficient depending on the harness, and argue that efficiency should be an important evaluation axis.
In briefCLeaR: A Unified Training-Free Framework Resolving the Leakage-Degradation Tradeoff in Style Transfer
SummaryStyle transfer renders target content in a reference image's style, but existing methods often leak objects, layouts, or semantics from the style reference into the output. The paper proposes CLeaR, a training-free framework that uses orthogonal subspace projection, ensemble inversion across multiple vision foundation models, and energy-guided calibration during diffusion sampling. The authors report that on StyleBench, CLeaR improves style alignment, reduces content leakage, and scores better on LLM-as-Judge evaluation than existing methods.
In briefMulti-agent flow matching that forces generated outputs to meet hard constraints, using decoupled guidance for each agent
SummaryThe paper introduces DeGG-Flow, a framework for multi-agent flow matching with decoupled generative guidance. It models the generation process as a control-affine dynamical system and derives guidance conditions for shared requirements that depend on several agents together and for private requirements that depend on each agent's neighbors. The authors establish feasibility conditions, finite-horizon convergence guarantees, and a Wasserstein bound on the distributional deviation caused by guidance. Experiments on multi-robot gap crossing and multi-object scene generation with affordance requirements show that generated objects satisfy all hard requirements, including at team sizes not seen in training.
In briefAchieving an O(1/N) optimality gap in average-reward weakly-coupled MDPs
SummaryThis paper studies average-reward weakly-coupled Markov decision processes, in which N smaller MDPs called arms share several per-step budget constraints. The arms are assumed to have identical parameters, multiple actions, and state- and action-dependent costs. The authors identify conditions, analogous to those for restless bandits, under which the optimality gap can be better than 1/√N. They design a policy achieving an O(1/N) optimality gap that, unlike priority-based approaches, induces locally linear mean-field dynamics.
WhyImproves the optimality gap from O(1/√N) to O(1/N) for general WCMDPs via a non-priority policy with proven locally linear mean-field dynamics.
In briefSmall Perturbations Can Rewrite a Vision-Language Model's Perception at ε≤4/255
SummaryThis paper tests how far adversarial perturbations can steer Vision Language Models (VLMs). Under a white-box threat model, it aligns each source-image stream with its target-image counterpart in the victim model's post-merger token space. Under a strict success criterion, complete replacement reaches 38% at ε=4/255 for images and 35.9% at ε=1/255 for video.
In briefHelixWorld: A Real-Time Interactive Audio-Visual World Model
SummaryThis paper presents HelixWorld, a real-time interactive world model that generates synchronized visuals and camera-aligned stereo sound as users interact with the scene. The authors build a stereo audio-visual dataset with metric camera poses and distill a bidirectional teacher into a few-step streaming student that runs at 24 FPS on a single GPU. They also introduce HelixBench to test whether generated sound fields track viewpoint motion. The authors report that HelixWorld matches silent world models in visual quality and responsiveness while outperforming baselines in spatial-acoustic immersion.</summary>
In briefWUSH-KV: Data-adaptive transforms for quantizing the KV cache
SummaryKV cache memory and bandwidth costs grow with context length and batch size, limiting long-context inference. This paper introduces WUSH-KV, a low-bit KV-cache quantization method that uses calibration data to build separate data-aware transforms for keys and values to reduce quantization error. It is integrated into SGLang for end-to-end evaluation. At 2 bits, WUSH-KV performs comparably to or better than the OSCAR transform across all evaluated models and downstream tasks.
Jiale Chen, Vage Egiazarian, Eldar Kurti\'c et al.
In briefStochastic world models for verifying vision-based neural feedback systems
SummaryThis paper proposes stochastic world models as perception surrogates, replacing GANs, for verifying vision-based neural feedback systems. The models use physically grounded latents built from operations that standard verifiers can bound, and reproduce held-out frames more faithfully than GAN surrogates with up to 130 times as many parameters. On an emergency braking benchmark, the verification procedure resolves the entire state space, where the prior state-of-the-art verifier left 38% unresolved. On the RGB version of the benchmark, it resolves over 80% of the state space.
I. Samuel Akinwande, Mykel J. Kochenderfer, Clark Barrett
In briefVideoLoop: A looping working-memory agent that reduces semantic thrashing in long video understanding
SummaryThis paper addresses semantic thrashing in long-form video agents, where append-only working memory causes attention to key evidence to collapse. The authors propose VideoLoop, which pairs an outer reasoning loop with an inner loop that retrieves past artifacts and rewrites a bounded working memory after each step. It improves four LVLM backbones by an average of 4.2 points on VideoMME (long), and reaches 88.3% on that benchmark with Gemini 3.1 Pro.
In briefGA-EIRFS: Geometry-augmented sampling to improve rare-class LiDAR 3D object detection
SummaryThis paper argues that long-tailed 3D object detection should consider how observable objects are in LiDAR point clouds, not only how frequently each class appears. The authors propose GA-EIRFS, which adjusts a frequency-based repeat factor using a geometry score from point count, surface-normal entropy, and surface coverage, changing only frame-sampling probabilities. On nuScenes, experiments with CenterPoint and PointPillars show mAP and NDS improvements, while KITTI results vary across seeds.
Taufiq Ahmed, Constantino \'Alvarez Casado, Daniel Herrera Castro et al.
In briefSelf-Aligned Forcing: Streaming Video Generation With Differentiable Noisy History
SummaryThis paper introduces Self-Aligned Forcing (SAF), a training scheme for streaming autoregressive video diffusion that aims to reduce error accumulation over long generation rollouts. SAF aligns each block's history with the noise level of the block being denoised, keeping the noisy history differentiable so training avoids a separate no-gradient rollout. The authors report training up to 1.8x faster with lower memory, and 49.1 FPS on 4 GPUs at inference, with better balance between visual quality and motion in long-horizon generation.
In briefIMPACT: Modeling socially interdependent movement in a multi-agent simulation of a Pompeian household
SummaryThe paper introduces IMPACT, an architecture that uses culturally specific roles and obligations to make AI agents' activities depend on one another, so movement is shaped by social coordination. It is instantiated in a five-hour simulation of a Pompeian dinner with ten agents. In an ablation evaluation, thirty-seven participants rated the full architecture as more socially coherent and believable than two reduced versions, and six archaeology experts found the movement patterns historically plausible while noting areas needing stronger historical grounding.
In briefReCIRC: Rectified Conformal Risk Control (a method to control per-input error rates in black-box predictions)
SummaryThis paper addresses black-box predictive models where task-relevant error rates, such as missed lesion pixels or missed labels, must be controlled. Standard conformal risk control (CRC) uses one threshold for all inputs, giving only a marginal guarantee that can overprotect easy cases and underprotect hard ones. ReCIRC inverts each input's estimated local risk curve to reparameterize the threshold as a shared target risk budget, then applies CRC to this family. Across three synthetic and five real-data settings, ReCIRC achieved the lowest average worst-group risk and mean positive group excess in every setting while keeping marginal risk close to the target.
Bruno Marcondes e Resende, Helton Graziadei, Thiago Rodrigo Ramos et al.
In briefFrom Routing Signals to Selective Review: Regrounding Vision in MoE VLMs
SummaryVision-language models can accept false visual premises, answering about objects that are not actually in the image. This paper presents a framework that uses internal routing decisions of Mixture-of-Experts vision-language models to detect target absence before generation. Separate linear detectors are trained for two models and used to selectively trigger a target-aware review prompt. The authors report improved end-to-end accuracy on two datasets without modifying model weights.
In briefHow local mixing encodes relative position in global NoPE attention without explicit position embeddings
SummaryThis paper explains how hybrid transformer models can implicitly encode position in global attention layers that use no explicit position encoding (NoPE). The authors argue that sliding window attention and gated linear attention induce a recency bias in the residual stream, which global attention logits propagate and select. Unlike models with only global NoPE attention, this recency bias can be maintained across long sequences.
In briefDo LLM agents actually carry out the plans they declare?
SummaryThis paper studies the gap between the plan an LLM agent declares and how it actually executes it. The authors propose Planning-as-Routing, where the LLM picks one of four planning modes (Predefined, Sequential, Hierarchical, or Search) and a deterministic router sends the task to a matching executor. Across four benchmarks and three LLMs, generic Plan+ReAct often fails to preserve the declared structure, while pattern-specific executors greatly improve success. However, current LLMs do not reliably choose the best mode for each task.
Subba Reddy Oota, Francisco Herrera, Jordi Cabot Sagrera et al.
In briefRight Answers, Flawed Reasoning: Verifiable Math Shows Chain-of-Thought Limits
SummaryThe paper uses iGSM, a synthetic grade-school math benchmark, to check model chain-of-thought traces step by step with programmatic verification. On the hardest instances, 31.6% of correct answers had invalid traces. Shuffling or swapping training trace content still preserved high accuracy, and the authors discuss implications for chain-of-thought monitoring and AI safety.
Ratish Puduppully, Pranabendu Misra, Paarth Iyer et al.
In briefPruning Speech-LLMs for speed can widen fairness gaps across demographic groups
SummaryThis paper studies how pruning the audio encoder of a speech LLM (SLAM-ASR) affects recognition error for different demographic groups, using the Fair-Speech and Common Voice datasets. Pruning does not affect all groups equally, and the gap between the best and worst groups grows, a disparity that the aggregate word error rate can hide in the largest model. LoRA adaptation improves every group but helps already well-performing groups more. The authors recommend reporting per-group word error rate and using the worst group's error rate as an explicit deployment criterion.
Ganesh Pavan Kartikeya Bharadwaj Kolluri, Michael Kampouridis, Ravi Shekhar
In briefExplore Broadly, Reason Sharply: Boosting Small Models via Sampling
SummaryThis paper proposes Parallel Power Tempering (PPT), an inference-time sampling method that improves reasoning in large language models without parameter updates or external rewards. It runs multiple interacting replicas at different sharpening levels, so low-power replicas explore diverse reasoning paths while high-power chains exploit higher-likelihood answers. The authors report that PPT substantially improves single-chain power-sharpened sampling and outperforms RL-post-trained models, with performance comparable to frontier models.
Panagiotis Theodoropoulos, Nan Jiang, Xintong Duan et al.
In briefPaper Showing AI Agents Can Be Radicalized Through LLM-to-LLM Conversation
SummaryThe paper simulates conversations between a target LLM role-playing a demographically defined persona and an influencer LLM trying to make the target's beliefs more extreme. It examines two pathways: resonance, which reinforces existing beliefs, and persuasion, which promotes beliefs the target initially considered unimportant. Both pathways radicalize the target, with resonance producing consistently stronger effects that also spread to related beliefs. The authors raise concerns about vulnerability of personalized AI agents and multi-agent systems.
Ozgur Can Seckin, Shalmoli Ghosh, Alessandro Flammini et al.
In briefTraining-free dense retrieval from in-context examples alone
SummaryThe paper asks whether decoder-only large language models can produce effective dense retrieval representations using only a few in-context examples, without retriever training. It introduces RICE (Representations from In-Context Examples), which conditions the LLM on examples that provide a shared context for encoding queries and documents. The authors report that RICE embeddings substantially improve the accuracy of prompt-based LLM embeddings.
In briefNeuronEye: Query-Guided Activation of Visual Concepts for Vision-Language Reasoning
SummaryVision-language models mix object identity, spatial layout, and attributes together in dense hidden states, making it hard to isolate the visual evidence a question needs. This paper introduces NeuronEye, a plug-in framework that builds a sparse concept-level neuron vocabulary and activates query-relevant visual concepts during inference. On Qwen2.5-VL-7B, it raises CV-Bench overall accuracy by 3.1 points, with a 9.5-point gain on Distance, and improves BLINK Multi-view by 8.3 points.
In briefTail-Influence Sampling for Estimating Rare-Failure Risk in Policies
SummaryThe paper studies how to allocate a fixed evaluation budget to estimate a policy's lower-tail conditional value-at-risk (CVaR) as accurately as possible. It defines a tail influence for each queryable conditional distribution and proposes Tail-Influence Sampling (TIS) with a visitation-anchored variant to reallocate queries. Experiments on CliffWalking and frozen language-model review workflows report lower mean squared error than learned-occupancy and full-rollout baselines.
Pauline Bourigault, Xiaotong Ji, Matthieu Zimmer et al.
In briefFaster, more stable zero-order training via probe-space preconditioning
SummaryThis paper addresses why zero-order optimization (ZOO) training converges more slowly than backpropagation. It shows that shifting compute from many steps to large batches with many perturbations lets 1SPSA outperform MeZO with less training compute. It then introduces 1.5-SPSA, which adds one clean forward pass per step to build a diagonal preconditioner in probe space, improving convergence. The authors report state-of-the-art results across six post-training datasets on Qwen3 and OPT models, and say OPT-30B can be trained in place on commodity GPUs such as the A100.
Francois Chaubard, Mykel J. Kochenderfer, Chris R\'e
In briefDimensionally consistent surrogate models built with dimensional analysis and harmonic expansions
SummaryThis paper proposes a data-driven method for building surrogate models that respect dimensional homogeneity, so their results do not change under unit conversions. The method derives Buckingham Π-groups from the dimension matrix of measured variables, builds dimensionally admissible prefactors, and approximates the remaining dimensionless dependence with truncated harmonic expansions. Coefficients are found by regularized linear regression. Tests on the simple pendulum, Planck's law, a double-pendulum Lyapunov field, and COBE/FIRAS black-body spectrum data show better conditioning, noise robustness, and sample efficiency than unconstrained baselines.
In briefTraining AI agents to be risk-averse through character traits
SummaryThis paper proposes using character training to make AI agents risk-averse, so that misaligned agents are less likely to cause catastrophic harm. The authors write a model constitution describing constant absolute risk aversion (CARA) over an agent's resources and instill it through on-policy distillation. Character-trained models perform competitively with baselines trained directly on the benchmark, despite never seeing its decision format, and generalise better out of distribution on two of four models.
Arav Dhoot, Punya Syon Pandey, Jamie Johnson et al.
In briefMira: Memory-efficient MoE inference via adaptive caching and predictive expert staging
SummaryThis paper presents Mira, a co-designed algorithm and system that enables high-capacity Mixture-of-Experts (MoE) model inference on a single GPU. It uses lightweight per-layer predictors to anticipate expert usage two layers ahead, combined with a two-tier GPU cache and a custom quantization format to reduce transfer stalls. The authors report speedups over state-of-the-art baselines of 5.71x in average throughput, 11.71x in Time-to-First-Token, and 3.84x in beam search inference on a memory-constrained GPU.
In briefNeural topology optimization of ship structures to reduce propulsion-induced vibration
SummaryThis paper extends neural-reparameterized topology optimization, using a convolutional Kolmogorov-Arnold network (KATO), to design ship structures under forced vibration from propulsion machinery. It targets active input power as the objective and tests a 100 Hz deck panel and an 18 Hz thruster frame. Results are compared with GCMMA, with KATO producing connected, feature-controlled designs and running 6.4 to 10.4 times faster.
Shengyu Yan, Muhammad Muztahidul Hakim Zareer, Jasmin Jelovica
In briefVISTA: Sharing Visual Discoveries Across Attempts to Train Tool-Using Multimodal AI Agents
SummaryThis paper presents VISTA, a training method for multimodal agents that use visual tools to gather task-relevant evidence while reasoning. It turns observations from multiple attempts on the same input into shared supervision, so discoveries in one trajectory can guide learning in another, using an experience-conditioned teacher. The authors report that VISTA achieves the strongest average performance among evaluated active multimodal agents of comparable size and outperforms same-backbone training baselines on fine-grained perception and general reasoning tasks.
In briefExploring neural network solution spaces with Hessian null space continuation
SummaryThe paper introduces Hessian Null Space Continuation (HNC), a method that uses local curvature to move through regions of weight space that preserve network function, and can be steered toward solutions with specified properties. Tested on RNNs with a memory task, ImageNet-trained Vision Transformers, and reinforcement-learning agents, HNC finds representations and strategies that differ substantially while keeping behavior the same. The results show a large amount of representational diversity exists near a single trained solution.
WhyIntroduces Hessian null-space traversal showing diverse functionally equivalent solutions, validated across RNNs, ViTs, and RL agents.
In briefMoFlow: Generating AI Agent Workflows That Balance Multiple Goals
SummaryThis paper studies generating agentic workflows that balance several objectives at once, such as accuracy, cost, latency, robustness, and consistency. MoFlow treats workflow generation as a multi-objective decision problem and uses Convex-Hull Monte Carlo Tree Search, where each node stores a set of reachable trade-offs. One search approximately covers the Pareto front, so a workflow for any preference can be retrieved without retraining. On six benchmarks spanning mathematics, code, and question answering, MoFlow achieves the highest average hypervolume.
In briefOmniTaskonomy: When Visual Generation Helps Visual Understanding
SummaryThis paper examines whether training models on image generation tasks can improve their visual understanding, and under what conditions. The authors compare controlled pairs of image-to-image generation and image-to-text understanding tasks, finding that under the correct training recipe, generation training improves downstream understanding, with larger gains as generation data increases. They introduce OmniTaskonomy, a taxonomy covering 19 generation tasks and 25 understanding capabilities, which reveals that transfer benefits are selective and task-dependent.
In briefMUGEN: Camera-controlled interactive panoramic 360-degree world exploration
SummaryThe paper introduces MUGEN, a large-scale dataset of over 1,300 hours of at least 4K panoramic videos with rich semantic and geometric annotations for interactive 360-degree world exploration. Building on it, the authors present Wan360, a camera-controllable panoramic video generation model that uses parameter-free ERP-aware components and a panoramic Plücker embedding for camera motion. Experiments indicate that MUGEN provides a data foundation for panoramic world exploration and that Wan360 produces high-quality, temporally coherent, camera-controllable 360-degree videos.
In briefRS-OPSD: A reliable privileged on-policy self-distillation method for ultra-high-resolution remote sensing visual question answering
SummaryThis paper proposes RS-OPSD, a framework that internalizes the benefit of zooming into image regions into a model for answering questions about ultra-high-resolution remote sensing images, without extra visual search or tool calls at inference. The authors build GeoEvidence-6K, containing 6,750 VQA samples with evidence-region annotations, and introduce Context-Preserving Visual Privilege and Correctness-Aligned Distillation. According to the paper, RS-OPSD achieves state-of-the-art results on XLRS-Bench, MME-RealWorld-RS, and LRS-VQA, outperforming comparable-scale models by an average of 4.0 percentage points.
In briefProbability Alone Is Insufficient: Counting Divergent Tokens to Quantify LLM Reasoning Uncertainty
SummaryThis paper introduces Divergent Token Confidence (DTC), a method that estimates an LLM's confidence in its answer by counting tokens where two models strongly disagree along the same reasoning trajectory. Divergent tokens are identified using Jensen-Shannon divergence between next-token distributions, and their count is almost negatively correlated with answer accuracy. Experiments across multiple model families and six math benchmarks show better calibration than probability-based and verbalized baselines in both white-box and black-box settings.
WhyCounts JS-divergent tokens between two models along one reasoning trajectory; tested on multiple model families and six math benchmarks with clear calibration gains.
In briefA foundation model for energy and radiation systems using heterogeneous scientific interfaces (GEODE)
SummaryThe paper presents GEODE, which connects task-specific scientific interfaces to a shared library of wavelet operators, so one jointly pretrained model handles cavity flow, radiation dose, and elastoplastic stress. New tasks are learned by training a private interface holding about 2.1% of the parameters. Parameter isolation keeps earlier predictions unchanged, while unrestricted fine-tuning degrades them substantially. The authors argue that preservation alone does not prove pretrained reuse, which requires norm-matched randomized controls.
In briefAlpha Diffusion Language Models: Factorization Alone Is Not the Problem
SummaryDiscrete diffusion language models generate several tokens in parallel, but fewer denoising steps can make their predictions inconsistent. The paper proposes AlphaDLM, trained with a sequence-level alpha loss that reduces to cross-entropy as alpha approaches zero and has a joint-mode optimum at alpha one. Trained on TinyGSM, it reaches 34.6% accuracy on GSM8K with only four model evaluations, and the method is also scaled to SDAR-1.7B and tested on code and math benchmarks.
Nikita Gushchin, Dmitry Baranchuk, Alexander Korotin
