In briefThe Trinity: A Framework for Capability-Driven Self-Improving Multimodal Agents
SummaryThe paper presents The Trinity, a recursive self-improvement framework in which Creator, Explorer, and Evolver roles couple simulated world generation with the evolution of agent harnesses. Tasks are reorganized by shared capability requirements into a hierarchical capability tree, and reusable capability modules are evolved along it. The evolved harnesses improve mean performance by 34 points across 28 training benchmarks and by 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
In briefJaxnasium: A JAX reinforcement learning library using PyTrees for cross-suite, multi-agent training
SummaryThis paper presents Jaxnasium, a JAX-based reinforcement learning library that addresses fragmented environment suites and single-suite algorithm implementations. It represents observation and action spaces as PyTrees, allowing near-single-file algorithms to run on discrete, continuous, image, composite, and multi-agent environments without modification. Through a Gymnasium-style interface and wrappers for 11 suites, over 300 environments are available, and the library also provides tools for multi-seed evaluation and hyperparameter sweeps.
In briefRollback: Rewinding World Models to Improve Robot Policies
SummaryThis paper introduces Rollback, a method that turns world models from forward-only simulators into counterfactual engines. It rewinds to key decision points, branches into alternative actions, and compares outcomes to credit success to individual decisions. The system combines distributed world-model search with asynchronous real-robot execution and uses a lightweight LoRA pipeline to correct the model's dynamics with real transitions. On three real-world tasks, success rates improve over the strongest baseline by about 10.67%, 11.11%, and 11.85%.
In briefProton structure mode by mode: a neural tangent kernel analysis of a DIS fit to parton distributions
SummaryThis paper analyses how a neural network learns parton distribution functions from deep inelastic scattering data, using the neural tangent kernel (NTK) to describe its training dynamics. It extends an earlier single-flavour analysis to the multi-flavour case. The authors find that only a few leading NTK modes are needed to reconstruct the target distributions, while data noise is carried by sub-leading modes learned later in training.
In briefRetriever-Free Retrieval: Document Navigation with Decision Models
SummaryThis paper proposes DecisionRAG, which uses pretrained language models' decision-making ability instead of lexical matching or embedding similarity for retrieval. The model jointly evaluates candidate branches in one forward pass without generating tokens, navigating from broad topics down to individual passages. According to the paper, without task-specific training or a first-stage retriever, it outperforms dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions, at 0.7–1.8 seconds per query.
In briefAuto-Research Arena: A benchmark for evaluating autonomous research systems in language modeling
SummaryThe paper introduces Auto-Research Arena, a multi-target benchmark for evaluating autonomous research systems and discovering ideas in language modeling. It first decides whether results are admissible based on model quality and resource constraints, then measures progress among admissible results using the selected target. The authors instantiate it for language model pretraining with nine targets and evaluate seven research harnesses, finding that a simple sequential search baseline outperforms several sophisticated methods.
In briefEvolution strategies as moving toward a reward-weighted center of mass
SummaryThe paper offers a geometric view of evolution strategies (ES) updates, treating perturbed models as particles whose rewards act as masses. It shows that the mean-subtracted ES update exactly equals a displacement toward the reward-weighted center of mass, scaled by a factor depending on mean reward, for any finite population. In large populations or under antithetic sampling, the centroid lies on the model itself.
In briefVibeVideoBench: Testing AI Agents That Make Videos by Writing Code
SummaryVibeVideoBench is a benchmark that tests AI agents on creating MP4 videos with sound by writing code, rather than with diffusion models. It contains 239 English and Chinese tasks across 12 categories, graded with program checks, about 13k rubric checks, and judgments from VLM panels validated by human preferences. Across nine frontier agent systems, the best still fails a core requirement on about a quarter of tasks, and most lost videos come from renders that were started late, ran slowly, or never started.
In briefLearning humanoid loco-manipulation that reacts to contact force
SummaryThis paper presents FLORE, a framework that teaches a humanoid robot to control contact force while moving and manipulating objects. It collects robot-free demonstrations with force-sensing handheld grippers and body trackers, then trains a diffusion policy that takes vision, proprioception, and measured wrenches as input and outputs actions and per-hand force commands. A hybrid compliance whole-body controller with force tracking executes these commands while maintaining balance. Across three contact-rich tasks, the force-reactive policy succeeded in 80% of rollouts, versus 26.7% for a position-only policy.
In briefDo Vision Models Know That Three Is Three? A Guide to the Paper
SummaryThis arXiv paper asks whether vision encoders represent abstract concept changes, such as going from 2 to 3 or from clean to dirty, as linear directions. The authors study 46 vision encoders on 10 concept families and find that concept directions are near-universal within a single visual regime but often differ across regimes, a pattern they call fragmented linearity. Models with stronger local linearity are more likely to escape this fragmentation, and in some cases concept directions transfer across regimes as linear probes.
In briefOn the stability of learning: unstable when gradient signals arrive only intermittently
SummaryThis paper studies whether learning stays stable when gradient information is noisy and only intermittently available. The authors show that even if average learning dynamics converge to the correct optimum, learning variance can grow to infinity asymptotically. Using analytically tractable settings such as single-layer and multi-layer linear networks, they derive explicit conditions for stable learning that differ conceptually from existing stochastic gradient descent conditions. They conclude that unbiased credit assignment alone is insufficient, and sufficiently frequent and informative feedback is also needed.
In briefFL4N: A Feed-Forward Latent Model for Rendering Dynamic Scenes
SummaryFL4N is the first feed-forward latent model for 4D novel view synthesis, which renders dynamic scenes from new viewpoints and timestamps. Instead of explicit 4D Gaussian primitives, it learns temporal dynamics as residuals on a pretrained static latent prior, using a lightweight dynamic encoder to separate object motion from camera parallax. Trained only with an image synthesis objective, it reports state-of-the-art 4D view synthesis quality while rendering in real time.
In briefOSWAM: A one-step world action model for faster robot and driving action generation
SummaryThis paper presents OSWAM, a world action model that generates actions in a single diffusion step for autonomous driving and embodied control. It initializes diffusion with imagination tokens instead of random noise and adds privileged information distillation with future conditioning. The authors report results comparable to or better than multi-step baselines, about 3x faster than direct-policy WAM, with 40ms latency on an NVIDIA RTX 5090.
In briefPatient-specific rubric benchmark testing LLMs across nasopharyngeal cancer care stages
SummaryThis paper introduces NPC-RubricBench, a benchmark that evaluates large language models on multistage clinical reasoning across the nasopharyngeal carcinoma care pathway. It is built from longitudinal records of 405 patients across two cancer centers, comprising 941 patient-phase units and 17,688 patient-specific criteria. Testing 30 general-purpose and medical LLMs showed that strong diagnostic performance did not carry over to downstream care, and the dominant failure modes changed across phases.
In briefPreserve to Expand: Source-Aware Adaptation and Model Merging for Multilingual LLM-Based ASR
SummaryThis paper studies how to extend a multilingual speech recognition system to new languages without replaying source-language training data, so existing languages keep their performance. It works with connector-based Speech LLMs and compares target-update construction and merging strategies. SSU-Wanda+TIES gives the best balance of learning a new language and retaining source languages in single-target expansion, while recursive expansion in multi-target settings is highly order-sensitive.
In briefWork Agents: Systems, Capabilities, and Optimization
SummaryThis paper proposes a unified, work-centric framework that views work agents as systems combining foundation models and runtime harnesses to execute tasks in stateful workspaces and deliver verifiable outcomes. It introduces work persistence and work openness as two dimensions for characterizing working conditions and classifies digital work into five forms. The authors identify five core capability families needed for reliable execution, review optimization approaches and evaluation metrics, and discuss future research directions.
In briefMedVTok: a general visual tokenizer that encodes medical 2D images and 3D volumes into one shared token space.
SummaryThe paper introduces MedVTok, a visual tokenizer that maps 2D images and 3D volumes from different imaging modalities into a unified token space for multimodal medical AI. It adds slice differential regularization to preserve inter-slice anatomical coherence and multi-expert representation alignment to integrate knowledge from several medical encoders while keeping modality-specific information. The authors report training on 28M images and 70K volumes across 9 modalities, and state-of-the-art results on 24 benchmarks across classification, retrieval, segmentation, synthesis, VQA, and report generation.
In briefEBench Evaluation of a Frontier Agent: Direct Control Versus Harness With Policy
SummaryThis report evaluates GPT-6-Astra, a frontier multimodal model, as a robot-control agent on EBench and compares it with post-trained VLA and WAM policies. Two configurations are tested: direct waypoint control, and a harness in which the agent can invoke a post-trained policy for execution. Direct control performs competitively and varies least across generalization dimensions, but underperforms on long-horizon, tabletop, and high-precision contact tasks. Combining the agent with a policy approaches the strongest policy's overall success while substantially reducing token cost, though it remains below the standalone upper bounds due to policy invocation and grasp-disruption issues.
In briefSeeing Through Distractions: Context-Stable Attribution via the Least Core
SummaryThis paper studies feature attribution in computer vision, where a feature can lower a target-class score when added to a full image even though the predicted label stays the same; the authors call these contextual distractors. Under a weak average marginal monotonicity assumption, they derive quantitative bounds on the gap between semivalue-based and least-core attributions, along with sufficient conditions. They show that shifting semivalue weights toward larger contexts reduces the signed gap. Controlled image-classification experiments and evaluations on Salient ImageNet support these theoretical findings.
Sai Ganesh Nagarajan, Toshinori Yamauchi, Hiroshi Kera
In briefDo Jev-Style System One Models Resist Selection Bias?
SummaryLarge language models can still show selection bias in multi-option questions even when told to ignore option order. The authors shuffle answer options while keeping questions and candidate content fixed, testing two Jev versions, twelve Jev reproductions, and six generative LLMs. Jev's accuracy changes by under two points, yet its cumulative failure set grows by 27 to 29 percent relative to the original order. The study concludes that stable accuracy alone cannot show resistance to selection bias.
In briefGeoUG: Using Geometric Generation to Improve Spatial Understanding
SummaryThis paper studies how image and video generators' spatial consistency can improve spatial understanding. It proposes GeoUG, a unified model that combines depth and multi-view generation with spatial understanding, and an adaptive interleaved reasoning method that drafts an answer and then decides whether to generate geometric evidence to confirm or revise it. GeoUG reports circular accuracies of 97.9% on CV-Bench-3D, 90.8% on 3DCueBench, and 94.2% on DA-2K, using fewer generation calls than always generating.
In briefBack to Basics: A Skill-Centric Approach for Web Browsing Agents
SummaryThe paper introduces MiniWeb, a miniature web environment of 65 interconnected sites with 369 annotated tasks, labeled by 37 reusable interaction primitives grouped into 7 skill families. Benchmarking ten open-weight agents, the authors find drag and gesture skills to be the weakest. They propose MiniWebGen to generate verified skill-focused training data and MiniWebAgent, which pairs a planner with the trained executor to improve task success on unseen sites and external benchmarks.
Minh Pham-Dinh, Md Farhan Ishmam, Hernan Yupanqui et al.
In briefA paper on teaching a robot to follow tasks by training a vision-language model in simulation rather than on real robot data.
SummaryThe paper introduces ZAP, a method that adapts a pretrained 27B-parameter vision-language model to control a robot through a simple tool interface. It uses supervised fine-tuning and GRPO on simple simulated tasks, with no real-world robot data during training. The authors report 96.7% success on real-world tasks, comparable to GPT-6 Astra and above the MolmoAct2 VLA (33.3%), and they also report generalization to unseen objects, goals, and an unseen single-arm Franka robot.
Benjamin Pikus, Daniel Saunders, Mircea Mironenco et al.
In briefAutomatic editing of multi-camera concert videos using audio and visual cues
SummaryThis paper presents a modular framework for editing multi-camera concert footage. It combines instrument activity from multitrack audio, visible instruments, image sharpness, beat-based timing, and probabilistic camera selection through an interpretable scoring model. Component tests and an exploratory expert study on a four-camera concert produced mixed assessments across editing variants. The authors suggest adding contextual and sequence-level criteria in future work.
Lennart Schiweck, Adrian Kempf, Michelle Retig et al.
In briefExploratory creative pre-ranking: hierarchical Bayesian method for choosing which ad creatives to show in real time
SummaryThe paper introduces a creative pre-ranker that scores every recalled ad creative and shortlists each campaign's pool before the per-user ranker runs. Creatives are organized in a hierarchy of audience segment, advertised item, campaign and creative, so new creatives borrow from their parent levels, with the borrowing set in closed form from historical logs. Replaying Roblox Sponsored-ads traffic, the shortlist raised predicted play-through rate by 8.14% and forwarded over 40% fewer creatives than the full pool, adding only 0.2% to serving time. The method now serves in production at Roblox after an A/B test.
In briefUsing causal graphs and past trial data to design reinforcement learning that helps people keep up physical activity over time.
SummaryThe paper presents an online reinforcement learning system for the upcoming ADAPT factorial trial, which gives participants personalized activity suggestions twice daily to improve their affective experiences with physical activity. An expert-specified causal directed acyclic graph guides state construction, estimation of partially observed states, delayed credit assignment, and representation of beneficial and adverse intervention pathways. A longitudinal simulator calibrated from a prior micro-randomized trial is used to compare candidate RL designs before deployment.
Xueqing Liu, Young Hyun Cho, Bhanu Teja Gullapalli et al.
In briefDoes VGGT Know When Its Camera Poses Are Wrong?
SummaryVGGT, a feed-forward 3D reconstruction model, predicts camera poses and geometry in one pass but gives no uncertainty for them. On HardMatch, a set of 1,000 difficult image pairs where VGGT fails on 35%, a linear probe on its frozen features predicts pose failure on held-out categories with 93.5% ROC-AUC, above VGGT's own confidence (79.7%). The probe also catches confident errors and transfers without retraining to rotation failures on other datasets.
In briefAnchored Stein variational gradient descent for sampling non-smooth densities
SummaryThe paper introduces an anchored variant of Stein variational gradient descent for sampling from a density proportional to e^{-U}, where U may be non-smooth. It uses a smoother reference potential U0 and an anchor weight a=e^{U-U0}, so the target can be sampled without computing the gradient of U, and standard SVGD is a special case. Under assumptions such as U being locally Lipschitz and U0 continuously differentiable, the authors derive a KL first variation formula and a nonasymptotic convergence bound in terms of an anchored kernelized Stein discrepancy. They also give guidance on constructing a suitable reference potential U0 from U.
In briefLearning When to Commit Tokens: Faster Diffusion Language Models via Schedule Distillation
SummaryDiffusion large language models can generate several tokens in parallel but often need many decoding rounds. This paper proposes dSpark, which distills unmasking schedules that commit tokens whose reveal makes other positions easier to predict, reducing uncertainty earlier. A KL-divergence regularizer helps preserve generation quality. On GSM8K, the method reports 10.12x and 5.12x speedups over baselines on LLaDA-8B-Instruct and Dream-7B-Instruct, respectively, while maintaining task performance.
WhyProposes schedule distillation to learn token commitment timing for diffusion LMs, with multi-model, multi-baseline evaluation on GSM8K showing substantial speedups.
In briefA game-theory study of how AI affects academic effort and hiring
SummaryThe paper examines how fields like mathematics face hiring challenges as AI makes research papers easier to produce. It builds a two-sided game-theoretic model of applicants deciding whether to deepen understanding or publish more, and recruiters deciding whether to test understanding. The authors classify all pure-strategy equilibria and use learning simulations to study the behaviors populations may adopt.
In briefFOCUS: An RL Method for Long-Horizon Language Model Agents
SummaryThis paper presents FOCUS, a reinforcement learning method that helps language-model agents on long tasks assign credit to earlier actions when rewards are sparse. It uses selective branching so the critic network samples continuations from states expected to most reduce its value-estimation error, and scores segments with a TD(λ) advantage. Under matched per-step generation budgets, FOCUS outperforms GRPO and VinePPO on legal reasoning, deep search, and code localization benchmarks. At Qwen3.5-9B scale, it scores 55.4 versus 48.2 on BrowseComp-Plus and 56.5 versus 52.4 on DeonticBench.
Zhengping Jiang, William Jurayj, Liaoyaqi Wang et al.
In briefMeasuring how LLM research agents over-smooth and overstate claims
SummaryThis paper examines how large language model research agents synthesize many documents into confident reports that do not faithfully reflect the underlying sources, a failure the authors call smoothing. To quantify it, they introduce SmoothScore, a pipeline that extracts conclusive claims from generated reports, retrieves and clusters supporting evidence by high-level aspects, and scores each claim against that evidence. Across 7 systems and 4 datasets, the agents consistently produced claims stronger than the evidence warranted.
In briefSelective Tie Learning: Reducing Spurious Correlations in Preference Optimization
SummaryThe paper studies how preference optimization can learn spurious correlations and how adding ties (equally preferred responses) to preference data can reduce this reliance. For log-linear policies, the authors derive each tie's influence on spurious reliance and show that selecting ties with the largest predicted reductions is first-order optimal. Across log-linear policies, neural models, and large language models, the proposed selection rules achieve comparable reductions with fewer ties than uniform selection.
Christian Moya, Purav Matlia, Elliott Thornley et al.
In briefA theory of how noise shapes what recurrent neural networks learn
SummaryThis paper studies how intrinsic noise, as found in biological neural networks, affects representation learning in recurrent neural networks. The authors use an analytically tractable model with a linear activation function and white noise inputs. In the rich regime, where weights start small, noise changes how the input, output, and recurrent weights align and breaks solution-space degeneracies. For sufficiently high noise, a non-degenerate solution is not optimal when parameter scales are limited.
In briefDo Coding Agents Still Need a Debugger? Fault Localization with Frontier Models
SummaryThis paper evaluates how accurately coding agents locate faults with no tools, with optional debugger access, and with debugger use encouraged. The authors seeded 610 single-line faults into 144 LeetCodeDataset solutions and rewrote each program with five behavior-preserving transformations. Without tools, agents localized 96.7 to 99.9% of faults, and encouraged debugger use removed most remaining errors at 14.7 to 27.0 times the token cost.
Christoph Daxerer, Johann Glock, Christian Macho et al.
In briefOmniBrain-0: A Lightweight Unified Agent for Robot Manipulation and Navigation
SummaryThis paper presents OmniBrain-0, a compact embodied agent that coordinates robot manipulation and navigation through one shared high-level decision model. It covers task understanding, planning, progress tracking, and recovery from failures. The model is trained with a progressive framework combining embodied pretraining, domain-specialized agent distillation, and multi-teacher on-policy distillation, plus a rubric-guided recursive self-improvement loop. The authors report strong results across several simulation benchmarks and real-world manipulation tasks, with much lower inference and execution overhead than its closed-source teacher.
Xiaomi Embodied Intelligence Team: Yuechen Luo, Chong He, Fang Li et al.
In briefDECE: A Unified Arena Testing AI Agents on Ten Strategy Board Games
SummaryThis paper introduces DECE, a unified competitive environment covering ten Eurogames with shared interfaces and evaluation protocols. The main study evaluates 12 model-harness configurations in 480 three-player matches, with 1,032 completed matches across all studies. Results show broad cross-game ranking agreement alongside game-specific strengths, and the same model can perform differently across execution harnesses. The authors also find that performance advantages do not consistently yield better token or cost efficiency.
In briefFree to Look, Free to Reason: Simple LLM Harnesses Beat Hand-Crafted Root-Cause Agents
SummaryThe paper compares three simple LLM harnesses that do not prescribe the model's steps with five published LLM root cause analysis agents and with statistical and causal methods. Evaluation uses RCAtlas, 90 validated microservice-failure cases from eight public datasets. The simple harnesses achieve the highest accuracy on every model where both families were run.
In briefSignals Improve While the Encoder Declines: A Known-Answer Test of Label-Free Monitoring
SummarySelf-supervised training is usually monitored without labels, using loss and feature statistics. The authors trained LeJEPA encoders on synthetic images with known generating factors and found that, in nine of ten seeds, factor readability abruptly dropped and stayed low. Meanwhile the loss kept decreasing and most label-free signals kept improving, even rating the declining runs higher than the others.
In briefDo similarity measures really measure similarity? How user activity creates phantom correlation
SummaryThe paper argues that a substantial share of observed item-item similarity in recommender systems stems from differences in user activity rather than shared taste. Even when users consume items completely independently and differ only in consumption volume, common similarity measures still detect strong associations, which the authors call phantom correlation. They model independent consumption with a Rasch model, derive how each measure depends on item popularity, and compare this baseline with real-world benchmark datasets.
In briefCan clinical notes reveal hidden mental health risks in chronically ill inpatients?
SummaryThis paper studies whether mental-health risks such as depression, dementia, and suicidal ideation that emerge during hospitalization in chronically ill patients can be inferred from clinical narratives. The authors build a weakly supervised cohort of about 4,000 MIMIC-III patients using mental-health evidence from discharge summaries. A hierarchical state-space model learns representations from day-wise narratives and tracks their evolution to estimate risk. Preliminary results show improved binary risk prediction and distinct risk trajectories between risk and no-risk groups.
Sudeshna Jana, Siddhartha Laha, Manjira Sinha et al.
In briefEEGaze: Predicting visual search attention by refining saliency maps with EEG brain signals.
SummaryThis paper introduces EEGaze 140, a public dataset of synchronized EEG and eye-tracking recordings from 36 participants performing free visual search across 140 realistic scenes. It also presents EEGazeNet, a method that uses EEG signals to refine the output of a state-of-the-art saliency prediction model according to the user's cognitive state. The authors report that EEGazeNet consistently outperforms all baselines and ablation conditions.
WhyFirst method integrating EEG into visual search attention prediction, with a large dataset and consistent gains over baselines and ablations.
Mansi Sharma, Dinar Orazgaliyev, Antonio Kr\"uger et al.
In briefVidEdit3D: Flexible 3D Gaussian Editing Using Video Diffusion Models
SummaryThis paper presents VidEdit3D, which treats 3D scene editing as a trajectory-based video editing task to keep multi-view consistency. It uses pre-trained video diffusion models, with trajectory-guided rendering for keyframe selection, training-free flow-aligned editing with editing consistency guidance, and region-selective 3DGS refinement to limit changes to target objects. The authors report stronger instruction alignment and better multi-view consistency than baseline methods.
In briefCLOSURE: Checking Whether Sensors Still Agree in Multivariate Time Series Anomaly Detection
SummaryThe paper proposes CLOSURE, a method for detecting anomalies in multivariate time series without labels. It treats recoverability and cross-channel consistency as two separately learned tasks: a diffusion reconstructor conditioned on a discrete memory of normal temporal bases, and a relation-closure operator that predicts each channel only from its peers. The two scores are combined only through calibration on normal data. Experiments on public benchmark streams report that CLOSURE outperforms recent reconstruction-, frequency-, and topology-based detectors.
In briefChoosing models for off-policy evaluation using pairwise Bellman checks
SummaryOff-policy evaluation struggles to measure errors of fitted Q-functions and occupancy-ratio estimates directly from held-out data. The paper introduces pairwise Bellman validation (PBV), a tuning-free minimax method that selects among finite candidate sets using held-out transitions without auxiliary models or separate critic optimization. It gives finite-sample oracle inequalities relative to the best candidate, even under candidate-class misspecification. Experiments on Hopper and Deep OPE are reported.
Lars van der Laan, Vasilis Syrgkanis, Nathan Kallus
In briefSEMA-Auto: An autonomous multi-turn jailbreak attacker that watches victim replies
SummaryThis paper proposes SEMA-Auto, a multi-turn jailbreak attacker that reads the victim model's replies and decides by itself when to end the session upon success. It is trained in two stages: synthetic bootstrap tuning, then decision-supervised reinforcement learning. The authors report state-of-the-art attack success rates across several datasets and victim models, achieved within few turns.
WhyProposes an autonomous multi-turn jailbreak attacker beating static and interactive baselines across datasets and victims via two-stage training.
In briefLUCID: Measuring Local Variable Importance Using Conditional Density Estimation
SummaryThis paper introduces LUCID, a model-agnostic local variable importance metric and estimator. When perturbing a feature, it draws plausible values from the full conditional distribution given the other observed features, avoiding unrealistic counterfactuals outside the observed data. The authors establish theoretical links between LUCID and existing importance measures and provide an estimator with convergence guarantees and uncertainty quantification. Experiments on simulated and real-world data show its stable performance and practical utility.
In briefPreference-guided cooperation for multi-agent learning with several objectives
SummaryThis paper studies whether the multiple objectives in multi-objective multi-agent reinforcement learning can be used to help agents cooperate, rather than being treated as a separate difficulty. The authors introduce objective-wise counterfactual credit assignment and propose MIRL, with static (MIRL-S) and dynamic (MIRL-D) variants, which uses meta-gradient information so agents learn how to prioritise objectives. Experiments in a multi-objective multiple particle environment show improved Pareto-front quality over baselines, with the dynamic variant performing best.
In briefMulti-objective reinforcement learning with routed ensembles and trajectory attribution
SummaryThis paper studies multi-objective reinforcement learning, where policies must balance competing goals such as helpfulness versus honesty or speed versus energy use. The authors introduce objective-conflict coefficients, derive near-optimal regret bounds for online learning, and propose routed ensemble policy learning, which keeps a small set of basis policies and routes trajectories to the highest-potential one. They also present a kernel surrogate trajectory attribution method that estimates how feedback from some trajectories affects learning on others. Reported results include a 4.3% average performance gain for routed ensembles on Meta-World and MO-Gymnasium, a 50% improvement in correlation with leave-one-out analysis, and a 27% reward gain in trajectory selection.
Zhenshuo Zhang, Ruoxuan Xiong, and Hongyang R. Zhang
In briefTRACE: Memory-efficient time series class-incremental learning via temporal-spectral reconstruction
SummaryThis paper proposes TRACE, a replay-free framework that lets time series models learn new behavior classes without forgetting old ones, while keeping memory use and privacy exposure low. It uses a compact frozen temporal-spectral representation and lightweight class-specific autoencoders, classifying samples by reconstruction-based scoring. The authors report competitive accuracy across five benchmarks, about 98.26% lower mean retained memory than TS-ACL, and about 33.43% relative accuracy improvement on GrabMyo.
In briefThe Watchtower Imperative: Lifespan-Calibrated Evaluation for Frontier AI
SummaryThis paper argues that the gap between AI capability deployment and safety assurance is structural and cannot be closed by internal evaluations alone. It analyzes documented harms from 2017 to 2026 across the human lifespan, attributing them to systems being evaluated against an implicit adult user model. It proposes a Lifespan Vulnerability Matrix across six life stages, along with design principles and a three-tier external evaluation framework.
In briefMatAnyone-E: Efficient High-Resolution Video Matting via Sparse Memory Propagation
SummaryThis paper presents MatAnyone-E, an efficient extension of MatAnyone 2 for high-resolution video matting. It first propagates at low resolution to find uncertain regions, then performs high-resolution matching only in those regions using selected memory, reusing low-resolution results elsewhere, and adds a lightweight decoder. The authors report matting quality comparable to MatAnyone 2, an estimated 7.4x inference speedup, and peak GPU memory at 7.6% of MatAnyone 2's usage at 4K. On a single NVIDIA H200, it reaches 73.07 FPS at 1080p and 28.62 FPS at 4K.
In briefCan JEV Support Clinical Trial Criteria Assessment? A Comparison Study
SummaryThis paper examines how to tell unmet clinical trial conditions from insufficient evidence during patient prescreening. It compares JEV, a non-generative decision model, with schema-constrained Qwen Flash on 2,048 patient-condition pairs from TrialGPT and MIMIC-derived datasets. In batch mode, the two models performed similarly on TrialGPT and MIMIC-IV, while JEV scored notably higher on MIMIC-III, with lower latency and cost. The authors suggest JEV as an initial assessment aid, noting its outputs must be checked against patient records before eligibility decisions.
In briefMeasuring how diffusion models fail to learn data symmetries, using a simplified linear model
SummaryThe paper asks to what extent diffusion models fail to learn symmetries in data distributions, such as the rotations of a dog in images, and which sources of symmetry breaking matter most. Studying linear diffusion models, the authors find that noise in the learning target causes the most robust symmetry breaking, persisting even with abundant samples. Using local laws from random matrix theory, they derive formulas for average symmetry error and show that smaller symmetries are harder to learn.
In briefAuditing Reward Hacks in GRPO with a Limited Strong-Verifier Budget
SummaryThe paper studies how to spend a limited budget of expensive strong-verifier checks during GRPO reinforcement learning, where cheap checks like unit tests can be gamed by the policy. It proposes low-pass budgeted audits, which audit groups with the fewest weak passes first. In LeetCode and Countdown-Code experiments, this approach drives reward hacking to near zero, outperforming random audits that use the same budget or more labels.
WhyProposes low-pass budgeted audits with a proof, validated across 33 real GRPO runs; a strong, practically relevant result on reward hacking.
In briefQuantitative generalization bounds for multi-layer Transformers: longer training strings can outperform more short samples.
SummaryThis paper analyzes how Transformer language models generalize both to novel samples from the same distribution and to longer input lengths, treating these two axes jointly. For fixed-precision Transformers, it derives generalization guarantees in terms of parameter sizes, the quantization space, and the alphabet size. The authors find that increasing training length can be exponentially more efficient than increasing sample size, and experiments support these predictions.
In briefDetecting bus delays and anomalies in Niterói using autoencoder features
SummaryThe paper compresses raw GPS data from Niterói buses into 64-dimensional autoencoder latent vectors to classify individual trip delays and flag anomalous days. The latent representation does not improve delay detection over simple per-trip summary features. A day-level anomaly signal is stable across three autoencoders, but detection depends heavily on the alarm threshold, and two external checks of the score were circular.
In briefCIVER: A Causal Intervention Framework for Verifying Code Review Findings
SummaryCIVER is a framework that checks automated code review findings to separate genuine defects from false positives. It combines a review module, an intervention module that generates repair-oriented perturbations, and a verifier module that scores review reliability, optimized jointly through expert iteration. On SWE-Review-Bench, CIVER identifies 77% of genuine defects through targeted repairs. Cross-project and downstream repair evaluations show it generalizes across repositories and improves coding agents' task success rates.
Xinyuan Song, Hasibul Haque, Jay Somasundaram et al.
In briefTriage: Skipping, Skimming, or Refining Attention Tiles to Speed Up Video Diffusion Models
SummaryThis paper presents Triage, a training-free attention method that speeds up video diffusion transformers by assigning each attention tile one of three approximation levels: skip, low-precision NVFP4, or NVFP4 with residual correction. A lightweight router predicts each tile's error and upgrades a tile to a more accurate level only when the predicted error reduction justifies the extra compute. Across Wan2.2, HunyuanVideo, and LTX-2.5, the method reports up to 2.32x end-to-end speedup over dense FlashAttention and higher PSNR than the best baseline in each setting.
In briefExplainable multimodal AI to help reduce unnecessary breast biopsies
SummaryThis study proposes a multimodal vision-language model with Grad-CAM explanations to predict benign tumors among patients recommended for biopsy, especially BI-RADS 4 and 5. Trained and evaluated on 1,632 patients with four mammography views and radiology reports, it reached a best test-set AUC of 0.72, outperforming image-only (0.66) and text-only (0.68) baselines. The authors suggest it could help reduce unnecessary biopsies.
Sadam Hussain, Mansoor Ali, Daly Betzabeth Avenda\~no-Avalos et al.
In briefFrom Self-Confidence to Statistical Accuracy: Scaling Human-in-the-Loop Evaluation with Direct and LLM-Assisted Auditing
SummaryThis paper examines how to evaluate large language and multimodal models when their self-reported confidence is often miscalibrated. It builds a unified statistical framework comparing direct human auditing using confidence-stratified sampling with LLM-as-a-judge evaluation that corrects for judge errors. Experiments on MMLU-Pro and MMMU-Pro show that proportional allocation reduces error under imbalanced confidence, while only a few judge configurations beat the human-label efficiency condition once inference cost is considered.
Harine Choi (1), Chenchen Kuai (2), Zihao Li (3) et al.
In briefMiRL: Error-Targeted Reinforcement Learning for Fixing Terminal Agent Misbehaviors
SummaryThe paper proposes MiRL, which first identifies misbehaviors, recurring wrong steps across many environments, and then trains a terminal agent with reinforcement learning starting from the points where these errors occur. A rollout is rewarded only if it passes the tests and avoids the targeted misbehavior. On Qwen3.5-4B, eight misbehaviors account for 52.1% of wrong steps, and after training the policy repeats them 43% to 57% less often on unseen prefixes. MiRL scores 51.8% on Terminal-Bench Lite versus 47.4% for TMax-4B, matches it on Terminal-Bench 2.1 (21.3% vs. 20.6%), and uses 35% of its rollouts.
In briefA unified framework (OptIDC) for choosing the best input-side settings in click-through rate prediction
SummaryThis paper proposes OptIDC, a framework that jointly learns which samples, features, and embedding units to select for deep click-through rate (CTR) prediction. It adds a selection-weighted optimal transport mechanism to coordinate selection decisions across these input levels. The authors state the method is model-agnostic and can be integrated into existing CTR models. They report consistently higher AUC than existing input-side selection methods on four real-world benchmark datasets.
In briefNot One Belief: How answer strings, not just pushback, change model answers in conversational QA
SummaryThis paper studies how conversational question-answering systems correct their errors when users push back. Each follow-up is split into a speech act (re-ask, agree, or challenge) and a payload (an answer string, correct or not), and four confidence signals are measured on the same transcripts. Answers change mainly when the user supplies an answer string, not from the speech act alone, and the four confidence signals disagree with one another.
In briefGNN-VIEW: A Browser Tool for Inspecting and Explaining Graph-Based Earth-System Models
SummaryThis paper presents GNN-VIEW, an interactive browser-based tool that helps developers build and understand graph-based machine learning models of the Earth system. Before training, it exposes the graph's information pathways so developers can debug the graph structure. After training, it uses attribution-based explainability to show which input features the model used for a given forecast. The authors demonstrate the tool on graph construction for an ocean forecasting model and on explaining poor precipitation predictions during Hurricane Laura.
Alessandro Lucatelli, Ferdinand Unterhuber, Katharina Krenzer et al.
In briefEvoRouter: Low-cost multi-agent routing that learns from execution insights
SummaryEvoRouter is a framework that decides how to configure multi-agent systems for web queries, covering the language models, collaboration modes, agent counts, and interaction rounds. An expert distills reusable routing insights from configuration trajectories, and a teacher uses them with execution feedback to supervise the router through on-policy distillation, while cost-aware reinforcement learning balances performance and compute. Evaluated on 19 in-domain datasets, it reports 68.69% average accuracy, 7.23 points above the strongest routing baseline, with 50.95% lower inference compute cost.
WhyCombines insight extraction with on-policy distillation for multi-agent routing, a meaningful combination; evaluated on 19 in-domain datasets plus out-of-domain benchmarks.
In briefTrySee: A modular system that helps vision-language AI plan robot tasks in specialized fields
SummaryTrySee is a modular pipeline that combines a general vision-language model's reasoning with domain-specific perception modules and standard robotics components for closed-loop robot task planning. It uses replaceable vision preprocessing and a dynamic preprocessing selection mechanism to ground plans online, without fine-tuning the VLM or requiring end-to-end demonstrations. The approach was evaluated on planning tasks across five different domains, including a real-world robot disassembly task, and field tested in a brownfield environment.
Elias Hofele, Thomas Trapp, David Timmermann et al.
In briefAdapting LLM-SC to Sacred Heritage Reviews: Domain Extraction, Semantic Networks and Human Evaluation at Pashupatinath Temple
SummaryThe paper adapts the LLM-SC framework to Pashupatinath Temple in Nepal, using a GPT-4o extraction prompt with 12 entity types, four relation types and nine aspect-sentiment dimensions. It applies this to 2,500 English Tripadvisor reviews and evaluates entity, aspect and relation outputs against a two-annotator reference for 200 reviews. The authors report component-wise scores and state explicit reproducibility and generalization limitations.
Hari Sitaula, Manish Sakhakarmy, Pujan Pant et al.
In briefOmniShowcase: Multi-View Image-Based Video Generation Keeping a Subject Consistent from All Angles
SummaryThis paper addresses generating videos of a specific subject from a few reference images when the camera reveals surfaces not visible in the inputs. The authors propose OmniShowcase, which combines an automated data pipeline producing videos along diverse camera paths, including full 360-degree orbits, with a geometry-conditioning method that feeds surface-normal maps reconstructed from reference views into the video model. They also release Showcase-4K, a training corpus of 3D-asset-based videos, and ShowcaseBench, an evaluation set with source meshes and camera parameters. The reported experiments show state-of-the-art subject identity and 3D consistency with competitive video quality and motion.
In briefPrinted and Scanned: Testing Single-Image Morph Detectors After Print-Scan Damage
SummaryThis paper studies detection of face morphing attacks, where the biometric traits of two or more people are blended into one face image. The authors created a print-scan version of the SYN-MAD22 dataset at 300 and 600 DPI and tested eight detectors trained only on digital images, using two scanners. They found that detectors performing near state-of-the-art on digital data had their average equal error rate rise more than threefold.
WhyBuilds a print-scan SYN-MAD22 dataset and evaluates eight S-MAD detectors, exposing a large robustness gap; two scanners add solid evidence and practical relevance.
In briefCounterFlow: Flow-matching method for generating visual counterfactual explanations of AI decisions
SummaryThis paper introduces CounterFlow, a framework based on conditional flow matching that generates visual counterfactual explanations. It steers a source image toward a user-specified target class using target-conditioned vector fields. The authors report that it produces realistic semantic modifications at lower computational cost than GAN- or diffusion-based methods. They state that it achieves state-of-the-art results on standard benchmarks and works directly in pixel space.
Paul Sevellec, Elisa Fromont, Romaric Gaudel et al.
In briefDuplexTune: Memory-Efficient Mixture-of-Experts Fine-Tuning via Coupled-State Replay
SummaryThe paper presents DuplexTune, a framework that replays coupled hidden states through pretrained attention and expert modules to adapt mixture-of-experts language models. Trainable width bridges connect two residual streams to the backbone, with bridge warm-up preceding backbone adaptation and router weights kept frozen. On Qwen1.5-MoE-A2.7B, it reports 38.8 GB peak memory versus 64.8 GB for checkpointed supervised fine-tuning, and higher throughput, while reaching 67.2% MMLU, 75.6% GSM8K, and 7.69 MT-Bench. Its multilingual score is weaker than the checkpointed baseline.
In briefIndependent safety test of DeepSeek V4.1 Flash-1017 for loss-of-control and manipulation risks
SummaryThis paper presents an independent third-party safety evaluation of the open-weight model DeepSeek V4.1 Flash-1017, following the EU General-Purpose AI Code of Practice across 34 benchmarks. It finds substantial safety gains over the V4 generation and results approaching or matching leading closed-source systems on several dimensions, though some concerning behaviors remain. The authors recommend integrating safety evaluation into incremental releases, using process-level monitoring of reasoning traces and action logs, and developing private benchmarks.
In briefSCAFFOLD with Hessian Similarity: Sharper Convergence Analysis for Federated Learning
SummaryThis paper analyzes the SCAFFOLD algorithm for federated learning, which corrects client drift using control variates, under Hessian similarity across clients. It proves convergence guarantees with linear dependence on Hessian similarity for both convex and nonconvex objectives, extending beyond earlier quadratic-only results. The authors also establish matching lower bounds for the analyzed methods with constant local stepsizes. They further show that under partial participation, even identical clients may not gain the full communication benefit from extra local steps.
WhyExtends SCAFFOLD guarantees beyond quadratics to convex and nonconvex objectives with linear Hessian similarity dependence and matching lower bounds.
Zhirayr Tovmasyan, Slavom\'ir Hanzely, Egor Shulgin et al.
In briefMedJev: Zero-Shot System-One Decision Making in Medicine Evaluation
SummaryThis paper evaluates MedJev, an open System One model that runs locally on a laptop and answers structured medical questions in a single forward pass without generating text. The model is tested on ten benchmarks covering medical knowledge, diagnosis, and patient-level prediction using MIMIC IV and MIMIC IV ED. It approaches published GPT-4 accuracy on MedQA, USMLE, and NEJM and outperforms XGBoost on in-hospital mortality, but trails supervised models on operational outcomes such as ED triage admission. The authors also introduce a two-parameter recalibration method that reduces calibration error using 50 to 200 labeled examples.
In briefA hybrid framework that extracts causal chains from aviation safety reports and combines them with physics-based failure quantification and intervention analysis.
SummaryThe paper uses a local large language model to extract ordered causal chains from 56,202 American accident reports and aggregates them into a causation knowledge graph. On a 2,000-report sample, the chains recover 90.8% of investigator-flagged causes. The authors then build physics-based occurrence models for each factor and combine them with a leaky noisy-OR into a failure mode probability, illustrated with carburetor icing as a cause of engine power loss.
In briefA provably efficient Q-learning method for robust reinforcement learning
SummaryStandard reinforcement learning can perform poorly when real environments differ from simulators, so robust RL aims to handle such uncertainty. This paper proposes a sample-efficient asynchronous Q-learning algorithm for robust MDPs whose uncertainty is defined by total variation distance, using a single Markovian trajectory. It derives an upper bound on the sample complexity needed for an entrywise ε-approximation of the optimal Q-function. The authors state the bound is nearly minimax optimal and improves on existing results.
In briefSpan readouts and selective adaptation for better JEV entailment reranking
SummaryThis paper studies how to improve JEV decision models that rerank answer options through separate natural language inference requests. The authors replace the option-ending pointer with mean-span pooling over each option's states and keep low-rank adaptation only in attention and gated DeltaNet projections. On MC11, averaged over eleven reranking configurations, selective adaptation raises NeoHorse-1-4B accuracy over the all-module baseline while cutting trainable parameters by more than half, and mean-span pooling gives further consistent gains at the same parameter count.
In briefActive learning for conditional generative models using a transport neural tangent kernel
SummaryMany high-throughput experiments output distributions rather than single values, requiring conditional generative models to interpret data and predict outcomes at new conditions. The paper shows that mean-based active learning ignores distributional shape and proposes the transport Neural Tangent Kernel (tNTK), which measures how sensitive a generative model's predicted distribution is to its parameters. Used as a Gaussian process covariance, it greedily selects inputs that minimize posterior variance, and experiments on synthetic distributions, intent-labeled sentences, and single-cell perturbation responses show it outperforms a mean-based NTK baseline.
In briefTRON: Reusing Biological Protein Networks as Blueprints for Evolving Neural Networks
SummaryThis paper introduces TRON, a neuroevolution framework that uses experimentally measured protein-protein interaction (PPI) networks as architectural priors instead of random network topologies. It combines classical evolutionary operators with speciation and a species quarantine mechanism that protects newly generated architectures during early development. On three classification and three regression benchmarks, TRON achieved the highest predictive performance among the compared methods on all classification datasets and most regression tasks, with lower variability across runs but higher computational cost.
Konstantinos Zacharopoulos, Alexandra Bekou, Jose M Serradell et al.
In briefProofs of strong forms of the Bovdi–Geche universality conjecture and a complex analogue
SummaryThis paper proves two universality theorems for neural networks made of a single neuron. First, using the Weil bound on multiplicative character sums over finite fields, it establishes a strong form of the Bovdi–Geche conjecture for neurons whose activation is a modified discrete logarithm. Second, it generalises Aizenberg's universality theorem to complex binary neurons whose weight vector does not depend on the target function.
WhyProves strong single-neuron universality over finite fields and generalizes Aizenberg's theorem, answering an open question with rigorous proofs but limited applied impact.
In briefVPR-Space: A Benchmark for Visual Place Recognition in Space-Analog Environments
SummaryThis paper introduces VPR-Space, a benchmark for testing visual place recognition on extraterrestrial-like surfaces such as the Moon and Mars. It contains sixteen database-query datasets built from existing data collected in Space Analog Domains, covering changes in environment, viewpoint, and illumination. The authors define ground truth based on whether a match would be useful for a downstream SLAM pose estimate, and evaluate nine state-of-the-art methods, finding considerable room for improvement.
In briefSeeRF: A generative method that registers and fuses multi-modal images to see beyond each sensor's field of view
SummaryThe paper addresses the problem that multi-modal sensors have different viewpoints and fields of view, which leaves gaps in fused images. It proposes SeeRF, a framework that combines diffeomorphic deformation-based coarse-to-fine image registration with a generative diffusion module to complete and fuse non-overlapping regions. The authors report superior performance on multiple public datasets and real-world scenarios.
In briefTest-time adaptation for planning, not just prediction: Optimistic Test-Time Adaptation (OpTTA)
SummaryThe paper addresses degraded planning from frozen latent world models under distribution shift by proposing Optimistic Test-Time Adaptation (OpTTA). Its objective combines the prediction loss on observed transitions with the planner's goal-reaching cost on candidate rollouts, biasing adaptation toward models that are optimistic about reaching the goal. Across visual and environment shifts on manipulation, reaching, and navigation tasks, adapting only the predictor raises final planning success by 13.5 percentage points on average, versus 6.9 for the prediction-only objective (AdaJEPA).
In briefOffline language-based support tool for air combat tactical decisions
SummaryThis paper presents MATE, an offline language-based tool that supports pilot training and tactical decision-making in air combat. It uses retrieval-augmented generation to answer natural language questions from doctrinal and tactical manuals, and is built entirely from open-source components so it runs fully offline. Three Brazilian Air Force fighter pilots rated answers from three open-weight models, with average scores from 4.00 to 4.40 on a five-point scale, and Mistral NeMo 12B scoring highest. Pilots found the response time fully adequate compared to manual search, and the relevance of retrieved content was identified as the main area for improvement.
In briefFrom Language to Drives: Aligning Autonomous Robots with Human Purposes via Multimodal Models
SummaryThis paper proposes a multimodal framework that uses large language models and vision-language models to turn human purposes, expressed in natural language, into missions and drives that guide an autonomous robot's behavior. The system has three layers for semantic alignment, mission generation, and drive modeling, forming a bidirectional loop in which feedback maintains alignment with human intent. Experiments in simulation and on a physical robot show that the process produces valid missions and behaviors.
Diana Cerdas-Vargas, Alejandro Romero, Richard J. Duro
In briefSynthetic Deformable Tubular Benchmark: A Simulated Dataset for Testing 3D Reconstruction of Bending Tube-Like Surfaces
SummaryThis paper presents a synthetic dataset of 5,580 RGB images of tubular scenes generated with the open-source LumenMorph toolkit. It provides camera poses, geometry, depth, deformation fields, and persistent material correspondences for tracking surface points across frames. The authors use it to evaluate structure-from-motion, optical flow, and mechanical-prior methods. The benchmark is a controlled reference and does not claim anatomical or clinical realism.
\'I\~nigo Cirauqui-Viloria (Universidad de Zaragoza)
In briefTurning a neural network into equations to predict when planetary systems become unstable
SummaryThe paper addresses when multi-planet systems become dynamically unstable by training a neural network on N-body simulations and then distilling its learned knowledge into closed-form algebraic equations. The method is two-stage: it first learns an interpretable variable basis, then searches for equations using symbolic regression. The resulting equations reportedly outperform existing analytic instability estimates across wide ranges of mass ratios, spacing, and orbital eccentricities.
WhyCombines neural networks with symbolic regression to derive closed-form instability equations that outperform prior analytic estimates across wide parameter ranges.
Simon Alford, Edward Gu, Omar Costilla-Reyes et al.
In briefHierarchical Skill Evolution for Coding Agents: Distilling Reusable Procedures from Coding Trajectories
SummaryThis paper proposes Hierarchical Skill Evolution (HSE), which distills reusable procedural skills from coding trajectories without modifying the underlying model. It aggregates evidence from individual tasks into repository-specific skill candidates, then consolidates patterns across repositories into transferable skills. An external writer-evaluator loop generates and assesses skills while the coding model stays frozen. On SWE-bench Verified and DeepSWE, 10 skills evolved from SWE-Gym improved performance across all evaluated settings.
In briefMaking Reasoning Visible: A Diagnostic Framework for Adaptive Scaffolding
SummaryThis paper presents a diagnostic framework that analyzes students' action sequences in an interactive grammar learning game to reveal their step-by-step reasoning. It computes five interpretable strategy dimensions, estimates coordination demand, and classifies reasoning into four states. Analysis of 36,234 interactions from 234 students found that strategies vary mainly by context rather than stable traits, and that students often fail repeatedly without intervention. An early warning model based on strategy patterns predicts validation success with good accuracy.
In briefVaLoP: Value Gradient Lookahead and Pullback for Steering Diffusion Models Toward Rewards
SummaryThis paper proposes VaLoP, a method that steers pretrained diffusion or flow models toward outputs preferred by a reward model at sampling time, without retraining for each reward. The authors find that existing value-gradient estimates are least accurate at high noise levels, and they reduce this error by evaluating the value one sampler step ahead and pulling the gradient back to the current state. A cheaper variant, VaLoP-Fast, approximates the correction with a small learned head, and the authors report that both methods outperform existing guidance across models, rewards, and benchmarks, including critic-guided robot policies on LIBERO and RoboCasa.
In briefArchiveLens: A framework that turns historical maps into queryable spatial knowledge
SummaryThis paper presents ArchiveLens, a framework that treats historical maps as entity-centric, machine-queryable geospatial databases. It extracts building footprints with YOLOv11-seg, classifies building materials by HSV color thresholds, and parses attributes such as building use with a prompt-constrained vision-language model. Applied to 445 georeferenced Sanborn map sheets of Chicago from 1918 to 1950, the framework reportedly achieves high geometric and semantic accuracy, and it introduces a hybrid ST-RAG retrieval architecture.
In briefRoboDoctor: Training-Free Self-Improvement for Robot Policies
SummaryThis paper presents RoboDoctor, a training-free framework that lets robot policies and language-model agents adapt during execution using feedback from physical interaction, with model weights kept fixed. It revises decisions through diagnosis, grounded action, and reusable records of past experience. On a curated set of 54 RoboDojo-Sim task entries, it achieves a five-dimension mean score of 57.10 and a native success rate of 49.44%. The authors note that controlled trials are still needed to isolate the contribution of cross-episode experience.
In briefEvolve on Graph: Self-improving knowledge graph reasoning for LLMs without question-answer labels
SummaryThis paper proposes Evolve on Graph (EoG), a framework that lets a large language model improve multi-hop reasoning over knowledge graphs without any question-answer annotations. A single model alternates between proposing questions and solving them, with verifiable questions generated through symbolic topology synthesis and a curriculum that adjusts difficulty based on the solver's score variance. The authors report that EoG outperforms prompting-, retrieval-, and agent-based methods on six benchmarks over Freebase, Wikidata, and an industrial knowledge graph.
WhyProposes annotation-free graph self-play with verifiable topology synthesis and boundary-aware curriculum, consistently beating baselines across six benchmarks.
In briefKAPPA: a tool that checks statistics paper proofs using Lean 4 and AI agents
SummaryKAPPA is open-source software that uses language model agents run by Claude Code to translate the proofs in a LaTeX paper into Lean 4 and check them result by result. Each numbered result gets one of four verdicts: verified unconditionally, verified conditionally, issue recorded, or unfinished. Applied blind to three published statistics papers before their public corrections, it found the error the correction names, along with further problems the authors confirmed by hand.
WhyCombines LLM agents with Lean 4 to verify statistics proofs, finding a known published error blind across three papers; evidence is solid and useful for the subfield.
In briefAgentic Algorithm Design: LLMs Proving Runtime Bounds for Evolutionary Algorithms in Lean
SummaryThe paper presents a staged pipeline in which an LLM agent proposes black-box optimisation algorithms, Lean filters out flawed designs, and the agent then defines a potential function and proves drift bounds guaranteeing expected runtime improvement over a baseline. No empirical comparisons are made in the pipeline. Across 8 problem classes, 12 of 15 runs with frontier models produced a machine-checked expected-runtime bound.
WhyNovel LLM pipeline designing evolutionary algorithms with Lean-verified runtime bounds; evidence from 8 problem classes, 12 of 15 completed proofs, and ablation, impact mainly within theory of evoluti
In briefBeyond Distribution Matching: Self-Supervised Representation Forcing for Few-Step Video Generation
SummaryThis paper addresses few-step distillation of video diffusion models. It argues that Distribution Matching Distillation relies only on output-space supervision, which degrades spatial detail and temporal coherence as steps shrink. The proposed Self Flow Forcing adds a predictive representation loss to DMD, giving the student and an EMA teacher different noise levels so the student predicts teacher representations from more corrupted inputs. Experiments report reduced temporal flickering and improved spatial fidelity and motion quality.
In briefSAP-ViT: a vision transformer for medical images that explains its predictions patch by patch
SummaryThis paper proposes SAP-ViT, a Vision Transformer for medical image classification that drops the global CLS token and aggregates last-layer patch embeddings directly. Class predictions can thus be decomposed into patch-level contributions, with sparsity regularization keeping the evidence concise. An input-space back-mapping method estimates which input patches contribute to influential output patches. Evaluations on several medical imaging datasets report strong classification performance and consistent, interpretable patch-level explanations.
In briefGEAR: A Tool That Generates Multi-Agent System Code for Different Frameworks from One Specification
SummaryThis paper presents GEAR, an open-source tool for configuring multi-agent systems and generating framework-specific code from a shared specification. It uses a feature-based representation that makes configuration options and their constraints explicit, so specifications can be validated before code generation. A preliminary assessment found translation strategies for 271 of 320 evaluated feature-framework pairs (84.69%).
In briefLinear probes can be reused from base to distilled models despite direction rotation
SummaryLinear probes are small classifiers that read a language model's internal activations to detect harmful requests. The authors train probes on four base models and apply them unchanged to successor models, a distilled reasoning model and an instruction-tuned chat model, measuring how the separating direction rotates and how well the probes transfer. The probes remain usable across these model families, but part of what they detect is the surface appearance of harm rather than harm itself.
