In briefSharP: LLM Data Augmentation with Soft and Hard Prompts for Low-Resource Fake News Detection
SummaryThis paper addresses fake news detection when only a small set of labeled training data is available. It proposes SharP, a framework that uses large language models guided by hard prompts (structured instructions and factual constraints) and soft prompts (learned semantic patterns) to generate task-specific synthetic samples. A boundary-aware strategy directs generation toward examples the classifier is less confident about, and a dual objective balances semantic alignment with classifier feedback to improve robustness and generalization.
In briefRethinking how relative policy differences transfer in on-policy distillation
SummaryThis paper studies a refinement of on-policy distillation (OPD) that transfers the relative difference between an expert policy and a reference policy, rather than the expert's absolute distribution. The authors find that differences from task-specific training, such as reinforcement learning or supervised fine-tuning, transfer consistently, while differences arising from model scale alone transfer weakly or not at all. They also conclude that transfer succeeds only when the induced update moves the target toward the expert and generalizes beyond sampled rollouts.
In briefS³R: Faster, Faithful RL Rollouts via Self-Sparse Speculative Decoding
SummaryThis paper addresses slow response generation in long-context reinforcement learning with verifiable rewards (RLVR). It proposes S³R, which uses sparse attention to draft tokens and full-attention verification with exact speculative sampling to produce responses. Drafting and verification share one policy, so no separate drafter needs training, and the authors state the response groups follow the dense-policy distribution. On long-context math reasoning with Qwen3, S³R reports 1.35–1.39× end-to-end training throughput over autoregressive rollouts with comparable task performance.
In briefIs training the recurrent memory needed in partially observable RL?
SummaryThis paper asks whether a frozen recurrent network can supply enough memory for partially observable Markov decision processes without being trained. The authors pair a fixed recurrent network with a trainable nonlinear output module and test it on tasks with different memory demands, observation types, and action spaces. Frozen recurrent networks recall over hundreds of steps, reach competitive control performance, and use less than half the trainable parameters of fully trainable models.
In briefJANUS: Co-Evolving World Models for Better Long-Horizon Autonomous Research
SummaryThis paper proposes JANUS, a framework that improves long-horizon autonomous research by pairing a research agent with a co-evolving world model. Before execution, the world model forecasts the value, information gain, and exploration utility of candidate ideas, guiding which branches to keep, refine, or prune. After receiving ground-truth feedback, the system distills experience into evolving skills for both components. The authors report that JANUS significantly outperforms state-of-the-art baselines on MLS-Bench-Lite and real research tasks.
In briefPrediction-Informed Landing Window Planning for UAV Landing on USV in High Sea States
SummaryThe paper addresses autonomous landing of a drone on a moving unmanned surface vehicle in rough seas. It proposes a hierarchical framework in which a pretrained model explicitly predicts the vessel's attitude to help select landing windows and touchdown timing, while a conventional flight controller handles stability and tracking. Real sea-condition USV motion data is transferred to a MuJoCo simulation to form a real-sim validation setup. Experiments indicate that explicit attitude prediction is feasible for improving landing-window selection.
In briefDestination-conditioned adaptation for smoother generative in-between frame synthesis
SummaryThis paper presents DSG-GI, a framework that adapts pretrained latent video diffusion models to generate intermediate frames between a start and a target frame. It combines destination-conditioned adaptation, which guides the sequence toward the target frame, with semantic-geometric alignment that matches token-wise spatio-temporal relations to a frozen video encoder. The authors report improved temporal stability, target-frame faithfulness, and perceptual quality on standard and large-motion benchmarks.
In briefWaSA: Unsupervised Offline Detection of Malicious Clients in Federated Learning Using Wavelet Autoencoders
SummaryFederated learning trains models collaboratively without sharing raw data, but malicious clients can inject poisoned datasets. This paper proposes WaSA, which has each client send one privacy-preserving spectral summary before training. A denoising autoencoder trained only on clean auxiliary data scores clients by their deviation from benign patterns, without needing attack labels. Experiments on MNIST, CIFAR-10, and CIFAR-100 with up to 90% malicious clients show gains over supervised baselines and robust aggregation methods, with negligible communication overhead.
In briefAdaC-GRPO: Adaptive Curriculum Learning for Long-Horizon Visual Sudoku Puzzles
SummaryThis paper introduces FG-Sudoku, a fine-grained benchmark that categorizes visual Sudoku puzzles by grid size and required deductive steps. Evaluations show that frontier multimodal models fail on complex long-range dependency puzzles such as 9×9 grids, causing reward sparsity during reinforcement learning. The authors propose AdaC-GRPO, which dynamically builds training batches by tracking the model's online rollout accuracy. Experiments report that Qwen2.5-VL-7B approaches Gemini-2.5-Flash and closes 62.2% of the overall accuracy gap to Gemini-2.5-Pro, with generalization also shown on InternVL-3.5 and Qwen3-VL.
In briefMixTok: Adaptive image tokenization combining global 1D tokens with sparse 2D refinement
SummaryThe paper introduces MixTok, a hybrid visual tokenizer that pairs a compact 1D global representation with sparse 2D tokens for local detail. A content-adaptive router ranks image locations by refinement priority and assigns 2D tokens per image under a target average token budget. On ImageNet 256 x 256, MixTok reports rFID of 1.195 and gFID of 1.95, using about 128 and 131 content tokens per image on average.
In briefFast generative video upscaling that runs on one consumer GPU
SummaryThis paper presents a zero-shot framework built only from pretrained components, requiring no end-to-end retraining, for generative video super-resolution. It separates spatial upscaling, handled by a time-distilled one-step diffusion model, from temporal stabilization, handled by a lightweight flow-based frame interpolator. A resolution-aware hybrid attention scheme swaps some high-resolution global attention layers for neighborhood attention, yielding up to 44% speedup at 4K. The system reaches about 32 FPS at 720p on a single RTX 4090, trading some reconstruction fidelity for speed.
In briefHidden Failures in AI Data Science Agents Beyond Accuracy Scores
SummaryThis paper argues that current evaluations of agentic data science only measure downstream predictive performance and code correctness, which can hide methodological errors such as data leakage and improper validation. The authors assess agents from three angles: a large-scale failure-mode analysis across 220 tasks, a reusable test suite targeting six specific failures, and a qualitative comparison with expert Kaggle solutions. They find that most agentic solutions contain at least one failure mode and that data leakage correlates positively with predictive performance. Compared with competition-winning data scientists, the agents generally lack problem-solving creativity and domain knowledge.
In briefCompactSplat: A Feedforward Method for Compact 3D Scene Reconstruction
SummaryCompactSplat is a feedforward framework that reconstructs 3D scenes from sparse views without per-scene optimization. It uses texture-guided spatial partitioning and hierarchical sampling to place dense Gaussians in complex regions and larger, scale-modulated Gaussians in flat areas. Experiments on RealEstate10K, DL3DV, and ScanNet report better standard metrics and rendering consistency than prior methods, with significantly fewer primitives.
In briefFixing the Piece-Matching Measure, Not the Solver: STITCH for Eroded Puzzles
SummaryThe paper argues that for puzzle fragments whose boundaries are eroded, the compatibility measure rather than the solver is the main performance bottleneck. It proposes STITCH, a self-supervised measure that compares frozen ViT patch tokens from piece interiors through a lightweight cross-attention head, trained without manual annotation. Plugged into existing compatibility-based solvers, it reports strong results on JPwLEG-5, an ImageNet 7x7 benchmark, and GAP datasets.
In briefPreCoG: Efficient Video Correction During Generation via Preview Editing
SummaryThis paper presents PreCoG, an iterative framework that corrects semantic failures during text-to-video generation. It diagnoses errors from low-cost previews and edits cached generation trajectories, limiting disruption to content that is already correct. Experiments report the highest VBench and T2V-CompBench scores, at about 0.56 times VSR's inference time, and outperformance of GENMAC on three metrics at about 0.86 times its inference time.
In briefRoleWorld: Transferring World Models via Functional Role Abstractions
SummaryThis paper introduces RoleWorld, a symbolic world modeling method that represents how objects behave using reusable role prototypes and environment-specific parameter values. A teacher LLM extracts role prototypes and executable transition rules from interaction trajectories, and an actor LLM applies them in a new environment by refining role assignments and parameter estimates online without updating model parameters. Experiments on the MARS benchmark reportedly show robust generalization to environments where object behavior is significantly altered.
In briefPreSTO: Speeding up LLM power sampling by prefetching future proposals
SummaryThis paper introduces PreSTO (Predictive SubTree Prefetching), a method that speeds up power sampling for large language models, whose Metropolis-Hastings updates are normally sequential. PreSTO generates proposals for a prefetched subtree of future accept/reject outcomes in one batch, then follows the realized path with the original acceptance rule, leaving the transition kernel unchanged. Across seven datasets and five base LLMs, it reports median speedups of 1.71x, 1.91x, and 2.33x on PowerMH and its follow-ups.
In briefScalable Data Synthesis for Long-Horizon Terminal Agents
SummaryThe paper introduces TerminalHorizon, a fully automated engine that synthesizes training data for long-horizon terminal agents, starting from real-world professional work. It uses independent execution to find key work omitted during environment construction and rebuilds environments accordingly, while progressively expanding tasks toward the same goal. Using 1.5K constructed task environments, the authors fine-tuned Qwen3.5-27B into TerminalHorizon-27B, which shows gains on all 12 benchmarks tested.
In briefSOAR: Source-aware RL with precise credit assignment for search agents
SummaryThis paper addresses evidence fabrication in reinforcement learning for search agents, where agents skip external information and invent incorrect answers from internal knowledge. The authors propose SOAR, which separates external and internal information into observable streams and uses a Source-Consistent Reward to guide training. Reported results show gains of 7.3% and 5.4% over baselines on 3B and 7B backbones, with evidence fabrication reduced by 74.8%.
In briefFrom Glance to Focus: Reducing Recognition Laziness in MLLMs for Fine-Grained Recognition
SummaryThis paper studies why multimodal large language models underperform on fine-grained visual recognition and identifies an inference-time issue called FGVR Laziness, where visual cues are weakly linked to category concepts. The authors propose LIVE, a training-free framework that refines candidate lists and recognition guidance to draw out the models' latent recognition ability. Experiments on 13 benchmarks against 24 methods show that LIVE is effective across different MLLM backbones.
In briefScoring coding-agent trajectories by process quality, not just test success, for fine-tuning data selection
SummaryOpen coding agents are often fine-tuned on interaction trajectories whose final patches pass tests, but this ignores inefficient process patterns within successful runs. The paper introduces TQS (Trajectory Quality Score), a structural scoring method that needs no LLM judge and runs in linear time. Using 155,144 successful trajectories, models trained on top-ranked subsets outperformed those trained on random or bottom-ranked subsets on SWE-Bench-Verified.
In briefLes Misérables Dialogue: a film-based network of social communication
SummaryThis paper uses the 1998 film adaptation of Les Misérables to build the publicly available Les Misérables Dialogue temporal text network. Each utterance records its speaker, addressees, other recipients, content, and timing. Using the ten social dimensions model, the authors also derive multilayer network representations in which each layer corresponds to a distinct relational dimension. The work additionally offers a general method for building similar communication datasets from other narrative sources.
In briefVIGOR: Verifiable Industrial Layout Optimization and Repair for Scene Programs
SummaryThis paper shows that scene programs generated by LLMs can run successfully yet still violate coupled engineering constraints such as boundary, collision, clearance, aisle, and structural rules. The authors build a dataset of 1,749 constructible industrial scenarios and propose VIGOR, which combines supervised program generation, execution-grounded verification, and staged geometric repair. Reported results include 96.86% Complete Feasibility and an Overall score of 91.20.
In briefCompetence-aware orchestration for LLM-based multi-agent systems
SummaryThis paper proposes a two-stage framework that automatically orchestrates LLM-based multi-agent systems by first estimating the backbone model's competence on the current query. It then uses this estimate to adapt role activation and the communication graph. The method reports an average score of 93.81% across six benchmarks. For higher-competence backbones, it raises average accuracy from 87.22 to 89.97 while cutting token cost by 36.26%.
In briefMeasurable semantic 3D Gaussian splatting for interactive building inspection
SummaryThis paper proposes a 3D Gaussian Splatting framework that uses mask-based region supervision and depth supervision from multi-view stereo point clouds to reconstruct buildings from multi-view images. Each Gaussian is associated with learnable semantic features, enabling component-level inspection and measurement. The framework is implemented in the AEC-SplatTwin platform, which supports real-time visualization and multimodal LLM-assisted interaction. On HKUPB, mean relative errors for height, width, and area measurements are 0.33%, 1.07%, and 0.60%.
In briefNeurograf v1: An End-to-End Romanian-First Language Model Family
SummaryThis paper presents Neurograf v1, an experimental pipeline for training language models from scratch for Romanian, a lower-resource European language. It includes a 635 GB Romanian corpus, an 80k byte-level BPE tokenizer, five decoder-only models from 160M to 6B parameters, three instruction-tuned assistants, and two embedding models. On the 1,000-question WWTBM benchmark, Neurograf-6B reached 69.89% accuracy on Romanian-context questions. The authors describe the system as an experimental demonstrator rather than a production system and discuss challenges such as diacritic preservation and corpus quality.
In briefLimits of Manipulation in Personalized Incentive Policies: When Can Agents Profit by Faking Their Profile?
SummaryThis paper studies personalized subsidy policies that aim to steer demand toward sustainable choices, and asks when agents can profit by simulating a different profile. The authors model the interaction as a Stackelberg game in which the regulator learns each agent's true profile from past behavior, so manipulation carries a cost. They find that classification uncertainty, though it lowers personalization accuracy, can deter manipulation and benefit the regulator. They derive an upper bound on the gain from manipulation and illustrate the effect with simulations on the London Passenger Mode Choice dataset.
In briefDP-MIND: Memory-guided differentially private inference for million-token contexts
SummaryThis paper addresses the privacy risk when users send long, sensitive inputs to LLM services. It proposes DP-MIND, which locally perturbs the input with differential privacy, processes it in chunks using a compact memory optimized by reinforcement learning, and uses a local denoising module to recover corrupted answers. The authors provide a theoretical privacy analysis and report experiments on long-context benchmarks up to million-token scale, where DP-MIND outperforms existing DP and long-context baselines in scalability and utility.
In briefLooking Ahead and Back: Anticipatory Generative Modeling for Road-Network Trajectories
SummaryThis paper studies generating vehicle trajectories on road networks, where choices at intersections strongly shape later routes. The authors note that existing methods such as diffusion models struggle with next-transition selection and can produce trajectories that leave the road network. They propose Anticipatory Generative Modeling and a model called Foresight, which uses an anticipate-select-update-re-anticipate loop. The paper reports that Foresight outperforms state-of-the-art baselines in extensive experiments.
In briefFaster gradient-based approximation of continuous time series alignment
SummaryDynamic Time Warping measures differences between time series but is slow and not differentiable. This paper shows that Continuous DTW is differentiable under mild conditions and introduces FastCDTW, an efficient approximate solver. It uses an exact closed-form method for piecewise-linear warping functions and an unbiased Monte Carlo estimator for smooth ones, both providing gradients and running on GPUs. The authors demonstrate its use in time series averaging, classification, and deep domain adaptation.
Pierre Houedry, Romain Tavenard, Nicolas Courty et al.
In briefLocate, Judge, Read: Role-Specialized Agents for Multimodal Document QA
SummaryThis paper presents KnowForge, a pipeline that splits open-domain multimodal document question answering into separate roles: a search-only Locator agent, a code-scheduled bridge judge, a Deep Reader for text, images and tables, and a synthesizer. It is built on a hybrid-retrieval knowledge-point index. The authors report higher F1 and accuracy than published systems on M3DocVQA, MultiHop-RAG and MMLongBench-Doc, and per-job ablations attribute the gains to each component.
WhySeparates search and reading into role-specialized agents with forced bridge judging, showing solid gains and thorough per-component ablations across several benchmarks.
In briefWhy heavy-tailed weight spectra can predict neural network quality
SummaryThis paper examines why the spectra of deep neural network weight matrices predict model quality, taking a statistical view. Using a probabilistic framework, the authors derive a marginal likelihood from only the final-layer weights and find it predicts quality across large convolutional networks and LLMs. They also derive the limiting log-marginal likelihood for heavy-tailed High-Temperature Marchenko-Pastur spectra, show it is convex in a heaviness parameter, and identify an optimal heaviness that empirically predicts over-fitting in large DNNs.
Joseph Wilson, Chris van der Heide, Liam Hodgkinson et al.
In briefGuidedNav: Guiding Vision-Language Navigation Representations with Pre-action Attention and Spatial Cues
SummaryThis paper proposes GuidedNav, a framework for vision-and-language navigation that uses training-time guidance to focus pre-action attention on the active instruction segment and relevant visual regions. Spatial guidance further aligns the policy's internal representations with geometric features capturing long-horizon and local spatial structure. The authors also release Guided-R2R and Guided-RxR, two annotation-enriched datasets. The reported results show success rate gains of 8.1% on R2R-CE and 13.1% on RxR-CE val-unseen over the strongest prior results among compared monocular RGB policies, with faster inference.
In briefBenCoDiC: A Large-Scale Contextual Cyberbullying Corpus for Bengali Political News
SummaryThis paper introduces BenCoDiC, a publicly available Bengali cyberbullying corpus built from 100 YouTube videos and their discussions, containing 50,250 texts with caption, comment, and reply relationships preserved. The corpus uses a nine-dimensional star schema and was annotated through a two-tier human process with expert verification, reaching Fleiss' kappa and Krippendorff's alpha of 0.86. Transformer experiments show that adding conversational context gives modest but consistent gains over classifying isolated texts.
Md. Nesarul Hoque, Rudra Pratap Deb Nath, Umme Salma et al.
In briefDecoding Visual Perception and Recall from EEG
SummaryThis paper introduces PAIR, an EEG dataset from 20 participants who watched video clips and then immediately recalled them visually. The authors benchmark several decoding models on semantic classification, meta-feature decoding, and a task distinguishing perception from recall. Results show similar performance for several model-task pairs, higher perception accuracy in two coarse-grained comparisons, and high separability between perception and recall at the state level.
In briefHow information value drives communication in evolved e-puck robots
SummaryThe paper studies when communication emerges between two autonomous e-puck robots searching for a feeding zone detectable only by passing over it. Robots are controlled by a GRU recurrent network, with fitness based only on time spent over food. When a compass pointing to the food is provided, signal use disappears completely, while without it, different runs adopted different strategies, and referential signaling generalized best as the arena expanded.
In briefDRYLAB-Bench: A benchmark for biosecurity risk in LLM agents designing biological mutations
SummaryThis paper introduces DRYLAB-Bench, the first benchmark for testing whether LLM agents can design high-risk biological mutations. It includes 30 hazardous scenarios drawn from biosecurity guidelines, in which models iteratively refine designs using biological design tools. Designs are scored against 187,695 wet-lab measurements from ProteinGym and ViroGym using a Biosecurity Risk Score. Across 15 frontier models, the authors report that tool access raises the pooled risk score by 0.116, with 23.5% of designs reaching the top risk decile.
WhyFirst comprehensive benchmark of LLM agent biodesign risk, validated against 187,695 wet-lab measurements and quantifying BDT-driven risk amplification.
In briefContinual post-training for small models via synthetic self-distillation
SummaryThe paper addresses how small models can learn new skills without forgetting prior knowledge, a problem where existing Self-Distillation Fine-Tuning (SDFT) fails at smaller scales. The authors link poor learning and forgetting to distribution shift after post-training and propose Synthetic Self-Distillation (SynthSelf), which outsources in-distribution rewriting to a slightly larger same-family aligner with off-policy supervision and rejection sampling. Evaluations on 4 acquisition datasets and 14 retention benchmarks reportedly show SynthSelf outperforms 6 methods across 3 small model scales.
In briefDoes Location Matter? A black-box causal audit of audio grounding in LALM hallucination
SummaryLarge audio-language models often affirm sounds that are not present. The paper proposes a black-box causal test that localizes the audio evidence for a claim using CLAP similarity, removes it with spectral gating, and compares against a random-window control across three models. When clip duration is held fixed, localized removal outperforms random removal in every condition, while raw detection performance varies by model. Contrastive decoding that combines blank-audio and language-prior contrasts raises Qwen2-Audio accuracy from 57.1% to 71.9%, beating Audio-Aware Decoding alone.
In briefWhat Visual-Head Discovery Scores Find: Attention Source, Necessity, and Random-Control Pitfalls
SummaryThe paper examines how attention-based scores select attention heads in vision-language models that focus on question-relevant image regions, heads often used for steering, grounding, and reducing hallucinations. Using Qwen3-VL-8B-Instruct, the authors vary query type, discovery corpus, and attention source, then compare the resulting head sets against layer-matched random controls. They find that attention source and query type reshape selected heads much more than the corpus, and that necessity distinguishes discovery settings. Steerability, however, is not specific to the discovered heads.
In briefScaling RNA Foundation Models via Iterative Refinement: an 80M-parameter model whose inference depth can be adjusted
SummaryThe paper presents Eco-RNA, an 80M-parameter RNA foundation model whose inference depth is set by how many times a shared Transformer core iteratively refines sequence representations, with parameters held fixed. On the 70-assay RNAGym fitness benchmark, raising iterations from one to four increases aggregate absolute Spearman correlation from 0.242 to 0.286. At four iterations, the model surpasses Evo 2 40B on all three aggregate fitness metrics despite having 500x fewer parameters. The authors also report gains over evaluated baselines in non-coding RNA family classification, secondary-structure prediction, and chemical probing profile prediction.
WhyIntroduces learned iterative refinement for RNA foundation models, outperforming Evo 2 40B on 70 RNAGym assays with far fewer parameters.
In briefFinding Where AI Agents Fail, Without Step-by-Step Error Labels
SummaryThis paper studies how to identify the step that caused an LLM agent to fail, using only failed trajectories without step-level labels. The authors propose SOAP, which builds a spectral reference space from LLM representations to score each step, then uses attention patterns to propagate downstream error signals back toward likely upstream causes. Experiments reportedly show that SOAP consistently outperforms competitive baselines on attribution.
In briefNavigation AI That Tracks Which Instructions It Has Already Followed
SummaryThis paper presents VLN-AIEP, a framework for Vision-Language Navigation that estimates how far an agent has progressed through a multi-step instruction. It suppresses the content already executed so that navigation decisions match only the pending instructions. The method learns progress directly from navigation decision labels and uses landmark-anchored trajectory representations. Experiments on R2R, RxR, CVDN, and R2R-CE report favorable results across multiple metrics.
In briefFinnish speech self-supervised learning from 2.1M hours of raw broadcast audio, filtering by recognition confidence
SummaryThis paper studies self-supervised speech pretraining on 2.1 million hours of unfiltered Finnish broadcast audio. The authors use automatic speech recognition confidence as a proxy for data quality, which yields 3-7% relative word error rate gains over language-identification-based filtering with a smaller training set. They also compare Transformer, Macaron-Net, Conformer, and E-Branchformer architectures at matched parameter budgets, and report improvements over prior Finnish SSL baselines, especially at Base scale and out-of-domain.
Yaroslav Getman, Tam\'as Gr\'osz, Tommi Lehtonen et al.
In briefVETR: An Execution-Grounded Agent That Self-Improves Symbolic Graphics Programs
SummaryThis paper introduces VETR, a framework in which visual agents refine executable scene programs by analyzing execution traces rather than only final renders. An execution refiner turns observed runtime state changes into constrained alternative programs, checked against captured runtime state. On BlenderGym and BlenderBench, the authors report that VETR consistently outperforms existing refinement methods, including BlenderAlchemy and VIGA.
In briefFLEX-MRI: stochastic few-step learning for faster MRI image reconstruction
SummaryThis paper presents FLEX-MRI, a method for accelerated MRI reconstruction. It divides the reconstruction process into a small number of finite intervals and directly learns the change across each interval, instead of integrating many fine-grained local steps. At inference, it randomly perturbs part of the estimated k-space coefficients to explore alternative transition paths, with each state reconciled to the acquired measurements through data consistency. Experiments on multi-coil MRI datasets report better results than competing methods, with computational efficiency comparable to few-step baselines.
Muhammad Usama Mirza, Esin \"Ozt\"urk I\c{s}{\i}k, Tolga \c{C}ukur
In briefWhen Faster Decisions Don't Yield Better AI Agents: A Study of Jev in Agent Workflows
SummaryThis paper examines whether faster, cheaper decisions from the Jev model translate into efficient task completion for AI agents. The authors evaluated five agents across 4,500 task executions in ALFWorld (embodied tasks) and AndroidWorld (mobile app use). Jev substantially reduced decision cost and latency, but task success stayed near baseline on ALFWorld and dropped more in two mobile agents, partly because extra reasoning and interactions absorbed the savings.
In briefVolume rendering with metric-phase fields for thin and sign-ambiguous shapes
SummaryThis paper presents PhaseSurf, a neural volume rendering framework built on Metric-Phase Fields (MPFs) for 3D reconstruction. MPFs split geometry into a nonnegative metric field for unsigned proximity and a continuous phase field giving a soft sign cue, combined through residual phase injection. The authors report that on multi-view datasets with thin structures and sign-ambiguous geometry, PhaseSurf yields more stable and higher-quality reconstructions than UDF-based methods and recent 3DGS pipelines.
In briefHuman vs. LLM Facilitation in Design Meetings: A Behavioral Comparison
SummaryThis paper compares facilitation turns by human project managers with those generated by Llama 3.1 8B Instruct across 199 paired contexts from the AMI Meeting Corpus. Human turns scored higher on contextual semantic distance and hedge-related lexical density, while the model scored higher on topic divergence, lexical elaboration, and phase-vocabulary alignment. The authors state these differences should not be read as a ranking of facilitation quality.
Mohammed Azeez Khan, Aaron Dsouza, Ashutosh Mishra et al.
In briefSWEAR: Efficient Audio Reasoning That Adjusts Effort by Difficulty
SummaryThis paper studies how large audio-language models can reason more efficiently by matching effort to task difficulty. The authors derive five-dimensional difficulty scores and compressed reasoning traces from offline assessments, then propose SWEAR, a framework that learns difficulty signals during supervised fine-tuning and uses reinforcement learning to allocate reasoning across stages. Experiments on MMAU-test-mini and MMAR report improved accuracy with substantially fewer generated tokens.
In briefAttention Recognizes, MLP Refuses: How LLM Safety Pathways Work
SummaryThis arXiv paper examines why safety-aligned language models can be jailbroken or over-refuse harmless requests. Across 18 Base-Instruct model pairs from Qwen, Llama, and Gemma, the authors find harmfulness is more readily decoded from attention updates, while MLP updates contribute more to writing refusals. Causal interventions support this pathway. The authors introduce Sentinel, which combines attention-based and MLP-based scores and achieves a lower aggregate error rate than single-signal routing across eight models.
WhySeparates attention-based harm perception from MLP-based refusal writing, validated causally across 18 model pairs, offering useful diagnosis of jailbreaks and over-refusals.
In briefLongHarness: A Framework for Reasoning Over Huge Document Collections by Tracking Evidence and Conclusions
SummaryThis paper presents LongHarness, a model-agnostic agent framework for answering questions across document collections too large for a model's context window. Its core mechanism, Evidence-Conclusion Coevolution, uses unresolved conclusions to guide searches for missing information, while new evidence updates those conclusions. Evidence Lifecycle Management keeps accumulated evidence current and traceable. On benchmarks including OfficeQA Pro, AA-LCR, and LongBench v2, the authors report large exact-match gains over direct reasoning with the same backbone models and over Claude Code and Codex.
In briefDouble-Take: Improving Long-Context Memory Capacity, Binding, and Selectivity
SummaryThe paper examines why fixed-size memory becomes less usable as sequence length grows, identifying single-step writes that accumulate interference and self-associative writes that cannot form cross-position bindings. It proposes Double-Take, which uses multi-step unpinned updates, cross-position mixing, and bidirectional change-sensitive writes. Trained only on 32 kb genomic contexts, it reaches 0.927 AUROC at 131 kb and 0.856 near 1 Mb, and transfers zero-shot from human to mouse.
Vakul Raghavan, Utkarsh Grover, Mingyang Mao et al.
In briefAV-FlowEdit: Training-Free Prompt-Driven Audio-Video Editing
SummaryThis paper presents AV-FlowEdit, a training-free and inversion-free framework for editing existing audio-video clips from text prompts. It uses differences between source and target flows to update both the video and audio branches together, supporting visual-only, audio-only, and joint edits. A cross-modal event-aware spectral trajectory preservation mechanism suppresses drift during sampling so that non-target regions are kept. The authors report strong prompt alignment and temporal consistency, with competitive audio quality and cross-modal coherence.
In briefDISCERN: Identifiable disease evidence for generalizable medical multimodal learning
SummaryThe paper argues that clinical multimodal models can learn enrollment shortcuts shared across data types and fail on new patient cohorts. It models the data with a latent structural causal model and proposes DISCERN, which jointly trains an information-lattice encoder, a latent-state model, and an outcome head. On reversed synthetic enrollment, conditioning raised state AUROC from 0.043 to 0.883 and from 0.265 to 0.825. Exploratory eICU-to-MIMIC-IV mortality experiments improved AUROC over matched empirical risk minimization by 0.036 and 0.035.
In briefCTR-Driven Ad Image Editing via Attribute Disentanglement and Preference Alignment (ECC)
SummaryThe paper proposes Evolving Creatives with Clicks (ECC), which treats e-commerce ad creative optimization as fine-grained editing rather than generation from scratch, modifying only attributes that hurt click-through rate. It identifies 28 attributes correlated with CTR using ANOVA and Random Forest regression, instruction-tunes a prompt model for editing suggestions, and refines it with real click preferences. It also introduces Consistent Preference Optimization (CPO) to mitigate DPO overfitting. The authors report that ECC significantly improves CTR over state-of-the-art models.
In briefCB-Orchestrator: Contextual-bandit-based adaptive workflow optimization for LLM agents
SummaryThis paper proposes CB-Orchestrator, which separates workflow generation from workflow selection for LLM agents. It first builds a diverse pool of candidate workflows via evolutionary search, then treats selection as a contextual bandit problem to reduce required training feedback. At inference, the model picks a workflow suited to each query. The authors report it outperforms all baselines across five benchmarks and reduces training API token overhead on MATH.
In briefKV-Repurpose: A Testbed for Reusing KV Cache Beyond Exact Prefixes
SummaryLLM serving systems usually reuse cached computation only for identical prompt prefixes. This paper presents KV-Repurpose, an open-source vLLM-based system for developing and evaluating segment-wise KV-cache reuse beyond exact prefixes. In a Zipfian RAG-style workload at the same 30 GB cache capacity, it reaches a 63.9% token-level hit rate versus 3.3% for prefix caching, and prefix caching reaches only 17.9% even with 1 TB of cache.
In briefCliQ-RRG: Clinical-Knowledge Guided Disease-aware Image-Text Alignment for QA-Style Radiology Reports
SummaryThis paper presents CliQ-RRG, a two-stage framework that automatically generates radiology reports from chest X-ray images. The first stage aligns image and text features using predicted disease embeddings and a prior-guided attention module that links current and previous scans. The second stage uses graph-based retrieval of clinical evidence and a large language model to rewrite the content as concise question-answer pairs with diagnostic summaries. The authors report better results than prior methods on two public chest X-ray benchmarks.
WhyCombines disease-aware alignment, prior-guided attention and GraphRAG retrieval for QA-style reports, consistently beating prior methods on two public chest X-ray benchmarks.
In briefRewinding to Move Forward: Retrospective Exploration for Long-Horizon Agents (ReTRAIL)
SummaryThis paper introduces ReTRAIL, a framework that lets language model agents rewind the environment to earlier states while keeping valid progress and carrying knowledge from abandoned exploration forward. It trains a shared policy using State-Anchored GRPO and a branch-advancement reward. Experiments on four interactive benchmarks (MineSweeper, Sokoban, ALFWorld, ScienceWorld) show that adding a rewind operation alone gives limited benefit and can hurt performance, while ReTRAIL helps agents recover from unproductive states and reuse insights from abandoned branches.
In briefCoBe: A Benchmark for Conversational Counterfactual Text Editing
SummaryThis paper introduces CoBe, a benchmark testing whether language models can revise a text under a hypothetical change while keeping facts not causally affected by that change. It contains 2000 questions across healthcare, science and engineering, business, and everyday scenarios. Frontier models average 58.05%, and the authors find models often default to associational generation rather than preserving causal structure. Performance does not consistently scale with model size or output length.
Arvind Raghavan, Inwoo Hwang, Shreyas Havaldar et al.
In briefA4-MR: Anatomy- and abnormality-aware MRI foundation model
SummaryThis paper introduces A4-MR, a 3D vision-language foundation model trained on about 50,000 MRI studies, including roughly 491,000 series, with about half paired with radiology reports. It aims to learn anatomy localization, align image and text at the anatomical level, and represent healthy and abnormal anatomy distinctly. Across diverse downstream tasks, it outperforms existing foundation models, and with a frozen encoder it matches or exceeds task-specific segmentation models for knee and abdominal organ segmentation.
In briefUniMergeTrack: Hierarchical Model Merging for General RGB-X Object Tracking
SummaryThis paper proposes UniMergeTrack, a model merging method for RGB-X visual object tracking, which combines RGB with auxiliary modalities such as thermal infrared, event, or depth. Modality-specific trackers are trained independently, then merged hierarchically: backbones are combined using high-pass-filtered merging with lightweight task-specific experts, and tracking heads are merged using optimal transport-guided channel alignment. A distribution alignment fine-tuning stage follows. The authors report state-of-the-art or competitive results on six benchmarks and strong generalization to unseen datasets.
In briefNOFACE3D: A training-free method to sharpen fine details in 3D generation
SummaryThis paper presents NOFACE3D, a training-free framework that improves fine, thin, or detail-rich regions in 3D assets generated from a single image. It extracts a selected region from a generated mesh, refines it with flow-based sampling under a spatially varying noise schedule, and integrates it back while keeping the rest of the asset unchanged. Evaluated on PAniC-3D, TexVerse, and BigDetail, a new 339-asset benchmark, it reportedly produces sharper geometry and finer texture than baselines.
In briefFailure-resilient multi-robot coordination under connectivity constraints
SummaryThis paper studies multi-robot coordination for tasks such as coverage, target tracking, and information gathering, where robots must stay resilient to failures while keeping the communication network connected. It introduces Resilient k-Connected Submodular Maximization, which jointly considers task performance, resilience, and connectivity. The authors use a Menger-based connectivity reduction and a connectivity-aware greedy algorithm, and provide a worst-case post-failure performance bound relative to the optimal connectivity-respecting assignment. Experiments on multi-robot target tracking validate the method and show the practical effect of connectivity constraints.
Yoonjae Hwang, Hojune Kim, Ellen Novoseller et al.
In briefRedirecting Teacher Supervision: Temporal-Response-Guided Latent Distillation for Video-Language Models
SummaryThis paper studies on-policy distillation for video-language models and argues that direct hidden-state alignment does not use the student's temporal responses to pick a supervision direction. The authors propose Temporal-Response Guidance (TRG), which measures representation responses to local temporal perturbations, projects the teacher correction onto their span, and restores the original norm. On TempCompass, Video-MME, MVBench, and Video-MMMU, TRG improves the 4B student's macro accuracy by 2.86 percentage points over the reported output-only OPD baseline.
In briefRAMP: Mixed-precision method that speeds up prefill in recurrent-depth language models
SummaryRecurrent-depth language models improve reasoning by repeatedly applying a shared network, but this raises inference cost, especially during long-context prefill. The paper finds that quantization errors introduced early in recurrent depth are attenuated by later steps, so they affect the final representation less. Based on this, the authors propose RAMP, which uses low-precision weight-and-activation computation for early refinement and restores high precision for terminal refinement. On contexts up to 4K, RAMP keeps near-full-precision prefill quality and achieves up to 1.16x end-to-end first-prefill speedup on an NVIDIA H100, a 13.9% latency reduction.
In briefLightweight vibration sensing that turns signals into time-frequency images for appliance monitoring
SummaryThis paper proposes a lightweight AI framework that converts vibration signals from home appliances into time-frequency images using wavelet and Hilbert-Huang transforms. It analyzes these images with compact deep learning models such as MobileNetV3-Small, ConvNeXt-Tiny and YOLOv11-cls. Tests on public bearing datasets compare random splitting with structured isolation to avoid data leakage. The authors report that CWT with YOLOv11-cls gives high accuracy, while MobileNetV3-Small offers low latency for edge devices.
In briefEasier-to-Tune Training Algorithms: New AlgoPerf Benchmark Results
SummaryThis report presents updated results from the AlgoPerf V1 benchmark, which measures wall-clock time to reach a target across nine workloads on fixed hardware. It focuses on the self-tuning ruleset, which requires one configuration for all workloads, and finds that Schedule-Free AdamW ranked first among 13 PyTorch and JAX submissions. The authors also note that learning rate and momentum are highly sensitive across domains and algorithms, pointing to a need for methods that are more robust to tuning and generalize to new tasks.
Priya Kasimbeg, Ahmed Khaled, Aahlad Chandrabhatta et al.
In briefGeoGR: Compressing visual tokens to make 3D scene vision-language models faster
SummaryThis paper proposes GeoGR, a two-stage framework that reduces the cost of running vision-language models on 3D scenes. At the projector, GeoSemZip merges multi-view observations into compact 3D voxel tokens and keeps discarded features as contextual anchors. Inside the language model, GroupRoute lets low-relevance tokens skip layers and rejoin later, and the authors report retaining about 95% to 97% of dense performance with roughly 6 times faster prefill on two 7B models.
In briefTarget-constrained pooling for extrapolating partial trajectories
SummaryThis paper addresses extrapolating the continuation of a partial target trajectory using completed related trajectories. The authors propose target-constrained pooling (TCP), which estimates the finite-horizon endpoint from two complementary source signals and converts them into corrections of the free forecast. Across 25,245 targets from five domains, TCP reduces mean RMSE by 35.8% to 42.1% for all ten target-only forecasters.
In briefWhat Git Doesn't Record: How Readable AI Use Is in Research Repositories
SummaryThe authors audit two single-investigator research repositories by the same author, containing 466 and 103 commits, to see what they reveal about AI agent involvement. The two kinds of agent traces are legible in opposite ways, and the Co-Authored-By trailer covers only part of the AI-assisted commits and varies over time. The paper proposes an auditable evidence ledger and requirements for systems that record research provenance as it happens.
In briefLow-Rank Kolmogorov-Arnold Generator Flows for Interpretable Operator Learning
SummaryThe paper introduces KAFlow (Kolmogorov-Arnold flow), which learns a physical operator's generator as a low-rank symbol built from univariate Kolmogorov-Arnold functions, integrates dynamics with exact exponential integrators, and automatically distills a closed-form expression. On synthetic benchmarks it lowers out-of-distribution loss by 3.4 to 11 times and recovers coefficients to within 10^-4. On long-horizon quantum dynamics it reduces state infidelity by up to 6 orders of magnitude.
WhyProposes a low-rank Kolmogorov-Arnold generator with automatic closed-form extraction; cuts out-of-distribution loss 3.4–11x and quantum infidelity up to 6 orders, with broad benchmarks.
In briefWhy brain MRI AI models fail on new datasets: image content is encoded in different directions per dataset
SummaryThis paper studies why frozen features from brain MRI foundation models lose accuracy on new datasets. Across 17 public datasets and eight frozen representations, the models recognize MRI sequences within each dataset but mostly fail on unseen ones. The authors attribute this to image content being encoded along different directions in each dataset, not to dataset encoding or dataset shift. Removing dataset signals did not improve transfer, and only 3DINO and SAM-Med3D encoded the sequence consistently enough to transfer.
WhyShows cross-dataset failure stems from content encoded along different directions per dataset, verified across 17 datasets and eight representations.
In briefEvo²: Co-evolving model weights and agent harness for self-improving AI agents
SummaryThis paper proposes Evo², a framework in which a language model agent improves both its weights and its harness, the structure that governs prompts, tools, and context. A single policy acts as a Proposer that writes harness edits and a Reflector that diagnoses their effects, and a training pipeline called EvoTIDE combines SFT with interleaved GRPO. The authors report 82.1% on Polyglot and 48.3% on SWE-bench Verified, compared with 62.6% and 41.0% for the strongest baseline.
WhyNovel weight-harness co-evolution framework; large gains over strongest baselines on Polyglot and SWE-bench with cross-model transfer.
In briefWhat Visual Representations World Action Models Need: A Systematic Study of Vision Features for Robot Manipulation
SummaryThe paper notes that World Action Models inherit priors from video generators, but their VAE representations are optimized for reconstruction rather than manipulation. The authors compare visual representations, five integration mechanisms, and feature combinations under controlled settings, finding that strong general-purpose representations give broader robustness and that channel-wise joint prediction achieves the highest average success rate. The resulting R²-WAM reaches 84.7% on RoboCasa GR1 Tabletop and 92.8% on RoboTwin 2.0, improving over VAE-only baselines by 5.9 and 2.2 percentage points.
In briefUnified generative video codec with autoregressive and bidirectional diffusion modes
SummaryThis paper presents UniGVC, a generative video codec that decodes a single bitstream in two modes. The bidirectional UniGVC-BI prioritizes reconstruction quality for random-access playback, while the autoregressive UniGVC-AR targets low-delay streaming. On standard 1080p benchmarks, UniGVC-BI achieves the best rate-distortion-perception trade-off among the evaluated methods, at rates as low as 0.0005 bits per pixel. UniGVC-AR decodes 1080p video at 36.85 FPS, over 250 times faster than UniGVC-BI.
In briefDon't Guess, Just Ask: Evaluating Interactive Requirement Clarification for Repository-Level Development
SummaryThis paper introduces ReqClarifyBench, a benchmark for interactive requirement clarification in repository-level software development. It contains 341 tasks across game, miniapp, and web development, each with a vague user intent, a human-validated requirement document, and code evaluation artifacts. Using LLM-based user simulators with personas derived from real conversations, the authors evaluate six LLMs on clarification effectiveness and requirement document quality. They find that stronger clarification does not always yield better documents, performance varies across personas, and coding-agent harnesses do not improve document quality over their base LLMs.
In briefEmbodiedCLBench: A Benchmark for Continual Learning in Self-Evolving Embodied Agents
SummaryThis paper introduces EmbodiedCLBench, the first benchmark in embodied simulation for testing whether self-evolving agents can keep learning over time. It centers on compositional generalization, checking if agents can learn basic tasks and then combine those skills to solve unseen advanced tasks. Experiments show advanced tasks are hard to solve zero-shot, while experience from basic tasks gives consistent gains that hold as complexity rises. However, the gains do not grow with more learning samples, and current agents fail to extract useful information from additional experience.
In briefTrajectory-aware policy optimization that corrects state drift in long LLM reinforcement learning
SummaryThis paper addresses how small per-step policy mismatches in long autoregressive LLM rollouts can accumulate into large state distribution shifts across trajectories. It proposes TAPO, which derives a state-distribution correction from the off-policy policy gradient and adds it to the RL objective, using a geometric-mean estimator for variance reduction and length-aware dynamic clipping. The authors report that TAPO outperforms existing LLM RL methods on eight competition-level benchmarks, with gains of up to 48% on mathematical reasoning and 12% on algorithmic coding over GRPO, and more stable training.
In briefFlyHack: Stress-testing mechanistic claims in a connectome-constrained fly vision model
SummaryThis paper examines how reliable mechanistic conclusions are in flyvis, a connectome-constrained model of Drosophila motion vision. The authors compare it with task-matched CNNs using a 50-model ensemble and replicated interventions. They find that attribution results, circuit conclusions, and cell-type criteria change with the intervention, representation, and individual trained model. The authors conclude that such models make hypotheses testable, but claims should be limited to the specific model family, intervention, representation, and trained solution.
In briefGeneralized Proximal Policy Optimization: unifying LLM RL algorithms via divergence
SummaryThis paper proposes Generalized Proximal Policy Optimization (GPPO), a framework that unifies reinforcement learning algorithms for LLMs through the lens of divergence. It establishes a generalized f-divergence trust-region theory with a monotonic improvement guarantee and introduces estimator topologies that give more faithful divergence estimates. PPO and variants such as DAPO, DPPO, and RIPO are shown to be special cases within a two-dimensional Divergence × Estimator design space. On eight competition-level math and coding benchmarks, the best GPPO instantiation outperforms GRPO by up to 50% on math and 14% on coding.
In briefEvidential Transformer with Dempster-Shafer fusion for emotion recognition
SummaryThis paper addresses multimodal emotion recognition models that stay overconfident when sensor data are corrupted or contradictory. It proposes EMT-DS, which represents predictions as Dirichlet-based subjective opinions, separating epistemic from aleatoric uncertainty, and fuses modality evidence with Dempster's rule while quantifying conflict. The authors report higher accuracy and lower calibration error on several benchmarks, and that the learned conflict metric flags most misclassifications in a small fraction of samples.
In briefLearning Features, Not Layouts: Feature-Axis Attention for Medical Transfer
SummaryThis paper argues that standard vision transformers apply global attention over spatial tokens, which suits natural scenes but may not suit medical images, where discriminative signals lie in feature correlations and texture. The authors propose FAD-ViT, which strictly decouples spatial and feature axes and makes feature-axis attention the primary global operator. Across 23 classification, segmentation, and low-data tasks, from-scratch performance is essentially unchanged, but transfer from ImageNet pretraining is markedly stronger, with fewer parameters and lower compute than baselines.
In briefUpdating AI beliefs by evidence source, not repeat count
SummaryThe paper studies repeated human-AI interaction, where one observation can return to a model as many quoted or summarized reports that look like independent confirmation. The authors propose Latent Quotient Evidence Training (LQET), which combines sparse Bayesian supervision with constraints that keep predictions stable when reports add no new evidence. On a controlled binary benchmark across two backbones, LQET reduces new-evidence update error and log-odds drift compared with task-only training, and the gains extend to a closed-loop simulation.
In briefScaling Empowered Agents: Maximizing Environmental Control Through Skill Learning
SummaryThis paper examines the connection between empowerment, an agent's capacity to actively control its environment, and skill learning. The authors propose a simple, scalable method that treats empowerment as the number of executable skills from a state, so maximizing it means reaching states that serve as good starting points for many skills. The approach works with arbitrary underlying skill learning methods. Experiments show emergent behaviors, such as building a house in Crafter.
In briefHindsight Structured Advice Distillation: a small advisor model that writes per-input guidance for a frozen agent, trained on its failures
SummaryThe paper addresses language-model agents that are capable but miss task-specific details in fields like finance, where retraining is costly or impossible. It trains a small advisor model to read each input and write structured advice that is appended to a frozen executor's prompt, using a teacher that sees grades and execution traces to produce helpful advice, which the advisor learns to imitate. On the FrontierFinance benchmark, an 8B advisor raised a proprietary research agent's qualification rate from 52.9% to 56.6% at 37% lower cost per query. The authors also report improvements on WebShop, BIRD text-to-SQL and ExpertLongBench.
In briefMeasuring LLM creativity by performance on rare input-output pairs
SummaryThe paper proposes that LLM creativity can be formalized as task-specific performance in the long tail of a reference input-output distribution. Grounded in the novelty-quality product formulation, it evaluates 15 benchmarks and 26 open-weight models, using n-gram absence from pretraining data as a novelty proxy. The authors introduce q@Tail(α), a lower-cost creativity diagnostic measuring mean task quality on the tail fraction of a benchmark.
Yichen Wang, Tenghao Huang, Vishakh Padmakumar et al.
In briefBenchmarking embedding shapes and clustering methods for classifying patent claims without labels
SummaryThis paper asks whether an unsupervised model trained on patent text can find relevant prior art that code-restricted search misses, and whether hyperbolic space helps. The authors trained hyperbolic and Euclidean Variational Autoencoders on 45,000 US patents. The hyperbolic model retrieved cross-category prior art more often (18.76% vs. 12.34% of top-10 results) but agreed less with official classification. The study also found that a domain's classification breadth predicts how well the approach performs.
WhyCompares hyperbolic and Euclidean VAEs for patent retrieval, showing cross-category gains with geometric explanation across 96 runs and 3,000 queries.
Sujith Subramanian, P N Bhargav Teja, Susen Sarkar et al.
In briefWhat SFT Should Learn: Conditional Information Gain for Generalizable Post-Training
SummaryThis paper proposes CIG-SFT, a supervised fine-tuning method for large language models that estimates how much information the task input provides for each demonstrated token. It compares token likelihoods from a frozen reference model with and without the task input, and uses the resulting conditional information gain to weight the training objective. The authors report that across math reasoning and code generation benchmarks, CIG-SFT outperforms standard SFT and recent adaptive baselines, especially on harder evaluation distributions, without needing extra data or online rollouts.
In briefTighter regret bounds for using offline data in linear bandits
SummaryThis paper studies how offline data can improve online learning in stochastic linear bandits, focusing on partially covered data where only k of d feature directions are poorly covered. By applying concentration of Lipschitz functions of Gaussian vectors, the authors obtain tighter control of reward prediction error and a phased elimination algorithm with regret of order O~(k√T). They prove matching lower bounds under partial coverage and uniform coverage, and also provide matching bounds on simple regret for pure exploration with offline data.
WhyImproves regret from √(dkT) to k√T under partial offline coverage, with matching lower bounds and rigorous proofs.
Kushagra Chandak, Volodymyr Tkachuk, Xiaoqi Tan et al.
In briefLanguage models contain interpretable components that shape narrative content.
SummaryThis paper examines whether language models have internal components linked to narrativity, the degree to which a text has narrative properties. The authors identify attention heads, MLP neurons, and sparse autoencoder features in GPT-2-Small, Gemma-2-2B, and Llama-3.1-8B whose activations correlate with narrativity. Ablating these components raises next-token prediction loss more for high-narrativity texts and reduces the narrativity of open-ended generations.
David Yiyang Liu, Aditya Joshi, Paul Dawson et al.
In briefPredicting how much masked-text predictions change across different text corpora
SummaryThe paper asks how much a masked-text reconstruction predictor's outputs would change if a second corpus were collected from the same population. It uses a smoothed predictor built from nearby token-pair counts over GPT-2's 50,257-token vocabulary and derives a first-order formula for expected disagreement between predictors fitted to two independent corpora. The formula can be evaluated from a separate reference corpus. On TinyStories, WikiText-103, and CNN/DailyMail, growing corpora from 512 to 2,048 chunks cut disagreement by only 5–40%, well short of the 75% implied by 1/n scaling.
In briefFast training for models with sudden random jumps in hidden states
SummaryThe paper introduces latent jump models (LJMs), in which a hidden state evolves by closed-form deterministic dynamics between Poisson-driven event times and undergoes a learned instantaneous jump at each event. The authors derive an event-augmented path-space evidence lower bound and propose Scan-Parallel Variational Inference (SPVI), which uses triangular affine scans to train in logarithmic temporal depth. On synthetic stochastic systems and real irregular time series, the authors report that LJMs are competitive with continuous-time latent baselines in forecast accuracy while training one to two orders of magnitude faster.
In briefNucleoNet: A Nuclear-Aware Wavefunction Model for Electron Energies Across Molecular Shapes
SummaryThis paper presents NucleoNet, a neural wavefunction for variational Monte Carlo that treats electrons and nuclei as one particle set and jointly evolves their representations. On equilibrium small-molecule benchmarks it reaches accuracy comparable to established neural baselines with substantially fewer parameters. A single model optimized across geometries describes the N2 dissociation curve, Li displacement in LiCN, and bond-dissociation minimum-energy paths for ethane and formamide.
In briefWork2Skill: Turning Human GUI Demonstrations into Reusable Agent Skills
SummaryThis paper introduces Work2Skill, a framework that records human desktop GUI experience and induces reusable skills for computer-use agents. The authors built WorkDemo-350, a synchronized multimodal dataset of 350 screened human demonstrations. Their agentic skill-induction method produces skills covering workflow rules, step dependencies, and completion checks. On MyPCBench Bounded Action and a 286-task OSWorld-Verified subset, skills improved mean scores over a No Skill baseline by 4.67–10.54 and 3.11–7.50 points.
In briefEvidence Anchoring Boundary: When Visual Interventions Help Multimodal LLMs
SummaryThis paper studies training-free visual interventions intended to reduce hallucinations in multimodal large language models. It finds that the intervention's effect drops persistently at a certain decoder layer, which the authors define as the Evidence Anchoring Boundary (EAB). The paper also compares counterfactual subtraction across presence, attribute, and relation questions, and proposes calibrated selective routing, which improves full-set attribute accuracy by 1.56 to 1.81 percentage points across three checkpoints.
In briefLearning Representations to Guide Pixel Diffusion Transformers
SummaryThis paper studies which visual representations make images easier to generate in pixel diffusion transformers. It finds that better representation quality does not guarantee better generation, and that specialized pixel decoders mainly redistribute semantic and spatial modeling. The authors propose Representation Guidance (RG), which uses a low-dimensional bottleneck to form generation-relevant representations and guides decoding by modifying the flow-matching target. RG reaches an FID of 1.55 on ImageNet 256x256 with RG-XL/16, and 1.74 at 512x512, and also extends to text-to-image generation with GenEval 0.89 and DPG-Bench 87.1.
In briefBeyond Semantic IDs: Continuous-query generation for recommendation that avoids beam-search limits
SummaryGenerative recommendation usually retrieves items by decoding discrete Semantic IDs with finite-width beam search, which can keep held-out targets out of the returned list. The authors propose CQG-Single, which generates one continuous retrieval query in a single pass, and CQG-AR, which autoregressively generates continuous residual updates, so every encoded item can be scored via dense or approximate-nearest-neighbor retrieval. Under a unified protocol, CQG-Single outperforms the TIGER-style baseline in all seven runs while using 55.3% as many trainable parameters.
In briefDiffusion-Based Daily Precipitation Generation for Medium-Range Weather Forecasts
SummaryThis paper proposes a plug-and-play framework that uses a separately trained diffusion generator to map non-precipitation forecasts from a frozen weather model into observation-aligned daily precipitation, avoiding costly retraining of weather foundation models. The generator is conditioned on atmospheric states and satellite-gauge precipitation estimates, using a multi-scale recurrent encoder and cross-attention. Global ensemble evaluation reports lower Brier scores than ENS and GenCast trained from scratch at all reported lead times, with lower CRPS in most comparisons. A heavy-rainfall case study using 2,288 stations in China further reports improved daily accumulation patterns and event detection.
