In briefDistribution matching distillation to speed up continuous diffusion language models
SummaryThis paper studies how distributional distillation can cut the number of network evaluations needed by continuous diffusion language models. It proposes Simplex-DMD, which uses continuous token relaxations and pathwise gradients, and Reinforce-DMD, which uses categorical sampling with REINFORCE and a learned density ratio, both with a reverse-KL matching objective. On OpenWebText with 1,024-token sequences, Simplex-DMD reaches a generative perplexity of 45.6 in 4 evaluations, a 49% reduction versus the strongest baseline at matched entropy. Reinforce-DMD reaches 14.9 perplexity with 256 evaluations, a 20% reduction.
Paul Le Van Kiem, Dario Shariatian, Umut Simsekli et al.
In briefEviRover: Agentic Perception Trained Beyond a Single Glance
SummaryStandard visual perception makes a one-shot prediction from a single view of an image, which often fails when fine details or up-to-date knowledge are needed. The paper frames perception as an agentic process that gathers more information through interaction, and builds training data plus EviLens, a human-verified benchmark of 688 instances. The EviRover model is trained with supervised fine-tuning followed by agentic reinforcement learning, and the 4B version outperforms its backbone by 30 points on average on EviLens.
WhyReframes perception as interactive agentic process with new data and benchmark, showing consistent gains across multiple evaluations.
In briefLOCI: Spatial Linear Memory for Streaming Video World Models
SummaryThis paper introduces LOCI, a hybrid memory architecture for video world models that keeps both a key-value cache of past observations and a recurrent linear-attention memory conditioned on camera geometry. This lets the model reproduce previously seen content more faithfully when the camera revisits a region. On the public MIND benchmark and held-out recorded trajectories, LOCI outperforms representative world models and a same-recipe full-softmax model. With full history, it lowers peak memory by about 30% at equal length, and with a bounded bank of observations it streams long videos at constant memory.
In briefPhantomEnvironments: Training LLM Agents in Fictional Worlds – rule-built fictional environments for training search agents
SummaryThe paper introduces PhantomEnvironments, multi-turn reinforcement learning environments built from fictional worlds generated entirely by rules, requiring no LLM and at zero marginal cost. Agents search a corpus of templated articles to answer multi-hop questions. The trained agents transfer to real-world multi-hop search benchmarks, often outperforming agents trained on real-world data on newer benchmarks. Ablations indicate that hop count drives transfer more than other environment features.
WhyRule-generated fictional environments train search agents that transfer to real multi-hop benchmarks, with ablations supporting the claims.
Anmol Kabra, Swathi Saravana Selvam, Albert Gong et al.
In briefLearning Skills from Past Action Trajectories: An Action Experience Dictionary for World Action Models
SummaryThis paper proposes an Action Experience Dictionary (AED) that encodes historical physical action trajectories into shared action embeddings to enable skill reuse across manipulation tasks. Historical actions are aligned with visual observations, retrieved from the dictionary, visually conditioned via cross-attention, and prepended to noisy action tokens. A motion-aware transition loss is also introduced to reduce reliance on irrelevant background cues. The authors report experiments on simulation benchmarks and real-world cross-embodiment settings.
In briefJevSpawn: Faster, Adaptive AI Agent Reasoning via Compositional Action Spaces
SummaryLLM agents produce reasoning and actions token by token, which makes long interactions slow and costly. JevSpawn is a compositional policy that links natural language task descriptions to finite probabilistic exploration, combining parallel action spawning, feedback-driven branch selection, representation revision, and recovery from retained alternatives. The authors evaluate it on eight benchmark tasks against seven agent baselines and a TypeSafe Jev variant. They report improved task performance and faster navigation.
In briefEach Batch as Its Own Validation Set: Leave-One-Out Gradient Matching for LLM Data Selection
SummaryThis paper proposes LOOM, a method for selecting useful examples batch by batch when fine-tuning large language models. The authors argue that matching gradients within the same batch biases the objective toward noisy examples, so full-batch training often wins. Removing each example's own contribution gives an unbiased error estimate, and the method outperformed full-batch training and other gradient-matching approaches across several tasks and models.
WhyLeave-one-out gradient matching removes covariance penalty with an exact half-batch superiority condition; broad backbone and task comparisons, impact confined to LLM fine-tuning data selection.
In briefIdentifying Urban Areas in Drone Visible-Light Imagery
SummaryThis study compares methods for distinguishing urbanized from non-urbanized areas using RGB images captured by unmanned aerial vehicles (UAVs). The methods tested are vegetation index thresholding and neural networks, applied to two study areas surveyed with different UAVs and cameras. Neural networks reached about 96% classification accuracy, higher than the best vegetation index, Excess Blue, at about 87%. Because the datasets are highly imbalanced, the Matthews correlation coefficient was also used for evaluation.
Edyta Puniach, Wojciech Gruszczy\'nski, Pawe{\l} \'Cwi\k{a}ka{\l}a et al.
In briefHow AI Agents Discover Scientific Equations: Rediscovering the Hydrotope and Finding New Water-Wave Amplitudes
SummaryThe paper studies how AI agents discover and check scientific formulas, using the recently found hydrotope, a geometric formula that merges polynomial pieces of nonlinear surface-wave scattering into one expression. Across 18 single-prompt runs with no hint, a false hint, or a true hint, only four recovered the formula in all regions, while others found correct regional pieces but failed to combine them. A workflow with one lead agent coordinating two research agents fully rediscovered the hydrotope and produced a new verified analytic expression for the six-point amplitude A6 in a harder negative-wavenumber problem.
In briefSpeechConversationBench: A Benchmark for Multi-Turn Spoken Math Reasoning
SummaryThis paper introduces SpeechConversationBench (SCB), which evaluates spoken mathematical reasoning using 103 sharded GSM8K problems. It compares problems delivered in one turn, as concatenated information shards, and as incremental spoken disclosure across turns. Four commercial speech systems and the authors' LEGO pipeline are tested, and sharded accuracy falls 5.0 to 25.3 points below concatenated accuracy for the commercial systems. LEGO scores 77.5 percent in all three conditions.
In briefMemLife: a memory system for curating and querying long-term first-person video
SummaryMemLife is a multimodal memory system that turns long-term egocentric video into entity-grounded, first-person text episodes and retrieves them with a time-indexed agentic reader. It needs no training and no query-time video access, and improves over the strongest training-free baseline by 4.6% to 12.0% across four long-horizon benchmarks. The authors also propose MemOpt, a reinforcement learning method that optimizes the memory writer for faithful, informative, and retrievable memories, adding 2.7% to 5.0% gains.
In briefNear-Linear Accuracy Bounds for Moreau–Yosida Unadjusted Langevin Sampling
SummaryThis paper proves near-linear accuracy bounds for the classical Moreau–Yosida unadjusted Langevin algorithm (MYULA), a sampling method for distributions of the form exp(−f−g) with f strongly convex and g convex and Lipschitz. Under an explicit step-size condition, the invariant-measure bias relative to the Moreau-smoothed target is O~(h), and about O~(ε^-1) iterations give a W2 error of at most ε for the N-th iterate. The analysis does not assume third derivatives or a Lipschitz Hessian, and it uses a Poisson-based estimate to turn a second-order stationary residual into a Wasserstein bound.
In briefCertifying Test-Time Scaling Curves: Making Accuracy Gains from Sampling More Answers Trustworthy
SummaryThe paper studies test-time scaling, where several answers are sampled per question and a verifier keeps the highest-scoring one, with accuracy plotted against the number of samples. It argues such curves are cheap to draw but hard to trust, and proposes confidence bands that cover all budgets simultaneously, deriving the minimax cost of certifying the whole curve. On 185 held-out score pools, the proposed paired audit used 0.74 times the answers of the cheapest competing certified audit at 64 budgets and 0.53 times at 1,024 budgets.
WhyDerives minimax cost for certifying whole test-time scaling curves and a paired audit using fewer answers than competing certified methods on benchmarks.
In briefA provably tractable engine for generating language under NFA constraints
SummaryThe paper studies sampling from language models under hard constraints defined by nondeterministic finite automata (NFAs). The authors propose NFA-LM, a polynomial-time generation engine with theoretical guarantees under mild assumptions. Experiments show it efficiently produces high-quality outputs with theoretically bounded approximation error.
In briefIndex-Translate: A Multilingual Translation Model Family for Text, Speech, Controlled Dubbing, and Long Documents
SummaryThis paper introduces Index-Translate, a family of multilingual translation models in 2B, 9B, and 35B-A3B sizes that supports 150 languages and multilingual instruction following. The authors report that it outperforms similarly sized models and performs comparably to 100B-scale translation models and frontier models on general and complex-instruction evaluations. The family also includes Index-Echo for end-to-end speech translation, Index-Homura for syllable-controlled dubbing, and Index-NativeLong for long-document translation with a dedicated benchmark.
In briefXOR-Trellis: Ultra-low-cost dequantization and curvature-aware LLM quantization without Hadamard transforms
SummaryThis paper studies trellis-coded quantization for compressing large language model weights at ultra-low bit widths. The authors introduce a structured, hardware-efficient trellis dequantizer that keeps diverse reconstruction choices for trellis search. They also reformulate trellis path optimization with a curvature-aware objective that reflects model sensitivity in the original coordinate space. Together, the techniques aim to deliver high-quality ultra-low-bit quantization with cheap, highly parallel runtime reconstruction and no Hadamard-based incoherence processing.
In briefMANET-GNN: A learned distributed optimizer for power allocation in multi-channel mobile ad hoc networks
SummaryThis paper proposes a learned, decentralized framework for transmit-power allocation in dynamic, multi-hop, multi-channel mobile ad hoc networks (MANETs). The authors formulate an end-to-end throughput maximization problem covering unicast, multicast, multicommodity, convergecast, and many-to-many traffic, and use it as an unsupervised training objective for MANET-GNN, a message-passing graph neural network. The model relies only on local, possibly noisy channel state information and a fixed number of neighbor message exchanges, and generalizes across topologies and network sizes. Numerical results report performance competitive with centralized approaches and robustness to channel uncertainty.
In briefLearning from Research: Toward Lifelong Agent Harness Evolution
SummaryThis paper introduces ScholarEvolve, a framework that improves the agent harness (the software handling tool use, memory, and task execution) while keeping the language model fixed. It draws on state-of-the-art research, organizing evolution directions into functional modules and using topic modeling to identify improvement strategies for each. Experiments on AppWorld and Tau2-Bench show gains, including Qwen3.5-27B goal completion rising from 49.6% to 63.6% on AppWorld Challenge.
In briefJunction geometry can be recovered despite unidentifiable blur
SummaryThis paper addresses recovering the arms of junctions, where thin filaments in microscopy or dispersed spectra on X-ray detectors cross or branch, under an unknown Gaussian-mixture blur. The authors prove a bound on the arm count, directions, weights and supporting lines in terms of the Euclidean difference in cell probabilities, with minimax error of order n^{-1/2}. On 360 crossings and branches built from measured single-molecule localizations, their fitting method more often recovered the arm count and all directions within 10 degrees than a tuned fixed-scale Hessian ridge detector.
WhyProves junction recovery survives unidentifiable blur with uniform bounds and minimax rates, validated on measured microscopy and Chandra data.
In briefPrefPI: Steering Robot Policies Toward Unseen Behaviors Using Preference Feedback
SummaryPrefPI (Preference-Guided Policy Iteration) is a framework that adjusts pretrained generative robot policies using only relative preferences between self-generated trajectories. It targets desired behaviors the initial policy rarely or never produces, by casting preference learning as preference-conditioned generative modeling with classifier-free guidance. Tested on diffusion policies and the PI0.5 VLA in simulation and the real world, it raises object transport height on real hardware from 10.7 cm to 19.8 cm using 150 preference-labeled trajectories.
In briefReinforcement learning guides graph changes to speed up sparse triangular solves
SummaryThis paper addresses sparse triangular solve (SpTRSV), a basic kernel in scientific computing whose data dependencies limit parallelism. The authors formulate graph transformation of the dependency graph as a sequential decision problem and train a reinforcement learning agent to learn matrix-dependent transformation policies. On real-world sparse matrices, the method reduces levels by up to 94% and the coefficient of variation of level costs by up to 80%, with policies also evaluated for transfer to unseen matrices.
In briefRole-Adaptive Policy Optimization (RAPO) for offline reinforcement learning
SummaryThis paper proposes Role-Adaptive Policy Optimization (RAPO), which adjusts policy-update coefficients separately for value learning and action execution in offline reinforcement learning. For TD3+BC, RAPO splits the bootstrap and execution actors and adapts their coefficients independently, while for IQL it adapts only the inverse temperature. Experiments on D4RL locomotion and AntMaze tasks show gains over both base algorithms, with larger gains for TD3+BC, whose RAPO version outperforms baselines on average.
In briefFrom Spectra to Joint Schedules: 3+3(+2) Scaling-Law Regimes in LLM Pre-training
SummaryThis paper studies how power-law learning curves change with learning-rate and batch-size schedules in noisy online SGD with linear random features. An exact Volterra equation splits the loss into a forcing term that propagates unresolved target error and a memory kernel that propagates stochastic-error injections, with conditions under which each follows a power law. Controlled nanoGPT experiments indicate that schedules with matched B/η paths are nearly equivalent in intrinsic time, and a forcing-memory surrogate predicts loss across schedules.
WhyExact Volterra analysis with spectral conditions characterizes when schedules preserve power laws; novel theory, validated with nanoGPT experiments and a surrogate model.
In briefWhy Howard's policy iteration can take exponentially long on deterministic decision problems
SummaryThe paper shows that Howard's policy iteration needs exponentially many iterations in the number of states for deterministic discounted and average-reward Markov decision processes, so it is not strongly polynomial. It proves that a Dantzig-style variant runs in strongly polynomial time on deterministic average-reward MDPs. Together with known results for Dantzig's rule on discounted MDPs, this establishes an exponential gap between the two methods.
WhyProves exponential lower bounds for Howard's policy iteration and a strongly polynomial Dantzig variant, establishing an exponential separation on deterministic MDPs.
In briefFrom DNA Design to DNA Slimming: An Auditable Agent-Discovered Deletion-Only Designer
SummaryThis paper introduces "sequence slimming," the task of selecting an exact-length, order-preserving subsequence of a functional DNA element that retains predicted activity. The authors build a quantitative benchmark modeled on NucleoBench that balances sequence reduction against functional retention, and require each slimmer to return source indices so its compliance can be verified. The coding agent ERA modified the substitution-only designer GrAdaBeam into GRADASLIM. On five transcription-factor binding targets at 400 and 100 bp, ERA achieved the highest mean in 9 of 10 settings compared with random and greedy baselines.
In briefLess is more: a data-efficient method for kilometer-scale downscaling of extreme heat
SummaryThe paper presents CASPER, a U-Net with a structure-preserving loss that downscales 32 km reanalysis data to 1 km temperature, humidity and wind. Across 24 configurations of one to eight months of training, held-out error grows linearly with climatological distance to the training data (RMSE = 0.83 + 2.95 d), explaining 90% of the variance. In Vancouver, 11 days of local simulation reduced held-out error from 3.8 K to 1.3 K, and training periods should span the target climate.
WhyShows held-out error scales linearly with climatological distance, validated across 24 configurations and a new region for data-efficient extreme-heat downscaling.
In briefGame-Guided Skill Discovery via Self-Play for Human-Playable Agent Control
SummaryThis paper presents GGSD, a framework that uses self-play in competitive games to discover motor skills that humans can directly play. A hierarchical agent selects from a small discrete skill set through a high-level policy, while a skill-conditioned low-level policy learns each behavior. After training, a human can replace the high-level policy and control the agent with the same discrete skills. Experiments in Ant, Franka-arm, and Unitree G1 environments show humans can compose these skills to solve unseen tasks such as Maze and CubePush.
In briefTouch-driven curiosity helps robots learn manipulation through contact exploration
SummaryThe paper introduces TacEx, a framework that adds tactile feedback to uncertainty-driven exploration, steering robot curiosity toward the sense of touch. Without task rewards or expert demonstrations, the robot learns to manipulate and grasp objects. The interaction data it collects supports offline learning of pick-and-place policies and post-training of vision-language-action models, which improves downstream performance.
Klemens Iten, Alexander Proshkin, Bhavya Sukhija et al.
In briefPrototype-rule regularization for tensor neural networks with scarce labels
SummaryThis paper examines whether a differentiable prototype rule can add a useful inductive bias to rank-constrained tensor neural networks when supervision is limited. The method adds a prototype-based regularizer to the Rank-R objective and optionally fuses prototype evidence with network logits at inference. Experiments on four hyperspectral benchmarks compare seven-fold stratified and spatially separated evaluation. Under spatial evaluation, full neurosymbolic inference raises Macro-F1 by 8.82 points on Botswana, 5.49 on Indian Pines, and 1.59 on Pavia University, and lowers it by 0.62 on Salinas. Most of the gain comes from training-time regularization, while inference-time fusion has a small, dataset-dependent effect.
Eftychios Protopapadakis, Konstantinos Makantasis, and Konstantinos M. Giannoutakis
In briefControlling Biology with Language: Offline Learning of Prompt-Conditioned Interventions
SummaryThis paper examines whether living systems can be controlled through natural-language instructions. The authors treat an archive of past interventions and observed outcomes as a fixed offline dataset, and use a vision-language model to judge whether each outcome matches a text description, training a mapping from instructions to interventions without new experiments or human validation. Applied to xenobots, synthetic multicellular constructs without a nervous system, the mapping generalizes to new instructions and reaches 80.0% accuracy on held-out archive data, versus a 66.7% chance baseline, matching a network trained on ground-truth labels.
Nam H. Le, Douglas Blackiston, Michael Levin et al.
In briefVR-JEPA: A method that guides video generation for visual reasoning using contrastive latent learning.
SummaryThe paper proposes VR-JEPA, a framework that aligns the V-JEPA predictor with task-specific reasoning logic through localized contrastive-state learning and skill-specific experts. Its predicted latent trajectories are used to guide video generation for visual reasoning. Experiments on the VBVR-Pro-Bench dataset show an 11.33% relative improvement over a generation-based reasoning baseline, with reduced physical artifacts and improved logical consistency.
In briefLearning Functional Subspaces for Compressing Neural Networks: Learns which subspaces to discard end-to-end to compress transformer models
SummaryThe paper introduces Learnable Subspace Projections (LSP), a method for compressing transformer models by low-rank factorization. Instead of choosing subspaces to remove with local criteria, LSP learns them end-to-end against a global objective while the pretrained weights stay frozen, then merges them into standard low-rank factors. Experiments on OPT, Qwen3, Llama-2 and ViT-B/16 show LSP outperforms baselines, with the gap widening at higher compression. At -70% compression on Llama-2-7B, it reports lower WikiText-2 perplexity and higher mean zero-shot accuracy than the strongest baseline.
WhyLearns orthogonal projectors end-to-end instead of local closed-form criteria, plus shared KV latents; thorough experiments across several LLMs and ViT show gains growing with compression.
Massimo Bini, Anders Christensen, Stephan Alaniz et al.
In briefDebias It Yourself: Teaching LLMs to Reduce Cognitive Biases
SummaryThis paper proposes Debias It Yourself (DIY), a framework that adapts five cognitive bias interventions from social psychology into debiasing procedures for large language models. These procedures are delivered through three paradigms: in-context examples, instruction tuning, and guided self-revision. Across three models, five bias benchmarks, and eleven baselines, the Train+Revise and Revise approaches rank highest and offer the best bias-reasoning tradeoff. They also reduce bias on unseen dimensions by up to 14.8%.
WhyTransfers human cognitive debiasing interventions to LLMs across three paradigms, validated on multiple models, benchmarks, and baselines.
In briefWhether AMR augmentation actually helps large language models
SummaryAbstract Meaning Representation (AMR) improved many NLP tasks in the past, but its value for modern large language models was unclear. The authors tried to reproduce recent work reporting large downstream gains and found those gains likely came from particular experimental choices; under a consistent hyperparameter protocol, text-only baselines matched or outperformed AMR-augmented models. They also introduced a perplexity-based probe and found AMR does not help LLMs better understand relational content in sentences, so it offers no clear downstream benefit.
Hoa Quynh Nhung Nguyen, Jacopo Staiano, Michael Sullivan
In briefFaster, scalable Cox regression via grouped risk sets and improved LogSumExp convergence
SummaryThis paper studies stochastic minimization of LogSumExp objectives to make large-scale Cox regression computationally feasible. It uses a softplus surrogate with one auxiliary scalar per normalizer, yielding unbiased single-sample gradients and improving the averaged objective bound from T^{-1/4} to O(T^{-1/2}). For Cox regression, neighboring failures are grouped to share auxiliary variables, achieving a mean-square rate of T^{-4/5} up to logarithmic factors. The compressed estimator matches the full estimator's asymptotic distribution.
In briefChainLoRA: a method that merges task vectors while preserving geometry, so language models can keep learning new tasks without forgetting old ones
SummaryThe paper presents ChainLoRA, a framework for continually fine-tuning large language models without replaying earlier data. It combines chain-updated training with post-stream adaptive SVD merging to retain prior knowledge, adapt to new tasks, and stay within a parameter budget. On the Large and SuperNI benchmarks it achieves state-of-the-art results among the evaluated replay-free methods and comes close to the replay-based method's average scores.
In briefReCAP: Persistent Context Graphs for Compacting LLM Agent Memory
SummaryThe paper introduces ReCAP, a memory compaction method that stores attention-derived importance scores and dependency links in a lightweight, persistent context graph. For each new request, it selects relevant messages and supporting context by following these links, without extra model calls. The authors report about 95% lower estimated compaction and cold-restoration latency than Codex's summarization-based compaction on Qwen3-Coder and gpt-oss, and roughly half the historical context per call on SWE-Together.
In briefBeyond Model Ranking: Diagnosing Why Industrial Forecasts Are Biased Across Demand Regimes
SummaryThis paper examines why time-series forecasting models that perform well on benchmarks show systematic bias when deployed in industry. It argues that standard loss functions embed fixed statistical assumptions that are violated in pathological demand regimes such as zero-inflated, skewed, or highly variable data. The authors propose the Regime-wise Relative Bias Vector, a metric-agnostic diagnostic, and report experiments over 13 loss objectives and more than 60,000 series from RetailShiftBench and M5 that distinguish optimization failures from bias failures.
In briefMemetic Trojans: Hidden Attacks Spread Through Agents' Social Sharing
SummaryThis paper introduces memetic trojans, a network attack that hides adversarial payloads in content that AI agents naturally choose to share, rather than forcing propagation through malicious instructions. Using posts extracted from Moltbook, a social platform for LLM agents, the authors found the most effective contagion was retransmitted in about 50% of later posts and upvoted at 2.5 times the average rate. Monte Carlo simulations show memetic trojans increase expected exposure by up to 3.19 times. The authors conclude that prompt-injection detection alone cannot stop this spread and that network-level defenses are needed.
WhyIntroduces endogenous social-contagion attacks distinct from agent worms, quantified via Monte Carlo simulation; experiments rest on a single platform, limiting scope.
In briefUnlearnable or Unmeasured? Reliability of Difficulty Labels in RLVR
SummaryThis paper revisits the observation that some hard prompts resist learning in reinforcement learning with verifiable rewards (RLVR). It finds these prompts do improve, at about one third the rate of learnable ones, but the difficulty labels used to define them are much less reproducible than expected. Part of the gradient-similarity evidence for unlearnability stems from difficult prompts yielding fewer correct rollouts for gradient estimation. The authors propose a sampling-based framework to quantify this instability and determine the evaluation needed for reliable difficulty assignments.
Chandak Chakma, Syed Nazmus Sakib, Nafiul Haque et al.
In briefAgent Error Dataset: 50,000 error-diagnosis pairs for studying and training from LLM agent failures
SummaryThis paper introduces the Agent Error Dataset (AED), with 50,228 error-diagnosis pairs drawn from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models. A five-stage pipeline (AET) collects failed agent runs, generates diagnoses and proposed corrections, and checks them against recorded evidence. In 3,062 matched replay pairs, first-proposal corrections raised verifier pass rates from 18.4% to 51.1%.
In briefOverdoseMoE: A Multi-Expert Model for Predicting Opioid Overdose Risk
SummaryThis paper predicts a patient's risk of opioid overdose within 180 days using one year of prior ICD diagnosis histories. The authors adapt language models through continued pretraining and fine-tuning, then combine models of different scales in a multi-expert framework called OVERDOSEMOE. It reports an AUPRC of 25.17 and AUROC of 69.49, and shows some robustness on an independent MIMIC-IV cohort.
In briefJuryFlow: Human-guided multi-agent AI evaluation that uses judge disagreement
SummaryThe paper proposes JuryFlow, a framework for evaluating AI-generated content with multiple LLM judges. Instead of discarding disagreement through majority voting, it treats claim-level disagreements as signals of uncertainty, and a human makes one minimal intervention to resolve a selected disagreement. Corrections propagate through the disagreement graph and are saved as reusable rubric entries. On MT-Bench and LLMBar, it reports better agreement with gold labels than single-judge and majority-vote baselines.
In briefAdjusting Passive Stiffness of Cable-Suspended Drone Loads by Moving Their Anchors
SummaryThis paper treats cable-suspended aerial vehicles as movable compliant anchors and develops a gravity-aware quasi-static theory to predict and shape the passive Cartesian stiffness of the suspended load. The theory applies to any number of vehicles connected to a point load by taut, straight, inextensible cables. A dynamic rigid-body validation framework is also defined to assess how far the derived stiffness remains predictive beyond the analytical model's assumptions.
In briefAutoDataBench: A Data-Centric Testbed That Evaluates How Well LLMs Improve Training Data
SummaryExisting auto-research benchmarks mix together training frameworks, hyperparameters, compute, and data, making it hard to tell which research capability explains an agent's performance. This paper introduces AutoDataBench, a controlled testbed that isolates data intelligence through three optimization tasks covering data diagnosis, organization, and construction. It also checks whether LLMs can predict the effects of data changes before training, and reports that reusing AutoDataBench trajectories for mid-training improves downstream coding performance.
In briefThunderEP: Faster expert-parallel communication for PCIe-connected consumer GPUs
SummaryThis paper presents ThunderEP, a communication design for expert parallelism on PCIe-connected consumer GPU systems. It removes the relay hops of the traditional ring algorithm, moves data through DMA engines to avoid competing with computation, and reduces polling overhead for completion flags in CPU memory. Integrated into vLLM and tested on RTX 4090 and RTX 5090 systems with three MoE models, it reports average speedups of 2.00x for dispatch and 1.53x for combine over NCCL, and up to 1.66x end-to-end speedup.
In briefGateSPINE: Gated Cross-View Fusion for Lumbar Spine MRI Report Generation
SummaryThis paper presents GateSPINE, a vision-language framework that fuses sagittal T1 and T2 volumes with axial images to automatically generate lumbar spine MRI reports. Its core is a gated cross-view fusion module that decides, per feature channel and location, how much of each view to admit, so the more informative view dominates. Evaluated on three lumbar MRI datasets, it achieves the highest clinical efficacy F1 and stays competitive on standard NLG metrics.
Hoang Nguyen Van, Cuong Vuong Tuan, Trang Mai Xuan et al.
In briefPTNO: A neural operator trained on noisy Monte Carlo data for particle transport
SummaryThis paper proposes the Particle Transport Neural Operator (PTNO), a surrogate model that learns particle transport directly from noisy, low-cost Monte Carlo labels instead of expensive converged solutions. It keeps labels in physical space, enforces positivity with a softplus output layer, and uses a stop-gradient pointwise relative L2 loss. On neutron transport in fusion reactors and radiative transfer tasks, PTNO is reported to be much faster than converged Monte Carlo and cheaper at matched accuracy.
In briefReplay on Demand: an emergent curriculum balancing adaptation and forgetting in continued pretraining
SummaryContinued pretraining helps language models adapt to new domains but can erase previously learned abilities. The paper proposes Replay on Demand (RoD), which sets replay allocation from the model's learning dynamics, ranking adaptation samples by remaining learning potential and replay samples by observed forgetting, with both competing for a shared training budget. Across models, scales, and adaptation domains, RoD matches or improves on the adaptation-forgetting frontier of tuned fixed-replay baselines and model merging.
Lukas Thede, Shengzhuang Chen, Stefan Winzeck et al.
In briefMeanVoiceFlow2: Faster one-step zero-shot voice conversion by jointly optimizing the flow module and content encoder
SummaryThis paper proposes MeanVoiceFlow2, a voice conversion framework that jointly optimizes a flow-based conversion module and a more efficient content encoder. It is trained with conversion distillation from MeanVoiceFlow and reconstruction of real data, plus diffusion-GAN training to improve realism and disentanglement. In zero-shot experiments, it reportedly achieves higher perceptual quality and about 9 times faster inference than MeanVoiceFlow, with comparable speaker similarity.
Takuhiro Kaneko, Hirokazu Kameoka, Kou Tanaka et al.
In briefBatSLAM 2.0: Sonar Mapping with Verified Loop Closures
SummaryThis paper presents BatSLAM 2.0, a sonar-only mapping system inspired by bat echolocation for robots navigating dark or cluttered spaces. It combines an updated acoustic front-end, a sequence verifier that checks loop closure candidates, and a pose graph built on a high-performance factor graph framework. The system was evaluated on both simulated and real-world recordings. The authors report that it builds robust topological maps, avoids map collapse, and scales robustly with map size.
In briefHow tissue detection changes false positives in diffusion-based histopathology artifact detection
SummaryOne-class artifact detectors learn normal tissue from a clean training pool built by a preprocessing step that includes tissue detection. On 16 annotated TCGA slides, the authors found that saturation-Otsu detection excluded normal tissue with large clear spaces, such as adipose and alveolar tissue. Switching to entropy-based detection reduced held-out clean-slide false positives from 0.102 to 0.016 without losing sensitivity, indicating the gain came from pool composition rather than size.
In briefLongEmo: Understanding and Reasoning About Emotions in Long Videos
SummaryThis paper introduces LongEmoBench, a benchmark for testing emotion understanding and reasoning in long videos, from continuous scene interactions up to complex episodic developments. It also proposes LongEmo, a memory-augmented agent that builds an Event Memory Graph to model long-range dependencies and emotional dynamics across events. Evaluations of 17 representative methods show that they struggle with long-video emotion reasoning, while LongEmo achieves state-of-the-art performance.
In briefRobust and learned online matching in growing trees
SummaryThe paper studies irrevocable maximum-cardinality matching in trees that grow by successive leaf attachments, where the growth law is misspecified or unknown. For deterministic affine attachment forecasts, the optimal threshold policy loses at most twice the cumulative expected conditional total-variation error relative to an oracle knowing the true growth law. For uniform-preferential attachment with unknown mixture parameter, the authors estimate it from the growing tree and update the policy at geometric times, achieving expected regret O(√n log²n). The exact minimax rate is left open.
In briefAn accelerated algorithm for sparse regularized partial optimal transport
SummaryThis paper proposes a new optimization framework for partial optimal transport, which relaxes the strict mass conservation constraint of classical optimal transport. It uses smooth, strongly convex regularizers such as quadratic or elastic net, reformulated through a penalty approach to allow efficient gradient-based updates. The authors design an accelerated first-order algorithm that alternates smooth updates with simple projection steps. Experiments on color transfer, domain adaptation, and point cloud registration report lower transport cost, higher sparsity, and faster convergence than established baselines.
In briefUsing digital twin structure to make time-series AI models better at predicting machine failure
SummaryThis paper benchmarks five time-series foundation models on remaining useful life prediction using the C-MAPSS dataset, finding that multivariate models clearly outperform univariate ones. The authors propose a fusion method that uses topological constraints from the digital twin's asset structure to shape cross-attention, so representations follow local physical connectivity. An ablation study shows that pretrained weights, target-task adaptation, and digital twin representations are complementary, and topology-constrained attention slightly outperforms unconstrained fusion. A frozen model informed by digital twin representations remains competitive with, and in some cases exceeds, state-of-the-art results on this task.
In briefInference Auctions: Letting users bid for priority in LLM inference
SummaryWhen demand for model inference exceeds available compute, providers must decide which requests to serve first. This paper proposes an auction that lets users bid for faster service, along with fast algorithms for pricing that encourages truthful bidding. It also introduces an autobidding agent that adjusts bids over time to maximize user utility within a user-set budget. Experiments show the auction increases system welfare while keeping the cache utilization and latency benefits of SGLang.
Keegan Harris, Siddharth Prasad, Asher Trockman et al.
In briefCommunity-driven API and AI writer design for scaling Community Notes openly
SummaryThis paper presents the design, operation, and impact of an AI Note Writer API for X's Community Notes, public since September 2025. Users, not the platform or AI, decide which notes display. The Community Writer, the largest AI API client, produced 52% of notes selected as Helpful and shown broadly. Among posts with Helpful notes, 42% had only AI notes and 30% had only human notes, so AI writing is additive. The writer is released as Apache 2.0 open-source software.
WhyDeploys a community-driven AI note-writing API on X at scale, with operational data analyzing human and AI note coverage; novelty and impact are solid, evidence is observational.
In briefFrom Tweets to Trades: How Public Mood Relates to Turkish Stock Market Performance
SummaryThis study analyses 610,422 posts from 176 curated X accounts in Turkey from 2022 to 2023, sorted into politics, economy and finance, and media and society domains. Turkish transformer models classify the posts, and the resulting public mood measures are compared with BIST100 and BIST30 market data using correlation, Granger causality, vector autoregression and impulse response analyses. The authors find public mood is not linked to the direction of stock returns but is linked to the size of price movements, with stronger relationships at longer time frequencies. Predictive relationships are concentrated in economy and finance communication.
In briefLARC: Low-Rank Adaptive Residual Connections for Learning in Frozen Models
SummaryThe paper proposes LARC, which adds a low-rank correction to hidden representations of a frozen model, using a slow state to learn shared starting factors and a fast state that adapts to feedback. In a four-candidate program-selection task, two feedback-gradient steps reduced expected query execution error substantially relative to resetting to the trained initializations. However, in a chronological replay of public continuous-integration jobs, retaining online updates increased half-Brier loss, so the authors separate residual capacity, adaptation from a starting point, and usefulness for later decisions.
In briefLess Data, Better Timing: Student-Curriculum Coupling for VLM On-Policy Distillation in Video Grounding
SummaryThis paper studies on-policy distillation for vision-language models in temporal video grounding and proposes Student-Curriculum Coupling (SCC), a closed-loop framework. A compact Anchor-Frontier curriculum defines candidate supervision, and the evolving student dynamically determines which supervision is active. Across three TVG benchmarks, SCC achieves a 5.1% relative gain in mean recall over Video-OPD while using 60.0% fewer training examples and cutting training time by 50.4%.
In briefEstimating causal effects with unmeasured confounders using proximal balancing
SummaryCausal effects cannot be identified from observational data when confounders are unmeasured. This paper proposes proximal balancing, which uses only proxies of the unmeasured confounders. It learns a low-dimensional summary of covariates and proxies that makes treatment groups comparable, then adjusts for that summary. The method needs no designated proxy roles, inverse problem, or latent-variable model. The paper gives identification theory, finite-sample guarantees, and an algorithm called PROBE, and tests it on low-dimensional, high-dimensional, and image proxies and on real data.
In briefWeighting pseudo-labels by quality to improve fundus image segmentation across domains
SummaryThis paper studies how to improve segmentation of the optic disc and optic cup in fundus images, which is relevant to glaucoma assessment. It extends Noisy Student training by using a multitask Dice predictor to estimate the reliability of pseudo-labels from the Teacher model, and uses these estimates to weight the Student's pseudo-label loss. Experiments on known and unseen domains report gains over the Noisy Student baseline, especially for optic cup segmentation.
In briefCoEvoWhen: Co-evolving Policies and Tools for Ultra-Long Video Temporal Grounding
SummaryThis paper proposes CoEvoWhen, a framework that jointly evolves high-level policies and executable media tools from a vision-language model's reasoning trajectories, forming a reusable skill without updating model parameters. An external skill updater distills transferable experience for temporal grounding in long videos and uses coding ability to upgrade or create tools. Experiments on five benchmarks and three VLMs show improved grounding accuracy on ultra-long videos and lower visual token cost, and the evolved skill also boosts general long-video QA performance.
In briefDistilling Gromov-Wasserstein embeddings into a model that maps new samples
SummaryGromov-Wasserstein multidimensional scaling learns low-dimensional representations from relational data, but it cannot embed samples it has not seen. The authors use a teacher-student approach: the teacher's optimal transport coupling is converted into targets by barycentric projection, and a neural student learns an explicit mapping for new samples. They extend this to multi-view settings. Experiments on synthetic and real data show the distilled models preserve geometry on unseen samples better than neural networks trained directly with the GW objective.
Rafael Pereira Eufrazio, Eduardo Fernandes Montesuma, Charles Casimiro Cavalcante
In briefMAGiDiff: Estimating Solar Surface Magnetic Fields from UV/EUV Images
SummaryThis paper presents MAGiDiff, a machine-learning method using denoising diffusion models to estimate vector magnetograms of the solar photosphere from stacks of UV/EUV filtergrams taken by SDO/AIA. The model is trained to reproduce disambiguated vector magnetograms observed by Hinode/SOT-SP. The authors report accurate agreement with Hinode ground truth and plausible active-region structures on full-disk images. They also state it generalizes across solar cycles and can be fine-tuned for STEREO/EUVI and GOES-R/SUVI, while noting it is not a substitute for a dedicated instrument.
In briefMGhana-ST: A Speech Translation Dataset for Ghanaian Languages and Multilingual Training Trade-offs
SummaryThe paper presents MGhana-ST, a speech translation dataset for four low-resource Ghanaian language varieties: Ga, Twi, Ewe, and Fante. Experiments use about 16.1 hours of paired speech and English translations and compare monolingual and multilingual training with Whisper-small. Under this data scarcity, flat multilingual training does not help any variety, and Ewe and Fante decline. The authors also report that an earlier single-run finding of positive transfer did not hold across seeds.
Frank Lawrence Nii Adoquaye Acquaye, Eric George Parakal, Jesse Johnson et al.
In briefRadian: Improving Autoregressive Video Generation with Representation-Space Adversarial Distillation
SummaryThis paper presents Radian, a distillation framework for few-step autoregressive video generation. It adds adversarial supervision on real video frames in the feature space of a frozen visual foundation model, complementing distribution matching distillation. The extra components are discarded after training, so the generator and inference denoising budget stay unchanged. Experiments on Wan2.1-1.3B report a VBench Total of 0.8444 and a VideoAlign Total of 0.8033 under four-step generation.
In briefIrekoGPT: Making Pruned LLMs Adjustable in Width at Inference
SummaryIrekoGPT is a post-hoc method that converts pretrained large language models into slimmable models whose width can be changed at inference time. Building on SliceGPT, it keeps the projection matrices unpruned so one model can expose nested subnetworks of different widths. Each layer is calibrated across multiple compression ratios, and downstream linear layers are corrected via gradient-free ridge regression. Preliminary results on Llama and Qwen models show gains over naive PCA-based slimming, especially at high compression.
Pietro Moriello, Pietro Buzzega, Angelo Porrello et al.
In briefOPTS-TTPO: Using Tree Search to Improve Policy-Gradient Learning from Limited Samples
SummaryThe paper introduces On-Policy Parallel Tree Search (OPTS) and Tree Trajectory Policy Optimization (TTPO), which sample new trajectory suffixes from the current policy to better cover rare high-return paths within a fixed budget while controlling gradient bias. Reported results include improved MuJoCo tail returns over PPO, a win-loss-tie record against PPO on Atari-57, and better avg@32 and pass@32 scores across four Qwen3 models.
In briefEfficient active auditing of multi-group fairness using targeted bias probes
SummaryThe paper introduces a bias probe framework that audits black-box models for multi-group fairness through targeted, adaptive queries, without reconstructing the model and so preserving its confidentiality. It proposes ALeBi, an active auditor that learns such probes, and provides sample complexity guarantees governed by a property-specific complexity measure. The authors also analyze adversarial settings where a model owner may strategically obscure bias, and experiments support the theoretical findings.|
In briefWorking Code, Unreliable Environments: Measuring Reproducibility in AI-Generated Software
SummaryThis paper finds that AI coding agents can produce functionally correct code while misspecifying the dependencies needed to run it. The authors propose a three-layer framework (declared, runtime-installed, and necessary-and-sufficient dependencies) and evaluate three agents across four languages and fifty programming tasks. Dependency set agreement for identical tasks was as low as 7%, and newer agents showed no meaningful improvement.
In briefWARP: A Benchmark for Testing How Well Invisible Image Watermarks Resist Attacks
SummaryThis paper introduces WARP, a unified framework for measuring how robust invisible image watermarks are. It covers 32 recent watermarking methods and 34 erasing techniques, including distortions, adversarial, purification, and re-embedding attacks. The results show some methods resist standard distortions but are highly vulnerable to re-embedding attacks.
Khaled Abud, Aleksey Yakushev, Aleksandr Akimenkov et al.
In briefFenchel Tilting: A weighted correction method for efficiently fine-tuning generative models
SummaryThe paper introduces Fenchel Tilt Flow Control (FTFC), a method for adapting a pretrained generative model to an arbitrary preference expressed as a utility function. It first fits an effective reward and density-ratio weights on pretrained samples using Fenchel duality, then uses these frozen weights to modify a diffusion or flow model in a single stage of importance-weighted training, without differentiating through sampling trajectories. The authors report that across image and molecule generation benchmarks, FTFC outperforms baselines on diverse preference functions and is up to 20 times more efficient.
In briefWho Verifies the Graph? Attacks on Causal Action Checks for AI Agents
SummaryThis paper studies causal action verifiers, which gate an AI agent's tool calls by checking whether each intervention is identifiable against a committed action-state graph. The authors corrupt only that graph and show that verifiers can execute many harmful actions while still issuing internally valid certificates. A randomized audit of executions detected both attacks with few false alarms, but beneficial actions rejected by misspecification still require separately paid auditing.
In briefAmortized Bayesian inference method for multilevel models of any structure.
SummaryThe paper presents a general method for amortized Bayesian inference on multilevel models of arbitrary structure. Given a generative model specified as a directed acyclic graph, it automatically derives valid posterior factorizations and matching neural network architectures through graph expansion and inversion. Unlike methods that simplify dependencies, it preserves all conditional independence and exchangeability assumptions. Across three case studies, it closely matches gold-standard samplers on models with over 6,500 parameters, reducing inference to a near-instant forward pass after training.
Daniel Habermann, Andreas Bulling, Stefan T. Radev et al.
In briefWeighted centroid search for exact incremental BPE
SummaryThis paper improves the local search step in exact incremental byte-pair encoding (BPE), which maintains the canonical tokenization after each appended byte. Each interval is weighted by the size of the recursive component it selects, so each move costs O(1+log(m/m')), and these costs telescope to O(log t) per append and O(n log t) over an n-byte stream. The authors also construct a vocabulary family where count-balanced search needs Θ(log² t) probes while the weighted search needs Θ(log t). The gain is a worst-case guarantee, not an average-speed improvement.
In briefCan We Anticipate Violence? Multimodal Cues Before Incidents
SummaryThis paper studies whether short-term risk of violence can be recognized from video signals shortly before an incident occurs. The authors build a Normal-versus-Risky classification task from 443 temporally annotated XD-Violence clips, with the incident itself excluded from the inputs. They compare facial appearance, synchronized audio, and body-motion features from tracked keypoints, and report that combining all three modalities performs best, reaching 91.21% accuracy on the held-out test set.
Sindhuja Penchala, Mohammed Yusuf Mujawar, Noorbakhsh Amiri Golilarz et al.
In briefMulti-link safety filter for VLA policies near moving hazards
SummaryThe paper studies how to keep a pretrained vision-language-action (VLA) policy away from moving hazards at run time without retraining it. The method covers the gripper, wrist, and forearm with five ellipsoids, filters every commanded motion through a barrier program, and uses sparse optical flow to track the hazard. In six simulated hazard-motion conditions, collision fell from 65.62% to 27.27%, and safe-success rose from 29.35% to 50.43%.
In briefDashVMC: Real-time control of a discrete world model in Geometry Dash
SummaryThe authors learn a compact, action-conditioned world model from about two hours of recorded Geometry Dash gameplay. A controller is initialized with behavioural cloning and then refined with PPO using only rollouts in the frozen model, with no further live-game interaction. The refined policies survive longer than their initializations on three official levels and a held-out community level. The system runs a 60 Hz capture-to-action loop on a consumer GPU.
In briefLimits of Causal Repair Depend on the Evaluation Target in a Gaussian Model
SummaryThe paper studies, in a scalar Gaussian causal experiment, how the gain of a learned causal repair differs from the best achievable gain. Under the usual 1/k learning scale, every feasible learner incurs a k^(-2) assessment floor. The authors characterize a sharp leading-log frontier in the magnitude-rich regime and propose a diagnostic-abstention rule that attains it. Results apply to this specific experimental setting and are not a general causal identifiability claim.
In briefLearning to Explain While Planning: Rule-Aligned Diffusion Planning for Autonomous Driving
SummaryThis paper proposes the Rule-Aligned Diffusion Planner (RADP), which adds differentiable driving rules to the diffusion training objective so the model learns rule knowledge beyond expert demonstrations. It also introduces Rule-Pressure Attribution (RPA), which uses gradients of rule losses to estimate online how strongly each rule shapes predicted trajectories. Experiments on nuPlan show improved closed-loop planning in safety-critical scenarios, and RPA attributions align over time with subsequent rule-specific risks.
In briefWhat Component-Replacement Evidence Can Establish: A Scoping Review of Local Decisions in LLM Agents
SummaryThis critical scoping review examines 348 studies and 90 comparison records to assess what evidence a component replacement in a language-model agent can establish. It separates task-level benefit from the contribution of local decision quality, using eight purposively selected cases. The authors conclude that online execution or simultaneous gains in local and task metrics alone do not establish that better local decisions explain task-level gains, and they derive eight claim-specific reporting items.
Shuyang Zhang (The Hong Kong Polytechnic University), Jianshuo Chang (The Hong Kong Polytechnic University)
In briefMid-Harness: Sampling and Verifying Actions Between Model and Execution Harness for Terminal Agents
SummaryTerminal agents act through model-generated commands, but a useful generated action may still fail to execute reliably. The paper introduces Mid-Harness, which samples several candidate actions, verifies them, and forwards one for execution without changing the generator or harness. On TerminalBench-Lite, a GPT-5.6 Sol verifier raises Pass@1 from 50.00% to 68.03% using 8 sampled actions.
In briefLocating Easily Confused Movie Scenes in Full Films
SummaryThe paper notes that many movie moments differ only in fine details, which existing temporal grounding benchmarks rarely test. The authors introduce CoM²-Bench, containing query pairs that share most attributes but differ in one distinguishing detail. They find that retrieval is the main bottleneck, often selecting confusable counterparts, and propose the training-free Agentic Window Search, which raises mIoU from 18.5 to 26.8.
In briefRichard: A Voice-First Mobile System for Persistent Tasks
SummaryThis paper presents Richard, a prototype mobile system that uses voice as the primary interaction method. It manages voice sessions, task execution, and result delivery separately, linking them through persistent request records so users can leave a conversation and later review, revise, or retrieve their work. The authors evaluate the design using Android functional records, controlled lifecycle verification, and execution records of a real programming request. Results show the backend continued progressing and returned failure feedback after the client disconnected.
In briefBioASQ 2026 overview: a biomedical semantic indexing and question answering challenge
SummaryThis paper reports on the fourteenth BioASQ challenge, held within CLEF 2026. The series covers biomedical language tasks such as semantic indexing, information extraction, question answering and summarization. In 2026 it ran six shared tasks, with 87 teams submitting more than 1000 runs. The authors note that several submissions reached competitive performance.
Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara et al.
In briefUBTree: Parallel Tree Drafting Using Unigram and Bigram Models for Faster Language Model Decoding
SummaryThis paper introduces UBTree, a parallel drafter for speculative decoding that pairs a unigram proposer with a bigram selector to build draft trees. The selector is trained with a renormalized KL objective on high-temperature data to encourage plausible alternative branches. Across seven benchmarks with Qwen3-4B and Qwen3-8B, the authors report average speedups of 5.84 to 6.94 times over autoregressive decoding, and they report advantages over DARTree and DSpark.
In briefHow VLA robot models fail when instructions and visuals conflict, and a training fix
SummaryThis paper studies why vision-language-action robot models fail when a task changes in a way that requires a different action. The authors find that models often replay a familiar trajectory tied to the instruction instead of combining language with vision to choose the correct action. They propose Equivariant Counterfactual Training (ECT), which uses paired demonstrations where the same instruction needs different actions and a loss that trains each pair together. Reported results include position-swap success rising from 36% to 59% on LIBERO-PRO and unseen-position success on a real UR5e rising from 8% to 88%.](en)
In briefTACTIC: Context-Aware LLM Planning for Roadside LiDAR Attacks
SummaryThis paper presents TACTIC, a framework that uses a multimodal large language model to coordinate the choice and parameters of roadside LiDAR attacks based on traffic conditions, under a gray-box threat model. It builds a semantic scene graph from roadside perception data and imagery, selects between a push-away primitive and a phantom-obstacle braking primitive, and runs reasoning and execution asynchronously. In 280 randomized CARLA trials, the full policy reached a 100% collision rate, compared with 35% for a fixed rule and 60% for random selection.
In briefWhat Limits Recursive Reasoning Models: Optimization, Architecture, and Test-Time Scaling
SummaryThis paper studies recursive reasoning models, which repeatedly apply a small shared Transformer block to refine a latent state, across six algorithmic domains under a unified pipeline. It finds that stable, generalizable performance depends mainly on an intermediate gradient horizon, large physical batches, and controlled updates of the recurrent state, with no explicit hierarchical architecture needed. Based on these findings, the authors build a stable 13.6M-parameter model that reports 71.2% Arithmetic OOD accuracy and 98.41% on Sudoku.
Yuliana Shakhvalieva, Dmitrii Kharchev, Viacheslav Bezrukov et al.
In briefJev-as-a-Judge for RAG Claim Verification: Confidence-Gated Cascading
SummaryThis paper compares a non-generative decision model that outputs a typed label with a probability, against a frontier LLM, on checking whether retrieved evidence supports claims. Both are evaluated on the same 495 human-labelled claims from LLM-AggreFact. The two are statistically indistinguishable on the conventional metric, but they make different errors, and the cheaper model admits more unsupported claims. A confidence-gated cascade that escalates uncertain cases reaches similar accuracy at lower cost, though a random escalation control performs equally well, suggesting the savings come from the escalation share.
In briefAIMS: An agentic AI framework for transferring multi-modal sensing and communication models from simulation to reality
SummaryThis paper introduces AIMS, an agentic AI framework for sim-to-real multi-modal integrated sensing and communication (ISAC). Given a natural-language deployment request, it derives a deployment-specific configuration and uses two cooperating agents to generate synchronized sensing and wireless data and configure mixture-of-experts learning. Experiments on the real-world DeepSense 6G dataset report improved vehicle detection and beam prediction over the considered simulation and fusion baselines.
In briefReconstructing the Dynamic World: A Unified View of 4D Scene Reconstruction
SummaryThis paper reviews 4D scene reconstruction, which recovers how the geometry, appearance, and motion of dynamic environments change over time from visual observations. The authors organize existing methods around scene representations, temporal modeling, reconstruction pipelines, and optimization objectives, and examine how design choices affect fidelity, consistency, motion representation, and efficiency. They also consolidate common datasets and evaluation metrics, identify limitations in current experimental practice, and discuss open challenges.
In briefThe life cycle of massive activations: how they arise, grow, and settle in transformers
SummaryThis paper studies how massive activations, very large residual-stream values linked to attention sinks in transformers, emerge, grow, and consolidate during training. Combining training-trajectory analysis with controlled interventions, the authors find that weight decay causally controls the turnover of their global scale. They propose a balance model in which AdamW-preconditioned growth opposes weight decay, explaining how peak timing and magnitude shift with the decay coefficient.
S. Aaron McClendon, Jorge Gallego-Feliciano, and Antonios Saravanos
In briefBetter Deck or Different Judge? Testing Agentic Harness Gains in Banking Presentations
SummaryThe paper studies an agentic system combining a 27B language model, financial calculations, narrative templates and validation checks to produce corporate and investment banking presentations. Multiple LLM judges scored the resulting decks against a baseline that generated decks directly from a short prompt. The full system scored 20.4 to 33.6 points higher out of 95, but margins against direct Opus generation ranged from -4.7 to +0.8 points. The authors also found that repeated grading shifts scores on unchanged decks, making small improvements hard to distinguish from judge variability.
In briefLearning When to Intervene in Coding Agents: A Lightweight Pre-Execution Sentinel
SummaryThis paper proposes HiSentinel, a hindsight-distillation framework that trains lightweight 0.6B and 1.7B sentinel models to decide before execution whether a coding agent's action should be intervened on to improve task completion. A privileged teacher uses recorded execution outcomes, and its judgments are distilled into a causal student that sees only the pre-action context and proposed action. The authors also introduce SWE-Intervene, an action-level dataset annotating whether actions should be allowed, redirected, or paused for human help. Reported gains reach up to 14% on SWE-bench Verified Mini and 10% on Ask or Assume.
In briefCoverage Before Control: Instruction-Guided Retrosynthesis Method That Follows Chemist Preferences
SummaryThis paper introduces RIGS, a two-stage framework for retrosynthesis that follows instructions about which precursor sets to favor. Stage A trains a language projector to learn which alternatives an instruction favors or discourages, and Stage B uses it to steer a frozen generative model through lightweight residual adapters. The authors train on nested one-to-many supports and report that broader support helps generate more alternatives and that RIGS can guide generation by instruction. They note the relationship between coverage and control is consistent across model scales but non-monotone.
In briefAdapting omni-modal LLMs to compressed inputs via ground-truth-free self-distillation
SummaryThis paper addresses the accuracy loss that omni-modal large language models suffer when their long audio-video token sequences are aggressively compressed. It proposes CAFD, a self-distillation framework in which a full-context version of the same model acts as a teacher providing soft targets to a compressed-context student, requiring no reference answers or rewards. Evaluated on Qwen2.5-Omni-7B across 125 conditions, CAFD improved 120 of them, with an average gain of 1.44 points and recovery of 26.9% of the accuracy gap.
In briefLearning to cover locally: neural optimization when each node sees only nearby information.
SummaryThe paper studies combinatorial optimization where each node sees only its k-hop neighborhood yet must contribute to a globally feasible solution, formalized as local set cover. It instantiates this on weighted multipoint relay selection in the OLSRv2 routing protocol and proves that one-hop-short deterministic selectors either fail coverage or lose a factor of the maximum degree, while GNN depth cannot overcome the horizon. A 3-layer GATv2 trained from CP-SAT optima reaches a cost ratio of 1.030 versus greedy's 1.138, and restricting it to one hop degrades it to 1.344.
WhyFormalizes local set cover under a hard information horizon, proves matching lower and upper bounds for GNN depth, and validates on real networks with substantial experiments.
Johannes F. Loevenich, Thies Moehlenhof, Laurin Holz et al.
