In briefLess from More: Learning Sparse Video Reasoning from Dense References
SummaryThis paper argues that video-language models need to reason reliably from limited temporal evidence, not just process more frames. It proposes SAVER, a post-training framework that uses dense video views as training-time references to guide sparse-frame predictions. SAVER is trained on only 1,250 randomly sampled temporal grounding examples and improves results across three temporal grounding and six video question-answering benchmarks. It can match or surpass dense-frame Qwen3.5 baselines while using substantially fewer frames.
In briefFactorization-guided training and inference to speed up diffusion-drafted speculative decoding
SummaryThis paper proposes DScout, a method that speeds up large language model inference through speculative decoding with diffusion drafters. It uses two complementary factorizations to guide drafter training and to select inference configurations that maximize throughput. Across six benchmarks, DScout reports mean speedups of 3.84x and 4.14x over autoregressive decoding on Qwen3-4B and Qwen3-8B, respectively, exceeding the state-of-the-art baseline by 9.1% and 7.1%.
In briefSPLIT-RL: Staged reinforcement learning that trains visual and language reasoning separately
SummaryThis paper proposes SPLIT-RL, a staged post-training method that trains visual reasoning and language reasoning in separate phases for vision-language models. It also introduces CLA-GRPO, which splits visual reasoning into atomic visual claims to give finer-grained advantages. The authors report that SPLIT-RL improves average accuracy over GRPO across several Qwen3-VL models and InternVL3.5-8B. They also state that answer-only GRPO leaves perception unchanged, while SPLIT-RL improves both visual and language reasoning.
In briefDecentralized collaborative continual learning: a multi-objective minimization method for agents that learn new tasks without forgetting old ones.
SummaryThe paper formulates decentralized continual learning as a multi-objective optimization problem, where agents compute locally and exchange information only with neighbors as tasks arrive sequentially. Agents store subsets of past-task samples in local memory buffers and incorporate them into updates to retain earlier knowledge while adapting to new tasks. A mean-square-error analysis indicates that cooperation among agents improves continual learning performance. Simulations show reduced forgetting and improved average mean-square deviation across tasks.
In briefStochastic Teacher Intervention for Agentic On-Policy Distillation (STI-OPD)
SummaryThis paper proposes STI-OPD, an on-policy distillation method for multi-turn agentic tasks in which a teacher model occasionally replaces the student's proposed action. Intervention is sampled stochastically, with its probability derived from the KL divergence between teacher and student policies, and training uses an Importance-Weighted Reverse KL objective to correct for mixed-policy sampling. The authors report that STI-OPD outperforms the strongest prior OPD baseline on every evaluated benchmark and student size across tool-integrated reasoning and long-horizon interaction tasks.
In briefEvaluating Generative Molecular Dynamics: Looking Under the Streetlight
SummaryThis position paper argues that generative molecular dynamics (GenMD) research is not evaluating its progress properly, so models can match reported results while being wrong about what they claim, risking overstated advantages over classical baselines. The authors attribute this to the high computational cost of the true reference, exhaustive molecular dynamics simulations, which pushes evaluation toward convenient proxies that become entrenched norms. They propose concrete observables, reporting standards, and computational and experimental benchmark designs.
Simon Olsson, Frank No\'e, Grant M. Rotskoff et al.
In briefBarron Optimal Transport I: a transport framework where neural network complexity sets the cost for generative modeling.
SummaryThis paper introduces an optimal measure transport framework in which the cost reflects how complex a neural network must be to represent the transport. It replaces the kinetic L2 energy with the Barron energy, defining a new metric on probability measures that complements Wasserstein and Stein geometries. Applied to generative modeling, the authors prove super-polynomial lower bounds on score approximation for diffusion models when the data are generated by neural network pushforwards of a Gaussian, quantifying their suboptimality in this geometry.
In briefSparse attention as matrix approximation, not picking large values
SummaryThis paper addresses the quadratic cost of attention in large language models, which grows with prompt length. It argues that existing sparse attention methods wrongly keep only large attention entries, ignoring that the attention matrix is multiplied with value vectors as a structured matrix. The authors propose Matrix Approximation Sparse Attention (MASA), which scores each sparse unit by how much it reduces matrix-product approximation error, and can be added to existing frameworks. Experiments across several methods, benchmarks, and backbones report consistent accuracy gains.
In briefVariance-reduced MCMC via learned change of variables: builds explicit control variates by transforming samplers into a latent space for more precise estimates.
SummaryThe paper addresses variance reduction for MCMC estimates, noting that good control variates require solving the sampler's Poisson equation, which rarely has a closed form. The authors train a bijection, such as a normalizing flow, so the target becomes close to a reference density in latent space, and show that Markov kernels and their Poisson solutions transform accordingly. Running samplers in latent space yields explicit control variates with consistency under mild tail conditions. Experiments on synthetic and real posteriors compare the method with state-of-the-art samplers and control variates.
In briefVoice-aesthetics model for conversational speech, trained on human listener judgments
SummaryThis paper introduces CVAM, a speech large language model that describes the voice aesthetics of real or synthetic speech replies in natural conversations. It points out salient moments in the voice and predicts nine categorical attributes, including gender, pitch, pacing, emotion, and delivery. The model was trained on about 3,000 CANDOR-derived responses, each annotated by roughly 10 humans, first through supervised finetuning and then with Group Relative Policy Optimization on human judgments. Experiments reported that CVAM agrees with human listeners more than Gemini 3.1 Pro and open-source speech LLMs.
In briefA sawtooth-shape measure of automaton rules that predicts network classification performance
SummaryThis paper addresses the costly task of choosing transition rules for Life-Like Network Automata, which are used to extract dynamic descriptors for classifying complex networks. The authors introduce a property called "jaggedness," which measures how closely a rule's transition function resembles a sawtooth shape, and argue it serves as a proxy for chaoticity and sensitivity. A heuristic search guided by this measure reportedly reaches accuracy within 5% of the global optimum while cutting computation by 90% compared with exhaustive search.
Lucas C. S. Oliveira, Michiel Rollier, Jan Baetens et al.
In briefSeparating Language and Voice Traits in Speech Models Using Routed Sparse Autoencoders
SummarySelf-supervised speech encoders mix linguistic content with speaker identity, emotion, and prosody in one representation. The authors combine a TopK sparse autoencoder with route-specific supervision and cross-factor adversaries to split these factors into separate routes. On frozen SPEAR and WavLM encoders, independent probes show linguistic information stays in the linguistic route while paralinguistic factors are kept in their own route and reduced in the other.
Beimnet Bekele Guta, Xiaoyu Yang, Guangzhi Sun et al.
In briefA mesh-based neural energy method for simulating heterogeneous composites
SummaryThis paper introduces the mesh-based neural energy method (M-NEM), which discretizes the displacement field on a mesh using shape functions to suppress non-physical oscillations seen in collocation-based neural energy methods. It computes strains with algebraic shape function derivatives instead of automatic differentiation and uses high-order Gaussian quadrature for energy integration. On a directly comparable benchmark, M-NEM reduces stress errors by up to three orders of magnitude and runs one to two orders of magnitude faster, and it is the only method that converges on two benchmarks with extreme stiffness contrasts.
Pius J. M. Wichmann, Stefan Hildebrand, Sandra Klinge
In briefUsing LLMs to predict product conversion rates from browsing sessions, comparing summary and session-level methods
SummaryThis paper studies how to use a large language model to simulate a user population and predict product-level conversion rates from e-commerce browsing sessions, without task-specific training. It compares summarizing sessions before inference against predicting per session and averaging, and evaluating products separately versus together in one prompt. Comparative summaries achieve the highest Spearman correlation with observed rates in 12 of 15 settings and cost a small fraction of session methods. With the best model for each outcome, they reach 66-89% of the correlation of a supervised benchmark trained on thousands of labeled products.
In briefPreference-driven unlearning for few-step distilled text-to-image diffusion models
SummaryText-to-image diffusion models are often distilled into few-step versions for fast inference, but they can still generate harmful content. The paper argues that existing data-driven unlearning methods rely on multi-step denoising, so they forget poorly in few-step distilled models, and unlearning the base model then re-distilling is costly. The authors propose a preference-driven framework that modifies Direct Preference Optimization (DPO) to fit few-step generation. Experiments on identity and NSFW removal, plus object-level unlearning, report consistent forgetting and strong retention of non-targeted capabilities.
In briefRFChipAgent: A Multi-Agent AI System for Designing Analog and RF Chips
SummaryThis paper presents RFChipAgent, a flow in which cooperating large language model agents automate end-to-end analog/RF circuit design under human supervision. It combines document-based knowledge retrieval, topology selection with schematic and testbench generation, closed-loop circuit sizing using TPE and CMA-ES optimization, and a trust-scored simulation database. The authors validate it on GF22FDSOI 60 GHz wideband mm-wave low-noise amplifier topologies and report substantial reductions in design effort while maintaining signoff-quality verification.
Awani Khodkumbhe, Yunfei Feng, Raj Rangarajan et al.
In briefSelf-supervised keyframe discovery for behavior cloning with long-horizon memory
SummaryThe paper addresses behavior cloning in non-Markovian environments, where policies must use long-horizon context. It proposes Keyframe Mnemonics, a self-supervised method that learns to select information-critical past observations as keyframes, then conditions the policy on them. On synthetic memory tasks the method reaches 100% success and generalizes to horizons far beyond training. On a memory-intensive robot manipulation benchmark across 23 tasks, it improves average success rate by 13.9 points over the strongest baseline.
WhyProposes self-supervised keyframe discovery for long-horizon behavior cloning, validated across synthetic memory tasks, 23 manipulation tasks, and a real robot.
In briefKDFP: A first-principles approach to general knowledge distillation for large language models
SummaryThis paper presents KDFP, a white-box method for general knowledge distillation in large language models, which trains small student models using representations from larger teacher models. The authors note that prior LLM distillation work has mostly targeted post-training abilities, leaving general knowledge distillation under-explored. They report that KDFP outperforms existing methods by 1.6% to 4.9% across 9 benchmarks and improves training efficiency by up to 99.1% through ephemeral parameter reduction.
In briefSimilar Prediction Accuracy, Different Learned Brain Dynamics in Personalized Disorder Models
SummaryThe paper uses a lightweight CNN-Transformer EEG foundation model to extract representations, then fits sparse transition models to each of 198 subjects (99 with epilepsy, 99 without) to build personalized dependency graphs. Epilepsy subjects show significantly denser learned dependency structure, while held-out likelihood and next-state prediction are nearly identical between groups. The authors conclude that comparable predictive fit does not imply comparable learned dynamics, so personalized clinical models should be evaluated on learned structure as well as predictive performance.please
In briefAmortized Off-Policy Evaluation for LLMs (PFN-OPE)
SummaryThe paper proposes PFN-OPE, a method for evaluating candidate large language models offline, using data from already-deployed models rather than exposing users to unvetted ones. It pretrains once across many contextual-bandit tasks, then estimates a model's value in a single forward pass without refitting for each new dataset or reward definition. On HelpSteer2 and UltraFeedback with Qwen, Llama, and Gemma policies, it reports 2.0 to 9.3 times lower error than the best baselines in the reward-shifted settings.
In briefReal long-term memory for AI: a 50M-token store that beats recomputation on speed and energy
SummaryThe paper tests galahad-kv, a public package that saves each roughly 16,000-token block of a model's key-value state to encrypted local NVMe disk and reloads it byte-exact without recomputing. On one NVIDIA H100 with Gemma 4 12B and 31B, across 50 million tokens of real public text, every probed block reloaded successfully, running 2.8 to 4.3 times faster than recompute with 8.8 to 12.3 times less GPU energy. The authors note limits: this is reuse of stored state rather than a wider attention window, answer quality depends on the model, and the store needs terabytes of local disk.
In briefTerraAgent: Multi-agent reasoning for long-context remote sensing visual question answering
SummaryThis paper addresses remote sensing visual question answering, noting that current models are judged mainly by accuracy, which cannot show whether answers actually use the image. The authors propose AVA, a training-free framework with modules for demonstration retrieval filtering, multi-agent decomposition of questions into evidence-tracked sub-questions, and counterfactual auditing that yields a visual-dependency score. Across several benchmarks, AVA reports higher accuracy and lower rates of prior-only answers and hallucinations.
In briefOn the Clock: Making Time-Budgeted AI Agents Punctual and Productive
SummaryThis paper studies whether small LLM agents can respect explicit wall-clock time budgets while using the available time productively. The authors test Qwen3.6-27B on five MLE-Bench Lite competitions and Qwen3-4B on Zork I. Injecting timing information through the harness improves budget adherence without losing performance, and GRPO reinforcement learning nearly perfectly meets budgets on Zork I but does not improve MLE-Bench results. The findings show a gap between respecting time limits and allocating that time productively.
In briefMotherTree: Meta-learning Improves Decision Tree Training on Synthetic Data
SummaryThis paper introduces MotherTree, a tabular transformer that meta-learns decision tree induction. Pre-trained on a synthetic prior without reference trees, it produces a hard, axis-aligned decision tree for a new training set in a single forward pass. On benchmarks with controlled sample sizes, it is competitive with size-matched trees from common algorithms. As an initializer, task-specific tuning of its tree outperforms from-scratch gradient-based learning.
In briefChecking fewer labels for conformal prediction without changing the prediction sets
SummaryConformal prediction gives prediction sets with finite-sample guarantees, but verifying labels for calibration can be costly. The paper proposes a partial verification method that returns exactly the same prediction sets as full verification. It characterizes calibration certificates, the verified information needed to determine the conformal threshold, and coordinates verification across calibration examples. Across retrieval, math solutions, and configuration evaluation, it cuts verification cost by 15-82% compared with verifying examples one at a time.
In briefTraceICL: A Training-Free Method to Make Multimodal In-Context Learning More Robust
SummaryThis paper addresses weaknesses in multimodal in-context learning, where demonstrations can mismatch images, redundant examples crowd out useful ones, and noisy answers get copied. The authors propose AVA, a training-free framework with three stages: thread-resolved context decomposition, influence-scored demonstration selection, and weak-to-strong consistency calibration. On five vision-language benchmarks with interleaved 4-shot prompts, AVA improves the strongest same-setting baseline by 1.7 points on average, up to 2.4 points on VizWiz, at a 26% latency overhead.
In briefLarge-scale grammar concept annotation: fine-tuned small models beat prompted frontier models
SummaryThe paper addresses how corrective feedback from lessons rarely becomes a usable picture of grammar mastery. The authors fine-tune Qwen3.5 small language models on filtered, rebalanced teacher-generated data and deploy a 0.8B model in a grammar mastery tracker for English learners. On two human-curated benchmarks, the deployed 0.8B model and a 4B reference model outperform prompted GPT-5.4 and GPT-5.6 Sol in precision and recall, while reducing serving cost about 16 times. An online experiment reports a 15.8% gain in learner engagement.
Marjan Celikik, Ana Peleteiro Ramallo, Javier Morales
In briefVelocity Scaling in Flow Matching: Why It Helps Image Generation
SummaryThis paper studies multiplying a learned flow-matching velocity field by a gain to improve generation quality. The authors argue that MSE training does not cause a velocity-magnitude deficit, and instead find that scaling mainly reduces population time lag, where sampled states resemble training states from an earlier time. Selecting the gain using measured time lag reduces FID from 28.0 to 12.2 on ImageNet-256 at NFE 25 without guidance, estimated by interpolation between neighboring gains.
In briefFixing robot skill handoffs when the scene is left out of place
SummaryThe paper studies why chained robot manipulation skills fail at the handoff, when each downstream skill starts from a state its predecessor left behind (Observation-Space Shift). Using simulator resets, the authors find the main cause is displaced scene objects rather than the robot's joints or the manipulated object. They propose a detect-restore-resume system that monitors task progress, restores displaced scene components, and fine-tunes the skill to resume. On the BOSS-44 benchmark, full-chain success rises from 7.6% to 26.5%, and a preliminary real-robot test on a Franka arm is also reported.
In briefMeasuring LLM-Generated Rubric Quality via Interventional Transfer
SummaryThe paper proposes Interventional Transfer (IT), a method for evaluating how well LLM-generated rubrics match expert rubrics, based on whether scores move together when responses are perturbed to pass or fail. In a HealthBench case study, degradations transferred between generated and expert rubrics, but improvements did not reliably transfer. The authors argue this asymmetry matters for using LLM-generated rubrics for performance monitoring and hill-climbing.
Erik Skalnes, Layne C. Price, Raviteja Anantha et al.
In briefA controlled study of when to request a third signal or abstain when three scores conflict
SummaryThe paper studies a benchmark where a policy sees two of three conflicting scores (audio, video, text) and decides whether to pay to request the third or abstain. It compares a threshold policy with majority and always-answer baselines under mechanism-specific rewards and matched budgets. The authors find that selection value depends on the budget and on how observed score pairs are ranked.
In briefSymbolic Density Estimators for Unnormalized Distributions: finding formulas for unnormalized probability distributions from data
SummaryThis paper proposes a framework that estimates symbolic (analytical) expressions for unnormalized distributions from observed samples. It uses domain-specific prior knowledge, such as the range of interactions and a predefined set of primitive functions. The authors combine deep generative models, including flow and score-based models, with symbolic regression, adding inductive biases like factorization to keep the problem tractable. Experiments on multivariate toy distributions and on the XY model and φ⁴ theory lattices show the approach works, including estimating compact symbolic Hamiltonian approximations at different scales for renormalization in φ⁴ theory.
Vikas Kanaujia, Riyansha Singh, Shashank Sharma et al.
In briefWhose Ground Truth? Embracing Ambiguity in Human-Centered AI
SummaryThis position paper argues that many human-centered tasks have ambiguous interpretations, with multiple reasonable judgments for the same input. It notes that conventional AI assumes a single ground truth and aggregates or discards annotator variability. The authors call for modeling the space of plausible human judgments while distinguishing meaningful ambiguity from annotation noise.
Jingyao Wu, Mohammad Tariqul Islam, Per R{\aa}dberg Nagb{\o}l et al.
In briefHow many repeated pairwise comparisons does ranking need when preferences differ?
SummaryThis paper studies how to rank options by their population-average utility from pairwise comparisons when preferences vary across users and tasks. The authors note that even with many users, a single comparison per user may fail to identify the option with the highest average utility. They propose an MLE-based algorithm needing Ω(1/Δ²) repeated comparisons per context, two MLE variants and a randomized Russian Roulette-style algorithm needing O(log(1/Δ)), and prove the logarithmic dependence is optimal. The Russian Roulette algorithm uses O(1) comparisons per context in expectation.
In briefCalibrating Ambiguity Sets with Diagnostic Transport for Robust Decision-Making
SummaryDistributionally robust optimization protects decisions against uncertain probability distributions by optimizing over an ambiguity set, but a poorly shaped set can force overly conservative choices. This paper proposes diagnostic-transport DRO (DT-DRO), which uses held-out calibration data to reshape the ambiguity set according to observed prediction errors, yielding a tractable dual formulation. The authors derive valid ambiguity radii and decision-risk guarantees, and show the method can remove the robustness floor caused by model misspecification. Synthetic experiments and a power-outage application indicate better decision quality under structural and tail misspecification.
In briefNavGPT-3: A framework linking language-model reasoning and robot action for navigation
SummaryNavGPT-3 is a harness that connects a language model's reasoning with a low-level action policy, running reasoning, acting, and monitoring as scheduled threads in an OS-like runtime. The action policy, NavGPT VLA, was trained on 19.28M examples. The authors report state-of-the-art success on R2R-CE (81.51 SR) and human-level success and path fidelity on RxR-CE, at 1 min 22 s per episode versus about 3 min for a human.
In briefPlan-and-Patch: Diffusion Language Models for Agentic Planning
SummaryThis paper introduces Plan-and-Patch, a framework in which a diffusion language model generates structured plans in parallel and repairs only the affected steps while keeping the surrounding steps fixed. On Natural Plan without task-specific training, the diffusion planner achieved a 53.7% plan repair success rate versus 27.0% for the autoregressive model Qwen3-8B. On ALFWorld and TextCraft after training, both planners showed similar plan generation success, while the diffusion model reduced mean generation latency by 39-46%.
In briefVICO: Co-evolving visual environments to improve vision-language model reasoning
SummaryThis paper proposes VICO, a framework that trains a vision-language model and an environment rewriter together, so the visual training tasks change as the model improves. The rewriter edits verifiable image structures such as scene graphs, chart tables, and protected masks, then re-renders them into training samples whose difficulty is tuned to the model's pass rate. Across nine multimodal benchmarks, VICO-8B gains up to +5.0% on out-of-domain tasks and is comparable to chart-specialized RLVR methods while using 16 to 160 times fewer labeled samples.
WhyProposes co-evolving actor and visual environment with pass-rate rewards; gains over baselines on nine benchmarks give solid evidence and a scalable path beyond static RLVR.
In briefMemoWM: How World Models Change What Agents Need to Remember
SummaryThis paper addresses how long-term AI agents can limit memory storage as they accumulate experience, using world models that capture reusable regularities. It introduces MemoWM, which compresses retained information with shared predictions, reconstructs omitted content, and allocates memory by balancing reconstruction error against storage cost. Across five long-term agent-memory benchmarks, the authors report 42.42% average accuracy and a 53.9% reduction in per-experience storage relative to MIRIX.
In briefNEMORA: learnable equivariant multipole operators for long-range atomic modeling
SummaryThis paper introduces NEMORA, a neural extension of the Fast Multipole Method that learns equivariant long-range tensor interactions for machine-learned interatomic potentials. It operates on an adaptive spatial hierarchy with linear time and memory complexity and can handle systems of hundreds of thousands of atoms. On non-local benchmarks, it reduces force and energy errors relative to short-range backbones by over an order of magnitude and up to three orders of magnitude, respectively.
WhyLearns equivariant FMM-style multipole operators with linear cost, retaining physical priors; reports large force and energy error reductions on non-local benchmarks.
Jay L. Kaplan, Samuel Varner, Rebecca Willett et al.
In briefAI-Mediated Self: How HCI Defines and Relates to the Self
SummaryThis scoping review analyzes 102 papers on how human-computer interaction (HCI) defines the self and how AI-self relationships are conceptualized. It identifies risks that arise when AI becomes entangled with selfhood, grouped into four ethical domains: agency/autonomy, identity/authorship, relational capacity, and meaning-making. The authors introduce the concept of AI-mediated self and the Inclusion of AI-Self framework, which places AI-self relationships along a spectrum of proximity.
In briefReinforcement learning for energy transition investment decisions to create long-term value
SummaryThis paper studies how energy companies should allocate investment across oil and gas, renewables, and CO2 reduction during the energy transition. The authors built a simulation of the energy landscape up to 2050 and a multi-criteria sequential decision framework with three goals: maximizing profit, minimizing CO2 social costs, and strengthening competitiveness in renewables. A reinforcement learning agent learns investment policies under uncertainty, and the paper reports that it consistently outperforms manually defined baseline policies in adaptability and long-term value creation.
Yasaman Cheraghi (1), Reidar B. Bratvold (1), Aojie Hong (2) et al.
In briefCan linear attention learn from nonlinear teachers in context?
SummaryThis paper extends the theory of linear attention for in-context learning from linear regression to nonlinear single-index targets y=f(x·w)+ε. The main result shows that linear attention extracts only the linear Hermite component of f, while the remaining nonlinear structure acts as effective noise in the generalisation error. This lets results from the linear theory transfer to nonlinear tasks, and the paper examines finite pretraining data and the shift from task memorisation to generalisation as task diversity grows.
In briefDOGS: Design-space sampling for prompt-driven logo generation
SummaryThis paper argues that prompt optimization for text-to-image generation, which usually rewrites text, is ill-suited to structured logo design. The authors propose DOGS, which mines a graph-structured design grammar from real-world logos and uses a GFlowNet sampler guided by a reward combining recognizability, aesthetics, and originality. According to the paper, compared with nine baselines on two open-source renderers, DOGS produces logos that are more recognizable, aesthetic, diverse, and less prone to trademark infringement.
In briefMESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Next-Frame Forecasting
SummaryThis paper addresses estimating 3D motion in dynamic scenes (scene flow). Earlier two-frame methods lack long-term temporal context and extrapolate poorly, while multi-frame methods are computationally costly and degrade over long horizons. The authors propose MESSENGER, which stores past flow estimates and latent states in a memory buffer and uses an uncertainty-aware reweighting module to filter unreliable retrievals. Experiments on nuScenes and Argoverse 2 report EPE3D reductions of 71.6% and 67.7%, respectively, in long-horizon future extrapolation.
In briefClarify, Then Focus: Normalizing Conversation Statements for Large-Scale Analytics
SummaryThis paper addresses the cost of repeatedly interpreting millions of enterprise conversations for many analytics questions. It proposes converting dialogue into short, speaker-attributed statements with source references and semantic tags, then selecting relevant evidence per question. In an offer-suppression task on customer-service calls, normalization improved a supervised classifier without selection, while weaker prompted readers benefited from both normalization and selection. The authors argue that shared preparation supports an inference pipeline built entirely from small models, making large-scale conversation analytics substantially less expensive.
Mikhail L. Arbuzov (Independent researcher), Karan Dave (Independent researcher), Evgeniya Dontsova (Independent researcher) et al.
In briefCORAL-KV: Compact Cross-Layer Prediction for Compressing Video Model Memory
SummaryThis paper addresses the growing key-value cache that makes memory a bottleneck in autoregressive video diffusion models. CORAL-KV predicts keys and values in each layer from earlier layers' cache states and quantizes the remaining residuals, using compact block-diagonal-plus-low-rank predictors. Evaluations on HY-WorldPlay and MAGI-1 report higher PSNR than QVG baselines while using 33.7% to 48% less cache memory. The authors also report better fidelity over long generations with comparable latency.
In briefTraceQuery: Learning which positions to recompute from full denoising trajectories in diffusion language models
SummaryDiffusion language models generate text through repeated masked denoising, but recomputing token representations at every step makes inference costly. The paper proposes TraceQuery, which uses full-denoising trajectories to guide fresh computation toward a small set of near-future, trajectory-relevant positions. The backbone model then selects which candidate to commit from these fresh predictions. The authors report that TraceQuery gives the best quality-computation trade-off among cache-based methods and generalizes across tasks, model variants, and families.
In briefDialogue-based task disambiguation over tables: leakage-aware framework, benchmark, and training
SummaryThis paper studies how an agent can ask clarifying questions to resolve an unclear user task before solving it over tables or databases. It proposes a framework for ambiguous verifiable tasks that separates the asking policy from the solution policy and defines formal oracle leakage and diagnostics. In text-to-SQL, the authors build AmbiTab, a benchmark unifying six ambiguous datasets, and train the asking policy with reinforcement learning. The trained asker improves disambiguation metrics on all six datasets and task success on five.
Nafiseh Ghoroghchian, Luis Scoccola, Tina Sedaghat et al.
In briefLossy Compressive Text Autoencoders: learning compact latent text representations at the intersection of compression and representation learning
SummaryThe paper proposes an autoencoder that downscales and upscales hidden text representations along the time axis, with a residual low-dimensional discrete bottleneck. It compares different quantization methods, training objectives, and datasets, evaluating reconstruction at the surface level with BLEU and at the semantic level with an LLM-based judge across compression levels. The models are also tested on downstream question-answering and semantic text similarity benchmarks. On web text, the authors report compressed representations at 2.24 bits per byte, on par with lossless text compression algorithms, with good reconstruction and downstream performance.
Vinko Sabol\v{c}ec, Angelos Katharopoulos, David Grangier
In briefCognitive Thermometers: Machine Learning and Logical Complexity
SummaryThis paper asks how the human mind represents semantic categories and why natural languages favor certain meanings. It argues that earlier explanations based on logical definability are sensitive to the choice of logical language, and proposes machine learning as a more agnostic way to measure semantic complexity. The authors review evidence that logic and machine learning often agree on relative complexity, and that where they diverge, learning better explains semantic typology. They propose treating machine learning models as "cognitive thermometers" to bridge symbolic logic and connectionist AI.
In briefLinSlot: Discovering object attributes without supervision using linear concept directions
SummaryThis paper studies learning disentangled representations of objects and their attributes from raw image data without labels. It proposes LinSlot, which uses the Linear Representation Hypothesis to model slots (objects) and blocks (attributes) in a probabilistic framework optimized via the Evidence Lower Bound. Experiments show disentangled object and attribute discovery, evidence for the hypothesis in slot space, image editing ability, and improved DCI scores over state-of-the-art methods on multiple datasets.
Sanket Gandhi, Utkarsh Giri, Varun Subramaniam et al.
In briefRemoving Eyeglass Reflections While Preserving the Eyes: A New Benchmark and OcuFlow Method
SummaryThis paper introduces OcuBench, a benchmark for eyeglass reflection removal with 10,280 controllable synthetic pairs, 732 real-input pseudo-pairs, and 458 real-world test images. It also proposes OcuFlow, which uses ocular-adaptive pixel MeanFlow to concentrate reconstruction on reflection-obscured eye regions while preserving observed details at native resolution. Experiments report advantages in removal quality, ocular fidelity, and efficiency, and code and dataset are to be released.
In briefInfinite context windows in recurrent networks via spatial neural computing
SummaryThe paper proposes a second-order recurrent model that replaces direct neuron-to-neuron communication with a spatially evolving field governed by discretized partial differential equations, so structured spatiotemporal patterns act as implicit memory. The authors show the model is equivalent to a structured infinite-order RNN with an effectively unbounded receptive field and a fixed parameter count. They also give constructive conditions for marginal stability that prevent vanishing and exploding gradients. Experiments report better performance than other recurrent models on long-horizon benchmarks with substantially fewer parameters.
Aleix Salvador-Pomarol, Arthur N. Montanari, Earl K. Miller et al.
In briefTetris3D: Generating 3D Scenes with Objects That Fit Together
SummaryThis paper presents Tetris3D, a framework that reconstructs 3D scenes from a single image so that objects are geometrically and physically consistent with each other. Each object's generation is conditioned on the shapes of neighboring objects and their physical relationships, which keeps its shape and pose plausible in the scene. The authors also introduce ComOb, a physics-simulation dataset of 1.2M scenes with per-object meshes and pairwise physical relation annotations. Experiments on synthetic and real scenes report coherent shapes and poses even under occlusion, with state-of-the-art generation quality and physical stability.
In briefNever Look Back: How Egocentric Video Lets Assistants Remember 3D Object Locations Over Time
SummaryThis paper studies how an embodied assistant can build a human-like memory of objects by observing a person's daily activities in egocentric videos. The authors present Ledger, a persistent 3D object memory that stores object locations, movement histories, and short descriptions, and keeps objects even after they leave view. The memory raises accuracy on HD-EPIC from 29.7% to 42.6% and on UCS-Bench from 33.8% to 38.5%.
Shravan Chaudhari, William Paul, Suchi Saria et al.
In briefExpDis: Separating Exploration from Optimization in RLVR
SummaryThis paper addresses how exploration and optimization interfere in reinforcement learning with verifiable rewards (RLVR) for language models. It proposes Exploration-Distillation (ExpDis), which trains explorer policies with a novelty bonus, filters their correct and high-quality trajectories, and distills them into a student policy trained without that bonus, alternating over several rounds. Across seven math reasoning benchmarks and two model families, ExpDis outperforms DAPO at the same wall-clock budget and improves pass@k, indicating more diverse correct solutions.
In briefWhy adversarially trained networks produce sparse saliency maps: a theoretical explanation
SummaryThis paper explains why gradient saliency maps of adversarially trained neural networks often look sparse, focusing on two-layer ReLU networks. It builds on the known equivalence between adversarial training and empirical risk minimization with weight decay and an adversarial total variation term. The authors prove that, as data and neurons grow and regularization parameters shrink appropriately, minimizers converge to a Bayes classifier with minimal gradient and Barron norm. Sparsity arises because the gradient norm under ℓ∞ attacks is anisotropic and favors axis-aligned gradients, and experiments comparing gradient ℓ1-norms and thresholded sparsity of naturally and adversarially trained models support these findings.
Yannick Lunk, Atell Yehor Krasnopolsky, Damien Garreau et al.
In briefEngramEdit: Updating LLM Facts Independently Through Conditional Memory
SummaryThis paper proposes EngramEdit, a method for updating factual knowledge in large language models that use conditional memory architectures such as DeepSeek Engram, while keeping the Transformer backbone fixed. It first computes target memory representations so the model predicts the updated fact across multiple expressions, then jointly updates shared n-gram embeddings, penalizing frequently reused embeddings more strongly to protect unrelated knowledge. Experiments report near-perfect editing success, with revised knowledge usable on unseen expressions and in multi-hop reasoning, and unrelated knowledge and general capabilities largely preserved.
In briefLong-WAM: Extending Context in World-Action Models for Robot Control
SummaryThis paper presents Long-WAM, a model and system framework that extends the visual history context of causal world-action models for real-time robot control. The authors find that longer histories help most when the video foundation model is pretrained autoregressively. On RoboCasa GR-1, extending context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%. The method reports the best results among compared approaches on LIBERO-Long, RoboTwin 2.0, and DOMINO, and it is deployed on real robots and several hardware platforms.
In briefDecentralized SGD under heavy-tailed noise and the role of gradient clipping
SummaryThis paper studies decentralized stochastic gradient descent (DSGD) under heavy-tailed noise, focusing on how clipping local gradients affects optimization and consensus. For smooth non-convex costs with bounded p-th moment noise, the authors show that clipped DSGD achieves order-optimal convergence rates both with high probability and in expectation. They also establish a linear speed-up in the number of agents. Unlike normalization, clipping keeps gradient magnitude information and remains convergent, and numerical experiments support the theory.
WhyShows clipped decentralized SGD achieves order-optimal rates and linear speed-up under heavy-tailed noise, distinguishing clipping from normalization; experiments only validate theory.
In briefRephrase Before You Act: Reducing Language Sensitivity in Robot Vision-Language-Action Models
SummaryThe paper shows that vision-language-action models for robots are highly sensitive to how instructions are worded, with a single-word change sometimes shifting success by tens of points. The authors distill rephrasing rules from many phrasings of a few training tasks using a large language model, then rewrite each incoming instruction once before execution. Applied to a frozen π0 policy, the rules improve success by 16 to 27 percent relative on twelve held-out tasks, without retraining and zero-shot to unseen tasks.
Mikey Watts (Independent Researcher), Yuchen Cui (University of California, Los Angeles)
In briefGRACE: Compressing video generation models for faster output without losing quality
SummaryThis paper introduces GRACE, a two-stage framework that compresses a pretrained video autoencoder for faster video diffusion while keeping it compatible with the pretrained Diffusion Transformer. It keeps a frozen base latent and learns a residual latent for information lost under stronger compression, aligning the compressed latent with the original in the DiT's feature space. According to the paper, GRACE cuts Wan2.1-I2V-14B's token count by 8x and latency by 11.1x at 480x832x81, while matching the pre-compression pipeline's generation quality on VBench.
In briefGraph Geometry Distillation: Closing the GNN-to-MLP Knowledge Gap
SummaryThis paper studies distilling a message-passing GNN teacher into a graph-free MLP student that is deployed without graph access. The authors argue that existing methods do not specify where the student should preserve the teacher's graph-induced geometry, which they link to two spectral failure modes on sparse and dense graphs. They propose G^2MLP, a training-time framework guided by Ollivier-Ricci curvature, which they report consistently outperforms graph-free distillation baselines on node-classification benchmarks.
In briefWhy forget-only machine unlearning requires memorizing training data
SummaryThis paper studies forget-only unlearning, where a model must delete selected training examples using only the trained model and those examples, without retained data. The authors show that different datasets can yield the same model yet need different outputs after deletion, which gives lower bounds on how closely unlearning can match retraining. They also show that handling arbitrary deletion requests can require memorizing a large amount of training information, up to the entire dataset for simple threshold learners.
WhyProves forget-only unlearning requires memorizing training data, with tight lower bounds for threshold learners; novel and rigorous, with notable subfield impact.
In briefRoboJEPA: Scaling Laws for Robotic Latent World Models
SummaryThis paper presents RoboJEPA, a latent world model built on the Joint Embedding Predictive Architecture and trained on data from 12 robot embodiments. The authors find that its imagination error follows a second-order power law in compute, and that downstream planning performance improves predictably and correlates strongly with this error. They also show the model can be deployed zero-shot on real robots to plan long-horizon tasks toward a single goal image, and release checkpoints and code.
Artem Zholus, Nicolas Beltran-Velez, Jianhao Yuan et al.
In briefSciExam for ENSO: Can AI Agents Build Climate Models?
SummaryThis paper introduces SciExam for ENSO, a benchmark where language-model agents build low-order stochastic models of El Niño-Southern Oscillation from real observations within a six-hour budget. Hidden graders assess each model on reproducing ENSO statistics, reconstructing unobserved variables, and forecasting held-out years, scoring a published model the same way. Six of twelve agent systems outscored the published model, mainly through better reconstruction and forecasting.
WhyProposes a benchmark grading open-ended agent-built ENSO models with hidden tests, with controlled runs supporting results, though scope is limited to one climate phenomenon.
In briefVideo-conditioned generative joint 2D-3D hand motion recovery
SummaryThe paper proposes JoHan, a unified generative framework that recovers 3D hand motion directly from video without relying on per-frame pose predictions. Trained from scratch, it jointly generates aligned 2D and 3D local hand pose sequences, learning their temporal dynamics and cross-representation correspondence. Experiments on challenging benchmarks report improved accuracy and speed in local hand-pose and camera-space reconstruction, with smoother motion while maintaining high per-frame pose accuracy.
In briefPrompts That Do More: How Retrieval Instructions Affect Embedding Models
SummaryThis paper studies how retrieval instructions change the query representations produced by embedding models in asymmetric retrieval tasks. The authors show that models can fail to follow even simple task instructions when query-side distractors are present in evaluation. They attribute this to current training and evaluation setups, and report that fine-tuning with added query-side distractors yields substantial gains with minimal effect on other tasks.
Amanda Myntti, Jenna Kanerva, Veronika Laippala et al.
In briefRECAST: A Method That Learns to Derive the Right Context by Adaptively Routing Evidence
SummaryThis paper presents RECAST, a framework that treats evidence construction as a sequential decision process over heterogeneous retrieval and computation operations, so evidence can be derived rather than only retrieved. A lightweight RouterLM selects or specifies operations, which a frozen CompilerLM turns into executable code, and once evidence is judged sufficient, a frozen AnswerLM produces the final answer. The RouterLM is trained with supervised fine-tuning followed by group relative policy optimization. The authors report a 75.6% mean success rate across six benchmark families and zero-shot generalization on three held-out benchmarks.
In briefParameter-free smoothed online learning with an ERM oracle, without knowing the base measure
SummaryThis paper studies smoothed online learning, where each covariate's conditional law has density at most 1/σ relative to a base measure μ, interpolating between adversarial and stochastic settings. Prior oracle-efficient methods needed sampling access to μ or labels perfectly predictable by a fixed hypothesis. The authors give the first oracle-efficient algorithm achieving sublinear regret in the agnostic setting without knowing μ, based on Gaussian Follow-The-Perturbed-Leader. It is parameter-free, needing no knowledge of μ, σ, or T, and achieves regret Õ(d√(T/σ)) for binary classes of VC dimension d with one ERM call per round, optimal up to a √d factor.
WhyFirst oracle-efficient agnostic smoothed online learner without knowing the base measure, with near-optimal regret and new techniques.
Sasha Voitovych, Adam Block, Alexander Rakhlin et al.
In briefEmbodiedRSI: Robots That Learn Continually by Testing Their Own Hypotheses
SummaryThis paper presents EmbodiedRSI, a self-evolving agent framework that lets a robot decide autonomously which physical experiments to run next and turns the results into better code and skills. It uses a fast-slow dual-system design, a Hypothesis Graph, and value-of-information experiment selection. On RoboCasa365 it reaches 77.0% overall success and on LIBERO-Pro 86.8%, and it also transfers zero-shot to a real robot with 71.3% overall success.
In briefQuadTok: A Quadtree Visual Tokenizer for Image Generation
SummaryThis paper introduces QuadTok, a visual tokenizer that uses a hierarchical quadtree structure to allocate more representation to complex image regions and coarser coding to uniform ones. Compared with a fixed 256-token grid, it saves about 10% of tokens on ImageNet and 9% zero-shot on COCO with comparable reconstruction fidelity. A 947M-parameter autoregressive generator built on it reaches a 2.08 gFID on ImageNet 256×256.
In briefEvolutionary Neural Architecture Search for Lake Chlorophyll-a Prediction from Sentinel-2 Data
SummaryThis paper revisits a Sentinel-2 algal bloom classifier for lakes and uses regularized evolution to search small multilayer-perceptron architectures while keeping the original features and lake-level train/test split fixed. Selecting on inner cross-validation AUC, the searched networks raise held-out AUC from 0.790 to 0.820 and accuracy from 0.733 to 0.748 with 409 trainable parameters, 26 times fewer than the strongest hand-designed reference. At about 1.6 kB, the model is small enough to serve as an onboard screening trigger.
In briefInsights from AutoResearch for Solar Panel Segmentation
SummaryThis paper studies AutoResearch, where a coding language model edits a training program within a one-hour GPU budget and keeps changes only if validation IoU improves. Applied to photovoltaic panel segmentation with DeepLabV3-ResNet-50 held fixed, three campaigns of 24 experiments all beat their one-hour baselines, but retained modifications did not transfer across hardware. The Qwen3-8B configuration trained on real images only reached a test IoU of 0.836, compared with 0.833 for the reference GAN-augmented schedule.
Justinas Lekavicius, Kursat Komurcu, Valentas Gruzauskas et al.
In briefBest-arm identification when all arm means shift together over time under adversarial perturbation.
SummaryThe paper studies best arm identification in stochastic bandits where the gaps between arm means stay fixed but a common shift can be chosen adversarially each round. The authors show that GLRT-based stopping rules, including Track-and-Stop, fail under such time-varying shifts. They propose ISM, an algorithm using importance weights, which is shown to be δ-correct with a sample complexity of order K(σ²+U²)Δ_min⁻² ln(1/δ). A matching worst-case lower bound up to constant factors is also provided, along with empirical evaluation.
Lukas Zierahn, Wouter M. Koolen, Shubhada Agrawal et al.
In briefBEANS-Next and ROOTS: Broader Audio-Language Tests and Training Data for Bioacoustics
SummaryThis paper introduces BEANS-Next, a benchmark covering bioacoustic tasks such as acoustic perception, biological category recognition, scene understanding, and in-context learning. It finds that existing audio-language models perform poorly beyond the task types emphasized by earlier evaluations. The authors also present ROOTS, a large training dataset built from curated real-world data, behavioral and acoustic metadata, and synthetic data, and report substantial gains across all BEANS-Next task groups. The benchmark, dataset, and data pipelines are open-sourced.
WhyBuilds a broader bioacoustics benchmark and large training dataset, with experiments showing substantial gains across task groups.
Christos Plachouras, David Robinson, Marius Miron et al.
In briefAbridgeKV: Compressing long reasoning into learned pause tokens' KV cache
SummaryLong chain-of-thought decoding is limited by memory, since the KV cache of generated tokens grows with test-time compute. AbridgeKV trains a model to compress each completed block of its output into the KV entries of a few learned pause tokens, then frees the block's original cache. On 446 competition problems, keeping a quarter of the generated cache matched a full-KV model in accuracy and output length, and in vLLM the cache was 3.5x smaller and problems finished 3.2x faster on a saturated H100.
WhyIntroduces a new trained pause-token KV compression paradigm, validated on competition benchmarks and real vLLM speedups with strong baselines.
In briefA fix for biased two-level softmax sampling caused by cluster size and dispersion
SummaryThe paper studies two-level softmax (2LS) sampling, which first picks a cluster and then an item within it to avoid costly exact softmax sampling. The authors show that 2LS introduces systematic bias because it ignores cluster size imbalance and intra-cluster similarity dispersion. They propose Size-Corrected 2LS (S-2LS) and Size- and Dispersion-Corrected 2LS (SD-2LS), which reduce this bias with little extra computation. Experiments on five large-scale datasets validate the improved sampling properties.
In briefAgentic RSR: Moving Robot Policies from Real Scenes to Simulation and Back
SummaryThe paper presents Agentic RSR, a framework that reconstructs a real robot workspace from video into a MuJoCo simulation and uses a coding agent to develop an executable manipulation policy. The agent refines the policy using execution feedback and transfers it to a real robot. All 18 reconstructed scenes across two robots pass a joint visual check, and real-robot task success reaches 80% of simulation success.
In briefPredicting Coding-Agent Performance from Base Models Before Post-Training
SummaryThe paper asks how to predict which base model is worth expensive agentic post-training. It replays successful post-trained trajectories to find the decisive step where a patch first turns failing tests into passing ones. Three screens are built at that step: Decisive-Action BPB, a Patch MCQ, and prefix-conditioned pass@K. Across ten base and post-trained model pairs, all three rank models in close agreement with post-trained SWE-bench Verified pass@1.
Tan Yu, Alexander Bukharin, Khushi Bhardwaj et al.
In briefLabel-free cell counting and viability prediction using brightfield imaging and deep learning
SummaryThis paper presents ViabiLens, an AI-assisted software that assesses cell viability without staining, using brightfield images. It combines a cell detection model with a CNN classifier to predict live or dead cells, plus an interactive UMAP-based viewer. On Chinese Hamster Ovary cells, it achieves a mean absolute error of 2.68% on unstained samples compared with fluorescence-based references. The authors also release a benchmark dataset for label-free cell viability analysis.
Amir Reza Vazifeh, Christian Zeigler, Sornanathan Meyyappan et al.
In briefA Society of Researchers: Institutional Design for Autonomous Research Agent Populations
SummaryThe paper proposes a society of persistent AI agents operating under explicit institutions, developed for science as a society of researchers built on six principles. Principal investigators compete for shared compute through proposals, independent review, and grants, while a human governor, the mayor, allocates resources without assigning tasks. In a running society of ten thousand researchers tasked with improving language model pretraining, one lab reported reaching the same quality with about 30% less compute, a result other labs have not yet agreed on.
In briefHow pre-class use of a voice AI discussion partner affects active student participation in class
SummaryThis preregistered field experiment randomly assigned 759 MBA students in ten course sections to prepare for two class sessions with a purpose-built voice-based AI discussion partner. After two uses, students made about 31% more voluntary contributions in each later class session. Students who used the AI partner more reported greater comfort speaking up and greater perceived learning, but not greater focus or motivation.
Dan J. Wang, Neelam Modi Jain, Vanessa Burbano et al.
In briefFoldBack: A Self-Correcting Policy for Long-Horizon Garment Folding
SummaryThis paper presents FoldBack, a self-correcting masked generative policy for folding garments over long action sequences. At inference time, it decides when to refine and verify grasps, how to roll back to a retryable state, and where to retry, so failed grasps can be detected and repaired. It requires no recovery demonstrations or base-policy retraining. On 33 real garments from six categories, it reached 75.2% final folding success and 0.837 final-mask IoU, compared with 45.7% and 0.689 for the strongest prior baseline.
In briefDelta distillation: transferring what each teacher learned for multi-teacher training
SummaryThe paper introduces Δ-MOPD, which transfers each teacher's logit shift relative to its base model, re-anchored at the student's frozen initialization, instead of the teacher's endpoint policy. It compares the two approaches in both common-domain composition and routed-domain settings. With three composed teachers, the authors report Δ-MOPD exceeds endpoint composition by 4.11 Math points and 1.95 points averaged over five benchmarks.
In briefMORCA: Offline-to-Online Reinforcement Learning for Adaptive Cache Reuse in Video Diffusion Acceleration
SummaryDiffusion Transformers generate high-quality videos but are slow because of iterative denoising. This paper proposes MORCA, a cache scheduling framework trained with offline-to-online reinforcement learning. It uses latent information to decide whether to reuse or recompute cached results under a user-specified acceleration target. Experiments on several video generation models and speedup ratios report better generation fidelity than existing caching methods at comparable computational budgets.
WhyApplies offline-to-online RL to diffusion cache scheduling, modeling terminal error and latent cues, with multi-model, multi-speedup comparisons against state-of-the-art caching.
In briefPHRBench: Evaluating How LLMs Reason After Encountering Hallucinations
SummaryThis paper introduces PHRBench, a controlled benchmark that examines how 18 large language models reason after encountering hallucinated premises, across four domains. It classifies reasoning trajectories as Hallucination Compliance, Hallucination Avoidance, or Heuristic Correction, independent of final-answer correctness. Across 4,820 instances, successful recovery is relatively rare and linked to more frequent belief updates, and properties of the hallucinated prompt predict recovery, with a lightweight predictor reaching an AUROC of 0.847.
In briefMemoCare: An Interactive Mobile System for Automated Multimodal Cognitive Screening
SummaryMemoCare is a mobile app that screens cognition using spoken answers, orientation questions, touchscreen tasks, and a drawing task, in English and Vietnamese. Software tests passed 151 of 151 predefined cases, and the drawing module reached 91.33% mean balanced accuracy on a locked 71-image test set. Four clinician co-authors reviewed the workflow, giving a pooled median rating of 4 out of 5 across eight criteria.
Duy-Cat Can, Mau Minh Phuc Le, Tuan-Khoa Hoang et al.
In briefA Self-Teacher That Matches the Student: Joint On-Policy Learning and Teaching (JOLT)
SummaryThe paper argues that self-distillation with privileged information does not guarantee the student improves, since the privileged teacher may use shortcuts unavailable to the student. The authors derive a necessary and sufficient condition for the teacher's distillation update to align with the student's reward gradient. Based on this, they propose JOLT, which trains one policy in two roles: a KL-regularized privileged teacher and an unprivileged student. The authors report improved training efficiency and performance on math reasoning, coding, tool use, and terminal use.
In briefRunningTab: Workspace Agents with an Environment-Side Task Tab
SummaryThis paper presents RunningTab, a framework for LLM agents that work directly with files in a workspace. The environment keeps a per-task record of requirements, read file excerpts with provenance, and listed but unopened candidate files. Agents resolve or set aside requirements and receive open items in a finish check. Tests on three benchmarks and three LLMs show it outperforms plain direct workspace interaction and baselines that keep the record inside the model.
In briefTeaching PPG to Track Change, Not Identity: ECG Fixed-Effects Distillation
SummaryThe paper finds that when ECG teacher models are used to train PPG-only student models, the students mainly memorise each person's identity (a per-recording mean) rather than changes in cardiovascular state. The authors propose fixed-effects distillation, which subtracts each recording's mean from predictions and targets so the identity component cancels out. Reported results include more than doubled state agreement and improved within-person labels, with the gain holding on two backbones and two further databases.
WhyFixed-effects distillation separates person trait from state change, showing ECG-to-PPG distillation mostly learns identity; validated across 34 runs, two backbones, and databases.
In briefSeq-Flow: Efficient Probabilistic Forecasting with Self-Rollout Error Control
SummarySeq-Flow is a conditional flow model that updates a distribution over future trajectories as new observations arrive. It transports samples from the previous forecast distribution to the updated one, enabling accurate updates with few flow evaluations. Self-rollout training controls error accumulation across recursive updates. On particle-accelerator beam spill forecasting, it reduces CRPS by 65% under a few-NFE budget.
In briefECHO: Embodied camera recordings of humans carrying objects to where they belong
SummaryThis paper introduces contextual object placement, a benchmark task that predicts where an object should go during an observed carrying episode. The authors release ECHO, a synthetic dataset with dense RGB-D scans of 115 HM3D scenes across 159 floors, 3,805 annotated episodes, and 198 objects, including trajectories and natural-language context. Their evaluation finds that no single input modality is sufficient, so models must jointly reason over scene structure, human activity, and contextual knowledge.
In briefQ-Learning with Scalar Adjoint Matching: a cheaper way to fine-tune flow policies
SummaryThe paper addresses the high computational cost of fine-tuning flow policies against a learned value function. The authors observe that the batch-averaged velocity Jacobian of pretrained flow policies concentrates on its diagonal, and derive a closed-form scalar adjoint that scales the final-action value gradient by flow time, avoiding per-step vector-Jacobian products. They propose SQAM, which combines this scalar adjoint with a value penalty on policy-generated actions, and report success-rate gains of 18 to 35 points over the strongest baseline on the four hardest OGBench domains. SQAM also improves over supervised fine-tuning when applied to a vision-language-action policy on a real bimanual robot.
In briefA knowledge-guided LLM that turns unclear network service requests into validated orchestration profiles for 6G networks
SummaryThis paper addresses natural-language service intents in integrated terrestrial and non-terrestrial (TN/NTN) 6G networks, which can be incomplete, ambiguous, or inconsistent. It proposes a framework that combines grammar-constrained semantic parsing with a knowledge-graph-driven clarification loop to produce validated service profiles for four service classes. On a 200-sample dataset, it reports 89% classification accuracy, 95% policy validity, and an average of 0.66 clarification turns per request. Compared with a pure-LLM baseline, it improves policy validity and contradiction detection while requiring fewer clarification turns.
In briefDetecting adversarial images via response profiles of vision-language models
SummaryThis paper studies how to identify adversarial images fed to frozen vision-language models, even when their confidence and image-text similarity look normal. The method summarizes how an image responds to a set of general semantic prompts, using category statistics, prompt relationships, deviations from clean reference distributions, and stability under weak transformations. A lightweight classifier uses this compact profile while the VLM stays fixed. Experiments across several datasets, CLIP-style backbones, and multiple attack types show the response patterns offer a useful complementary signal for detection.
In briefGenerative AI model rapidly producing 3D wind and temperature fields for urban microclimate
SummaryThis paper proposes Conditional Flow Matching (CFM), a generative framework that uses building geometry and mean flow as conditions to produce three-dimensional instantaneous wind and temperature fields for urban microclimate in seconds. A shared-noise initialization over overlapping pixel-space patches enables parallel generation while preserving spatial continuity. Compared with reference large-eddy simulation data, the model achieves wind and temperature first-order statistics NRMSE of 2.99% and 1.77%, respectively.
In briefSGF+: Separating Gradient Flows for Autoregressive Video Generation
SummaryThis paper finds that autoregressive video generation shares parameters between denoising current frames and writing context for future predictions, causing conflicting gradients that hinder joint optimization of visual quality and temporal consistency. The authors propose SGF+, which assigns separate parameters to these two roles while preserving their interaction through causal attention, optimized with the original generation objective. According to the paper, it improves visual quality and long-horizon consistency in framewise and chunkwise generation, and supports continuous generation up to 24 hours after training on only 5-second rollouts.
In briefDerivative Gaussian processes with two directions per gradient: a cheaper surrogate model for using gradient observations
SummaryThe paper proposes a derivative Gaussian process that uses only two directions per observed gradient, one capturing direct contributions to prediction and one aggregating indirect contributions through correlations with function values. Within a Vecchia approximation, each prediction costs O(m³) in dense factorization. In simulations, it matches the accuracy of a leading exact gradient-reduction method at equal conditioning set size while handling much larger conditioning sets with less time and memory.
