In briefEvoAlloc: A self-evolving agent that allocates limited compute to promising programs during LLM-driven program search
SummaryThis paper proposes EvoAlloc, a resource-allocation agent for LLM-based program evolution that learns from search experience to revise how it assigns evaluation resources to candidate programs. It periodically consolidates past search and allocation outcomes into reusable experience, and uses counterfactual exploration to evaluate candidates denied resources to enrich that experience. On coding and agent-harness optimization benchmarks, it needs 59-82% fewer full evaluations and 61-89% fewer LLM tokens to reach baseline performance, and achieves 8.7-12.0% higher final performance under the same full-evaluation budget.
In briefDoes Context Change What Models Memorize? A Study of Prefix-Based Extraction
SummaryThis paper studies whether conditioning language models on retrieved context reduces memorization revealed by prefix-based extraction. Across three open-weight instruction-tuned models, the authors compare suffix probabilities under an empty prompt and under retrieved contexts of varying relevance. They find that context does not simply erase memorization: a robust core stays extractable, while context suppresses some boundary samples and enables others.
In briefNot All LLM Judge Verdicts Are Reliable: A Reliability Audit
SummaryThis paper audits the reliability of LLM-as-a-Judge, the standard method for NLP evaluation, by testing six frontier models across four benchmarks under varied prompts, presentation orders, temperatures, and repetitions. It finds serious weaknesses: verdicts change across repeated runs at temperature zero, swapping presentation order flips most verdicts on hard tasks, and the most consistent judge simply repeats wrong verdicts, matching ground truth only 51% of the time. The authors introduce a trustworthy verdict rate (T) metric and report that holistic rubric scoring improves trustworthiness more than pairwise win-rate or single prompt changes.Q
WhyIntroduces a unified reliability metric and derives an accuracy bound under position bias; broad multi-model, multi-benchmark experiments give solid evidence with notable influence on LLM evaluation.
In briefPerception Test 2026: Challenge Report and New City-Scale Audio-Visual Reasoning Tasks
SummaryThis report covers the fourth Perception Test challenge, held as a workshop at ECCV 2026 and focused on spatial intelligence across four tracks. It introduces two new benchmarks built on city-scale walking-tour videos, KilometerAudio and KilometerVision, and summarises the winning solutions in each track. The paper notes that the winning city-scale solutions relied on expensive agentic pipelines, while standalone multimodal models still struggle with complex spatial and multimodal reasoning.
Fedor Kitashov, Jo\~ao Carreira, Shiry Ginosar et al.
In briefUsing gradients to speed up Bayesian inference for costly simulators
SummarySimulators based on differential equations are used to infer input parameters from observed outputs, but each run can be computationally expensive. The paper adds gradient information of simulator outputs with respect to inputs to a Gaussian process surrogate used in Bayesian optimization-based active learning. Under a limited simulation budget, this significantly improves convergence speed. Reverse-mode differentiation keeps its efficiency gains after accounting for extra cost, while forward-mode differentiation does not outweigh its computational overhead.
In briefA compact optical-flow method for classifying tactile rotation under different gravity conditions
SummaryThis paper examines whether a compact representation of tactile motion can classify object rotation direction under Earth, Mars, Moon, and orbital gravity. Dense optical flow from a simulated GelSight Mini is aggregated over a 7x9 grid into 126 features. Models trained only on Earth data drop from 94.4% accuracy on Earth to 75.9% in orbit, while a single model trained across all four gravity domains reaches 96.3% overall accuracy.
Oscar Martinez-Bernal, Mario Cavero-Vidal, Francesco Grella et al.
In briefEnriched holomorphic neural networks for 2D and 3D Laplace problems
SummaryThis paper studies Laplace boundary-value problems, where singular behaviour near corners, edges and junctions is hard for smooth approximators to capture. The method combines holomorphic neural networks with singular functions built from local geometry and boundary conditions, so the representation is harmonic by construction and needs no interior residual. The approach is extended to three dimensions and tested on several problems, where the authors report it is the most accurate method on every problem. On a CdZnTe radiation detector with many singular sites, it is about 1.7 to 2.1 times more accurate than the next-best method in both 2D and 3D.
Kalp Shah, Andrei Vlad Nicula, Mihnea Andrei B\^arsan et al.
In briefLIVIN: A Benchmark for Robot Spatial and Embodied Intelligence Using Digital Twins of Real Homes
SummaryThis paper introduces LIVIN, a benchmark built from digital twins of 30 real, lived-in homes, preserving their room layouts, furniture arrangements, and everyday belongings. The twins were built with a human-in-the-loop workflow that reviews and corrects intermediate results against the source observations. The benchmark evaluates four tasks: 3D detection, 3D reconstruction, navigation, and loco-manipulation. The authors report that current methods still struggle with the dense object arrangements, occlusions, limited free space, and constrained interaction regions found in real homes.
In briefMAST: Motif-guided diffusion with tree search for inferring molecular structures from spectra
SummaryThis paper presents MAST, a method for determining both 2D and 3D molecular structures from spectroscopic data. It adds interpretable motif priors as intermediate guidance during denoising and frames diffusion sampling as reward-guided tree search to produce a compact set of spectra-consistent candidates. On the QM9S multi-spectra benchmark, MAST reports 94.89% exact recovery and improved 3D fidelity.
In briefILM: An AI-Powered Tool for Learning Islamic Stories
SummaryThis paper presents ILM, an interactive platform for teaching Stories of the Prophets, combining Arabic NLP, a knowledge graph, and retrieval-based question generation. A knowledge graph constructor extracts entities and narrative relations, which learners explore through a visual story map and answer generated questions about. For open-ended answers, an LLM acts as a judge, comparing responses against retrieved passages and reference answers. The authors report that the system shows the feasibility of this approach.
In briefWhen Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation
SummaryThis paper examines whether LLM-based agents need to reason before every action, since earlier reasoning may still support later steps. The authors find that drops in the likelihood of subsequent reference actions after removing extra reasoning indicate whether those actions remain recoverable from earlier reasoning. They propose RACE, a training approach that uses a LoGiC procedure to identify reasoning turns that can be removed, and incorporates this signal into supervised fine-tuning and agentic reinforcement learning. Experiments on four agent benchmarks report substantially reduced reasoning cost while maintaining or improving task performance.
In briefLook Back, Think Ahead: On-Demand Visual Memory for Efficient Multimodal Reasoning (ViMoD)
SummaryThis paper addresses the high computational cost of multi-step reasoning in multimodal large language models, which must process long visual token sequences from high-resolution images. It proposes ViMoD, a lightweight framework that keeps a compact visual context and retrieves original fine-grained tokens as needed, using DART for content-adaptive aggregation and TRACE for selecting fine-token groups based on decoding history. On Qwen3-VL-4B at a 20% visual token budget, the authors report outperforming all baselines on eight reasoning benchmarks, with only 0.0546% additional trainable parameters.
In briefStructure Tax: how structured output changes LLM accuracy
SummaryThis paper re-examines the belief that forcing large language models to output structured formats like JSON or XML inherently reduces accuracy. The authors tested several models, datasets and schemas, measuring accuracy, confidence calibration and internal representations. They found the accuracy loss depends on schema design: putting reasoning before the answer matches or beats free-form output, while answer-first ordering causes steep drops, especially in smaller models. The authors conclude that well-designed structured formats can match or exceed free-form performance, so the key question is how to structure outputs.
In briefDiffusion Models Avoid Langevin's Condition-Number Penalty in Gaussian Sampling
SummaryThis paper uses Gaussian distributions to study why diffusion models outperform classical score-based samplers. It proves 2-Wasserstein convergence bounds showing diffusion sampling error depends on dimension, step count, and the largest eigenvalue of the target covariance, while unadjusted and underdamped Langevin dynamics carry an extra square-root condition-number factor. The authors state these bounds are sharp, and argue the benefit of noising arises during sampling rather than during learning.
In briefPreserving Early Uncertainty to Reduce Overconfidence in Trained Models
SummaryThe paper observes that early-stage models are better calibrated, while continued training or fine-tuning brings small accuracy gains but substantially increases calibration errors. The authors attribute this to models gradually losing uncertainty awareness in their predictions and features. They propose EUA-Cal, which uses the early model as an uncertainty anchor through early prediction and prototype structure regularization. Experiments on image classification and multiple-choice question answering across eight models show it outperforms state-of-the-art calibration methods.
In briefMPGE: A Multi-Perspective Graph Explainer for Molecular Classification
SummaryThis paper introduces MPGE, a graph explainer that gives three complementary explanations for a frozen molecular classifier: factual support, counterfactual sensitivity, and exemplar tolerance. Tested with a GCN on MUTAG, Mutagenicity, AIDS, COX2_MD, and BBBP, factual masks kept 8.6% to 15.5% of input edges on average. Counterfactual coverage ranged from 4.8% to 67.6%, and exemplar preservation coverage from 98.9% to 100.0%. The authors state that these describe model behavior and are not validated chemical mechanisms.
In briefEfficient, generalizable archetype analysis for discrete data like counts and categories.
SummaryArchetypal Analysis describes data points as combinations of extreme typical profiles, giving interpretable summaries. This paper extends it to discrete data such as binary, count, and categorical observations using likelihood-based models for Bernoulli, Poisson, and multinomial distributions. The authors propose an optimization method using quadratic approximations and sequential minimal optimization, and choose the number of archetypes with cross-validated predictive likelihood. Experiments on synthetic data, single-cell RNA sequencing, microbiome, and somatic mutation data are reported as efficient and interpretable.
A. Emilie J. Wedenborg, Jesper L{\o}ve Hinrich, Morten M{\o}rup
In briefTraceable World State: A Framework for Tracking Robot Knowledge Origins and Replaying Actions
SummaryThis paper introduces Traceable World State (TWS), a method for recording where each piece of a robot's world knowledge came from, how confident it is, and how it was revised, instead of keeping only the latest estimate. Updates are immutable and ordered, and a SHA-256 hash chain makes logs tamper-evident, allowing deterministic replay. In tests on simulated Unitree G1 robot trajectories, the system replayed terminal states exactly and detected all injected corruptions, with about 1.72% storage overhead over plain JSONL.
In briefNatural Language to First-Order Logic: A Survey of LLM-Based Autoformalization
SummaryThis paper addresses the lack of a unified task definition and systematic survey for autoformalization targeting First-Order Logic (FOL). It defines the task by distinguishing Ontology Extraction from Logical Translation and shows how conflating them obscures cross-study evaluation. The authors review existing datasets, evaluation metrics, and LLM-based methods such as fine-tuning, prompting, and verification-based refinement. They also identify open challenges in benchmarking, semantic evaluation, ontology-aware methods, and end-to-end applications.
Andrea Brunello, Cristian Curaba, Luca Geatti et al.
In briefPolicy design for finite agent groups under probabilistic reach-avoid limits on group fractions
SummaryThe paper studies how to design policies for a finite population of Markov decision agents so that, with high probability, a required fraction reaches a target while the unsafe fraction stays below a threshold. It tracks the variance of the empirical density with a discrete-time Lyapunov recursion and uses the Cantelli inequality to turn chance constraints into tractable moment conditions. These conditions feed a gradient-based sequential convex approximation, and additional bounds give a finite-population certificate. The method is tested on a gridworld and an EV-charging aggregation problem and compared with a deterministic population-level linear programming baseline.
In briefHWAM: A Humanoid World Model That Jointly Generates Actions and Body States
SummaryThis paper proposes HWAM, a Humanoid World Action Model for robot manipulation that generates reference actions together with the robot's post-execution body state. This addresses the gap between a policy's reference actions and the motion the robot actually performs. The model is trained through three conditional paths: a policy path, forward dynamics modeling, and inverse dynamics modeling. On three real-robot tasks with the LimX OLI humanoid, HWAM achieved the highest success rate among the evaluated baselines, reaching 70.6% on Candy Picking versus 43.3% for Fast-WAM.
In briefInterviewPlayground: A Validated Simulation Environment for Testing AI Interviewers
SummaryThe paper presents InterviewPlayground, a simulation environment that evaluates AI interviewers using simulated participants whose behaviors are grounded in social theory. Simulated studies produce an InterviewReportCard based on validated measures. To test validity, the authors compared 15 real qualitative studies with five AI interviewers and 450 human participants against simulations, finding an average Pearson correlation of 0.86 across 12 measures.
Jonathan Ivey, Aimee Liang, Arthur Y.S. Wang et al.
In briefTesting How LLMs Assign Social Responsibility and Blame
SummaryThis paper presents the first systematic study of how large language models attribute responsibility and blame to agents in social situations. Guided by Attribution Theory, the authors build a benchmark of 7,639 judgment questions drawn from classic vignettes and real-world narratives. They evaluate 32 LLMs and 5 non-LLM baselines, and use probing to examine internal representations. Findings show partial agreement with human judgments that grows with model size, and some attribution dimensions are decodable from model hidden states.
WhyFirst systematic study of LLM social attribution, with a 7,639-question theory-based benchmark, 32 models, and hidden-state probing analysis.
In briefCausalDreamer: Predictive World Models via Latent Disentanglement
SummaryThis paper proposes CausalDreamer, a world model for control that keeps a pretrained video tokenizer frozen and re-encodes its latent into four factored groups, separating controllable from uncontrollable and reward-relevant from reward-irrelevant information. The dynamics model is then fine-tuned to predict this factored representation. On 20 MMBench2 tasks, it scores 14% higher than the pretrained world model on clean tasks and 25% higher on manipulated variants, while neither model meaningfully beats a random policy in new environments.
In briefPulseBound: Forecasting Future Heartbeat States With Strict Information Limits
SummaryThis paper presents PulseBound, a photoplethysmography (PPG) representation learner that predicts nine rhythm and morphology descriptors for up to four future heartbeats. A fixed cutoff ensures the model only sees the visible prefix, so withheld later samples cannot affect its inputs. On MIMIC and VitalDB held-out groups, it reduces transformed-space MAE relative to last-visible-beat persistence by 28.06% and 22.22%. It also achieves the best mean on most of 13 downstream tasks.
In briefGhost tasking: parametrized Gaussian processes for solving linear differential equations
SummaryThis paper proposes a systematic method called "ghost tasking," which uses auxiliary tasks so that even non-parametrizable systems can yield parametrized Gaussian processes, while keeping output dimensions and latent functions few. The authors test the method in three experiments and report that it performs especially well in inverse problems with very little data. They compare it with the only other currently available method that applies to all the experiments.
In briefComplexity of grounded and preferred semantics in finitary argumentation frameworks
SummaryThe paper analyzes the computational complexity of grounded and preferred semantics in Dung's abstract argumentation frameworks restricted to computably finitary representations. It classifies credulous and skeptical acceptance, extension existence, uniqueness, and non-empty existence problems. The results identify where finitarity reduces complexity to the arithmetical hierarchy and where second-order quantification pushes problems back into the analytical hierarchy.
WhyGives precise arithmetical and analytical hierarchy classifications for grounded and preferred semantics in finitary frameworks, refining known results with completeness proofs.
In briefREACT: Rolling denoising and dual decoupling for more reactive robot control with VLA models
SummaryThis paper addresses a trade-off in chunked robot control with flow-based vision-language-action (VLA) models: long action chunks are smooth but slow to react, while frequent replanning causes discontinuities. REACT keeps a persistent action buffer with staggered flow timesteps, denoising the full horizon with the latest observation at each step and executing the cleanest block. It also separates sensing, VLM encoding, DiT denoising, and action execution to support real-time control. The authors report better task success, lower reaction latency, and smoother trajectories than frequent-replanning and asynchronous baselines on the RoboTwin 2.0 benchmark and real-world tasks.
In briefTest-time compute methods for tabular foundation models: what helps and what doesn't
SummaryThis paper studies how extra computation at prediction time can improve pretrained tabular foundation models, across adaptation, aggregation, and context construction. It introduces DiagScale, which trains only 0.003-0.03% of parameters yet achieves gains comparable to full fine-tuning. Selective aggregation and adaptation give consistent benchmark gains, while context construction benefits depend on the task and data regime.
WhyProposes DiagScale and systematically studies three test-time compute axes, matching full fine-tuning on TabArena with solid evidence, useful for tabular foundation model research.
Kanghui Ning, Marin Bilo\v{s}, James T. Wilson et al.
In briefSUDO: A top-level ontology linking each paper's research narrative into global scholarly discourse
SummaryThe paper introduces SUDO, the Scholarly Upper Discourse Ontology, which defines how knowledge elements in a publication are organized into its research narrative, or scholarly discourse. Instead of fixing a single extraction unit in advance, it builds each paper's discourse bottom-up in a source-grounded way and connects papers into a global scholarly discourse. The authors demonstrate the approach across several scholarly knowledge-modeling use cases and report performance gains on an LLM-based scientific question-answering task.
Nipun D. Pathirage, Henrique Santos, Deborah L. McGuinness
In briefPolytopal Neural Network: Making Deep Network Representations More Interpretable
SummaryThis paper proposes Polytopal Neural Networks (PNNs), which enforce a polytope-based structure in each layer that is used directly in subsequent processing, making layer-wise representations interpretable. The authors scale the approach with learned corpus representations and amortized simplex inference, and show how it connects directly to vector quantized (VQ) training. Reported results indicate that the constraints preserve meaningful latent structure with minimal performance loss, and yield more compressed representations than VQ in unsupervised learning.
A. Emilie J. Wedenborg, Anders V. N{\o}rskov, Teresa Dorszewski et al.
In briefA Method for Interpretable PDE Solution Discovery Using Structural Experience Distillation
SummaryThis paper addresses discovering explicit symbolic expressions for unknown physical fields from observations under known physical constraints. The authors note that existing methods rely on a single terminal score, which gives little insight into why candidates succeed or fail. They propose SED-MCTS, which distills structural experience from evaluated expressions and reuses it to guide later search, using counterfactual subtree interventions to estimate local contributions. Across a suite of PDE benchmarks, the approach performs strongly under a fixed evaluation budget and improves efficiency and robustness under noisy or scarce observations.
In briefAgentic-TTT: Learning a Test-Time Policy for Test-Time Training
SummaryThis paper introduces Agentic-TTT, which learns a policy that decides when a language model should perform test-time training, which algorithm to use, and whether an existing skill can be reused. It turns TTT procedures into callable tools and trains the policy using the observed utility gains from its decisions. According to the authors, on their benchmark it nearly doubles the utility of the backbone model, trades off utility against compute, and generalizes to domains unseen during training.
In briefNeural network verification for deep joint source-channel coding decoders
SummaryThis paper presents the first bound-propagation framework for verifying the decoder of deep joint source-channel coding (DeepJSCC), bounding worst-case reconstruction error over a wireless channel's noise region. The authors extend linear relaxation methods to parametric ReLU, replace transposed convolutions with an upsample-then-convolution form, and model Rayleigh fading as a structural perturbation at the decoder input. Combined with Lipschitz-regularized global robustness training (GloRo), the method yields tighter certified bounds on an image transmission model, and over-the-air OFDM tests on software-defined radios show the certificate holds on real hardware.
WhyFirst bound-propagation verification for DeepJSCC decoders, handling PReLU, transposed convolution, and Rayleigh fading, with over-the-air OFDM validation.
In briefDataVista: Diagnosing Multimodal LLMs on Data Video Understanding
SummaryDataVista is a new benchmark for evaluating how well AI models understand data videos, which combine charts with narration. It contains 961 real-world data videos and 6,775 questions across three levels: data perception, temporal reasoning, and narrative understanding. Testing 19 multimodal LLMs, the best model reached 70.0% overall accuracy, well below human experts, with the weakest results in causal reasoning and narrative structure. Adding more frames and subtitles mainly helped perception and temporal reasoning, with limited gains for narrative understanding.
In briefAdaptTok-Indic: Script-aware adaptive tokenization for Indic languages, to split Indian-language text into fewer tokens.
SummaryMultilingual tokenizers trained mostly on English text tend to split Indic languages into too many pieces, making sequences longer and widening the gap across scripts. The paper proposes AdaptTok-Indic, which adds Unicode normalization, explicit conjunct handling, script-aware routing, and learned byte-level boundaries trained jointly with a small Transformer. On Bengali and Hindi it yields lower token fertility than SentencePiece BPE, with competitive downstream scores. It is a small-scale prototype, and no statistical significance testing or code-mixed evaluation is reported.
In briefMetaOPD: Learning Token Weights for Better Student Distillation
SummaryThis paper proposes MetaOPD, a bilevel optimization framework for on-policy distillation that jointly learns the student model and a lightweight token-weighting network. The inner step updates the student with weighted distillation, while the outer step optimizes the weighting network using validation loss on reference solutions after a virtual student update. Experiments on six mathematical reasoning datasets and three out-of-domain datasets report gains over standard OPD for two student sizes.</summary>
In briefFearCaut-Qwen: Steering a Vision-Language Model's Fear Direction to Change Its Hazard Decision Threshold
SummaryVision-language models used for post-disaster building damage assessment often under-report hazards, showing low recall despite reasonable accuracy. The authors use signal detection theory to separate perceptual ability from decision criterion, then locate an affective fear-related direction in the model and steer its activations at inference time. Injecting the fear direction raised Red (unsafe) recall to 75.7% without improving discrimination, indicating a criterion shift, and no retraining was required.
WhyCausally injecting a fear direction shifts VLM decision criterion via signal detection theory, with random and happiness controls; novel, though validated on one model and dataset.
In briefLearning shape parameterisations from example designs for Bayesian optimisation
SummaryThis paper proposes learning a shape parameterisation for Bayesian optimisation from a collection of existing designs, by applying principal component analysis to the deformations between shapes. The resulting linear, interpretable search space lets the number of components trade expressivity against dimensionality. The authors test it on aerofoils, wings, and radio-frequency cavities, from 2D geometry to 3D aerodynamics and electromagnetics, and report improved sample efficiency and exploration beyond hand-crafted baselines.
Gabriel Diaz-Aylwin, Joseph Neighbor, Abiel Malkani Talwar et al.
In briefSpecialized Decision Model Jev vs. General LLMs: Benchmark Results on Knowledge, Reasoning, and Multilingual Tasks
SummaryThis paper evaluates Jev, a specialized model that selects an answer from given options rather than generating text. It tests Jev on 13 multiple-choice benchmarks against 19 general-purpose large language models. Jev is competitive on knowledge and commonsense tasks and achieves the top score on MMLU-Redux and ARC-Challenge. It performs poorly on the MathQA math word problems, scoring below all 19 models.
In briefFast computation of quadratic entropy using distance sketches
SummaryThis paper presents scalable methods for approximating the quadratic entropy p^T d p for arbitrary distributions p under common negative-type distances, namely Euclidean and spherical geodesic distances. It uses random feature embeddings and projections, together with amortized large matrix multiplication and control variates, to reduce computation and memory costs. The authors demonstrate the approach against direct pair sampling and on Open Graph Benchmark datasets, identifying papers, fields, and institutions with narrow or broad interdisciplinary reach from citations and text features.
In briefCAPABLE: Adapting Robot Policies to Joint Faults Using Learned Capability Estimates (Paper)
SummaryVision-language-action robot policies assume the hardware they were trained on, so joint faults that change how commands are executed can cause failures. The paper proposes CAPABLE, which adapts a frozen policy by inferring each joint's actual motion capability from command-response history and adding bounded residual corrections learned with reinforcement learning, without fault labels. On 28 LIBERO tasks, success on an actuator excluded from fault training rose from 24.8% to 59.3%. The authors also report recovery on a physical Franka Panda robot.
Mohammad Khoshnazar, Mohammad Dehghani Tezerjani, Deyuan Qu et al.
In briefLearning Which Correspondences to Trust: Confidence-Weighted Event-Camera Localization in LiDAR Maps
SummaryThis paper studies localizing an event camera against a pre-built LiDAR map by estimating dense optical flow between a rendered depth view and event images, then solving for pose with PnP. The authors propose CELL, which learns a per-correspondence confidence end-to-end through a differentiable probabilistic PnP and uses it to reweight flow supervision, select correspondences at test time, and refine edge matching. On M3ED and DSEC, the full system improves over the LEAR baseline on most evaluated sequences, reducing median translation error by up to 26.9% and median rotation error by up to 15.8%.
Panagiotis Kiousis, Kuangyi Chen, Jun Zhang et al.
In briefMindFlow: A Framework for Generating Research Ideas via Controllable Thinking Flows
SummaryThe paper presents MindFlow, which models research idea generation as graph-structured thinking flows built from modular thinking operators and represented by a probabilistic mind supernet. Given a research topic, a controller samples these flows to produce candidate ideas, and a tournament-based relative ranking optimizes the controller toward higher-quality flows. The authors also propose an evaluation protocol that assesses both problem finding and problem solving, and report superior results across diverse topics.
In briefCan LLMs Fix Bugs Without Code? Automated Verification of No-Code Fixes
SummaryThis paper proposes an automated, execution-based pipeline that tests whether large language models can resolve invalid bug reports through no-code fixes, such as changing settings or updating versions, in a real browser. The authors evaluate 322 generated fixes using executor agents and an issue-specific checker. Even under the best executor, fewer than half of the fixes resolved the reported bug, and results varied substantially with the executor used.
In briefMiMo-V2.6: Scaling Reinforcement Learning for Model Self-Improvement
SummaryThis report presents the MiMo-V2.6 series, an omni-modal model family that aims to improve intelligence by scaling reinforcement learning compute. The authors perform multimodal mid-training before RL and scale RL along three dimensions: larger asynchronous batches, more diverse environments spanning code, general, visual, and cyber domains, and groupwise agentic grading for long-horizon tasks. They describe freezing the MoE router and a multi-layer defense against reward hacking to keep training stable. The training dynamics, RL environments, and RL framework are open-sourced.
Xiaomi LLM-Core Team: Zongming Qiao, Ziyue Hua, Zirui Ou et al.
In briefStochastic Grouping Conformal Prediction: A Method for More Reliable Subgroup Uncertainty Estimates
SummaryThis paper proposes Stochastic Grouping Conformal Prediction (SGCP), a method that improves how uncertainty sets perform across subgroups when standard conformal prediction only guarantees coverage at the population level. SGCP learns a stochastic grouping map so each sample draws calibration information from others with similar calibration behavior, producing a local score distribution. The authors prove it keeps the standard coverage guarantee, and experiments on synthetic and real-world benchmarks report smaller subgroup coverage gaps with prediction set sizes that are comparable or smaller than baselines.
In briefReliability-aware future conditioning for time-robust robot manipulation
SummaryThe paper shows that generated videos of a robot's upcoming task can harm performance when their timing does not match the robot's current phase. It proposes RAFC, which at each step estimates how much to trust the received clip, selects a nearby temporal hypothesis, and falls back to a static branch when neither fits. RAFC learns from task reward alone, without shift labels. On CALVIN and a Franka robot, it substantially improves success under temporal mismatch.
WhyFrames generated future guidance as reliability-aware control learned from task reward, with gains shown on CALVIN and a real Franka.
Mohammad Khoshnazar, Mohammad Dehghani Tezerjani, Zhiyuan Gao et al.
In briefActive view selection for occluded CAD reconstruction
SummaryThis paper presents SightCAD, a framework for parametric CAD reconstruction that treats the feasibility of camera views inside realistic indoor scenes as a constraint. A learned view selector picks K feasible views for a vision-language model that generates executable CadQuery code, and both are trained jointly using the geometric fidelity of the executed solid as reward. The authors report gains over baselines on DeepCAD, Fusion360, T-LESS and MP6D, along with the highest rate of executable programs.
Kartik Bali, Mahish Guru, Yiderigun Borjigin et al.
In briefInterval-Valued SHAP for Testing How Robust Tree Model Explanations Are
SummaryThis paper studies how stable Shapley value explanations are for decision trees and random forests. The authors use an imprecise Dirichlet model to randomly add a few unlabeled instances to tree leaves, producing interval-valued Shapley values. They derive theoretical results for efficient computation and show the approach extends to Banzhaf values. Case studies and experiments illustrate the method and its use in debiasing uninformative features.
Chenrui Zhu, Vu-Linh Nguyen, Marie-H\'el\`ene Masson et al.
In briefCrowdH-Edit: Reference-guided, instruction-based synthetic Hajj crowd images for crowd counting
SummaryAnnotated images of Hajj crowds are scarce, and existing crowd-counting datasets and general generative models do not capture Hajj-specific conditions. This study introduces CrowdH-Edit, which uses real Hajj images as structural references and text instructions to control crowd and lighting changes, and selects Qwen Image Edit for final dataset construction. From 6,122 synthetic images, 387 were annotated to form CrowdH-Edit-Ann-387. Adding this synthetic subset to real training data reduced mean absolute error for three of five crowd-counting models and root mean squared error for two, with benefits depending on the model and metric.
Amirah F. Alshammari, Bander A. Alzahrani, Nahed A. Alowidi
In briefWhen History Helps and Hurts: Selective Use of Dialogue History in Multimodal Conversations
SummaryThis paper introduces ReTurn, a benchmark of 7,000 base tasks with visual and audio evidence for testing whether models selectively use conversation history across multimodal turns. Across 13 models, median open-ended accuracy drops from 93.7% with direct input to 72.3% in conversation. Probes show models can recall earlier questions yet apply the task weakly, and competing current media can pull answers away from historical targets. Supervised adaptation gives only partial gains.
In briefScore-based learning of cluster DAGs from interventional data
SummaryThis paper addresses causal abstraction, where a low-level causal DAG over many variables is summarized as a smaller high-level DAG whose nodes are clusters of those variables. The authors introduce COARSE, a score-based method that, under linear Gaussian assumptions, replaces the constraint-based edge-learning phase of prior work and uses interventions to identify a causal order over clusters. They report that COARSE matches state-of-the-art edge recovery on synthetic and real data given enough samples, with a much faster edge phase.
In briefTACROSS: A low-cost system that learns robot touch from human tactile gloves
SummaryTACROSS is a system that collects touch demonstrations with a low-cost piezoresistive glove worn by humans and transfers them to robots. It aligns human and robot tactile signals at the level of contact events, mapping them into a shared 256-dimensional latent space, with robot demonstrations providing the only ground-truth action supervision. On four contact-rich manipulation tasks, the authors report a 3.5-fold efficiency gain over conventional teleoperation and a 95.7% reduction in demonstration equipment cost.
In briefRight Screen, Wrong Transition: World Models as Verifiers for GUI Agents
SummaryThe paper argues that safety for GUI agents depends on the transition caused by an action, not just the screen shown, so monitors that inspect screens alone can be fooled by reused legitimate screens. The authors present LGWM, a decoder-free, action-conditioned world model that predicts the next screen's representation directly, trained on 1.85M real GUI transitions without semantic annotation. On the RSWT-BENCH diagnostic, its training-free score reaches 0.987 AUC at 17 ms per decision.
WhyNovel decoder-free action-conditioned world model verifying GUI transitions in representation space; strong results on a self-built diagnostic benchmark, though comparisons are limited in scope.
In briefEfficient Survival Analysis Learning When Outcomes Determine Which Data Are Kept
SummaryThis paper studies training data where an observation's inclusion depends on its outcome label, a form of truncation. The authors prove that Cox regression under truncated and right-censored data can be learned efficiently with first-order optimization when the baseline hazard is known, despite nonconvexity. When the baseline hazard is learned jointly, they show the objective can contain strict suboptimal local minima. Experiments on four survival datasets show that accounting for truncation reduces bias from naive Cox regression, extends to neural networks, and is orders of magnitude faster than standard expectation maximization.
Itai Zilberstein, Ioannis Anagnostides, Arman Kilic et al.
In briefDoes matching a target mean recovering it? Evidence study of adversarial claims on encoders
SummaryAdversarial attacks on vision-language models often cite the attacked model's similarity score as proof of success, and this paper tests whether that score predicts recovery of the target by an independent model. Comparing six contrastive encoders under matched attacks at three perturbation budgets, the authors find that robustly trained encoders transfer more independent evidence than vanilla CLIP or SigLIP, yet no setting reaches the reference level derived from a 50% target-image blend. The authors conclude the score is informative only within a fixed robust encoder and propose a reporting protocol in its place.
In briefEliteTable: A self-improving method that searches for code-based robot control policies
SummaryEliteTable is an LLM-guided program-search algorithm that keeps a fixed-size table of top-scoring solutions to inform new proposals, edits, and combinations. It was tested on six Gymnasium control tasks with ten independent searches each. Candidate scores generally improved over generations, and held-out results improved in some tasks, but six of ten Acrobot-v1 searches declined. The authors note that comparisons and ablations are needed to establish its relative performance and component contributions.
In briefVocabulary extension for Nepali stalls at pre-tokenizer; a repair that preserves English, tested in continued pretraining
SummaryMany open LLMs use a pre-tokenizer regex that splits Nepali words at Devanagari vowel signs, making Nepali cost 2.6 to 5.8 times as many tokens as English. The authors propose a small edit to this regex plus 32,000 new merges, bringing Nepali to about 1.01 times English tokens, roughly 54% fewer than standard extension. In small-budget continued pretraining, the original tokenizer gave the lowest held-out Nepali bits per byte under every recipe tested.
WhyIdentifies a pre-tokenizer regex flaw splitting Devanagari at combining marks and proposes an English-preserving fix, but evidence is limited to small continued-pretraining runs.
In briefUnified forecasting backbone linking multivariate signals and narrative text flows
SummaryThe paper proposes MIDAPN, a unified spatiotemporal forecasting backbone that handles both numerical multivariate time series and pre-aligned textual narrative flows. It combines a Multimedia Identity-Aware Graph with Contextual Identity Modulation to model relationships across media, and Spectral Prism Convolution to perform hierarchical temporal analysis. The authors report evaluations on 13 multivariate and 12 multimodal datasets against 16 state-of-the-art models, plus long-context comparisons with 14 time series foundation models, claiming consistent superiority and shared-backbone compatibility.
In briefNeural Decoding as Cognitive Inference: Recovering Stable Cognitive States from Variable Neural Signals
SummaryThe paper reframes neural decoding as cognitive inference constrained by brain-intrinsic priors, rather than mapping neural activity to external labels. Across five neural recording modalities and three cognitive domains (motor, perception, and internal mentation), the method reorganizes the geometry of neural observation representations. It yields meta-neural semantic representations with consistent geometric relationships across tasks, allowing stable cognitive states to be recovered from variable neural observations.
In briefProject Greenhouse: Building Fully Open, Sovereign Models for Agentic Search
SummaryThis paper explores whether fully open and sovereign models for agentic search can be built with modest compute. The authors train a competitive pointwise decoder-only reranker from scratch, followed by supervised fine-tuning, using commonly available datasets. They avoid third-party open-weight backbones and run most experiments on a handful of GPUs. The release includes data, code, configurations, and checkpoints for their Gaggle model family.
Jimmy Lin, Sahel Sharifymoghaddam, Lingwei Gu et al.
In briefQGMem: Event-centric memory with query-aware graph augmentation for long-term dialogue agents
SummaryThis paper proposes QGMem, a memory framework for conversational agents that organizes long dialogue histories into event-indexed atomic memory units and consolidates them into dynamic memory traces that track state changes. For each query, hybrid retrieval and query-aware reranking activate the most relevant units as a compact working memory, which is organized as a local graph and provided to the LLM as a graph token alongside text. The authors report consistent gains across six benchmarks in retrieval, multi-hop evidence composition, conflict resolution, and ultra-long dialogue reasoning, with moderate inference cost.
In briefFrom Surface to Depth: Cognitive Appraisal Reasoning for Multimodal Emotion Understanding
SummaryThe paper argues that multimodal models relying mainly on observable emotional cues can reduce emotion understanding to superficial cue-label links. Drawing on appraisal theories of emotion, it frames the task as a progression from perception to cognitive appraisal and introduces the CogEmo-40K dataset, the sparse CogEmo-MoE model, and the CogEmo-Bench benchmark. The authors report that their approach leads on CogEmo-Bench and generalizes well across domains..
In briefNot Every Change Is Necessary: Recoverable Drift in LLM Unlearning
SummaryThis paper proposes PTP-U, a framework for machine unlearning in large language models that removes target knowledge while recovering non-target abilities. It first makes local analytic edits to weaken associations with the target knowledge, then aligns non-target output distributions with the original model under fixed forgetting constraints. Across three benchmarks, PTP-U reports the best forgetting-retention trade-off among evaluated methods, with 81.22%-91.03% forgetting and 94.20% non-target utility retained on average.
In briefPuffin: A probabilistic method for learning fine spatial detail from coarse data
SummaryThis paper presents Puffin, a probabilistic framework that uses high-resolution satellite embeddings as covariates to split coarse regional totals into finer subregions. Rather than predicting one value per subregion, it learns a probability distribution trained with an aggregation-aware likelihood. The authors evaluate it on German and US census, employment, and election data, and study when the disaggregation succeeds or fails across regions and targets.
Chaitanya Jobanputra, Sebastian Vollmer, Gerrit Gro{\ss}mann
In briefCost-aware mixture-of-experts coordination for model marketplaces
SummaryCurrent model marketplaces trade models as indivisible units, which limits the use of complementary strengths among different experts. This paper proposes a market framework that uses Mixture-of-Experts gating, coordinated by brokers, to combine heterogeneous experts into a composite service, with a cost-adjusted revenue allocation rule. Experiments on fifteen tabular and image benchmarks report the highest mean welfare on all datasets and lower expected cost than Standard MoE. The allocation method also has much lower computational overhead than exact Shapley allocation.
In briefBeyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations
SummaryThis paper addresses hallucinations in Large Vision-Language Models (LVLMs), where models generate content not grounded in the image. The authors find that LVLMs naturally tend to attend to visual evidence and that textual context can also help reduce hallucinations. They propose AIMS, a lightweight training-free framework that builds prototypes for visual, prefilled textual, and generated contexts and adaptively steers each attention head based on query affinity. Experiments across several LVLMs and decoding strategies show AIMS reduces object hallucination while keeping general multimodal performance competitive.]
In briefRobustLDS: Learning linear dynamical systems despite adversarial data corruption
SummaryThis paper studies how to learn linear dynamical systems from a single trajectory when some observations are contaminated by adversarial outliers. The authors propose several estimators based on relaxations of least-trimmed squares, solved with an alternating minimization algorithm, and two estimators that exploit group-sparsity of the outliers. For the group-sparse penalized estimator, they derive non-asymptotic error bounds showing robustness to outliers, and experiments indicate the methods work well in practice.
In briefCan Decision Models Understand Stance? Testing Jev Against General-Purpose LLMs
SummaryThis paper evaluates Jev, a decision model built for structured decision-making, on stance detection, which identifies an author's attitude toward a target. It compares Jev with four general-purpose large language models and two fine-tuned models on the English VAST dataset and the Chinese ZS-CSD conversation dataset. Jev performs competitively on VAST, matching GPT-5.6, but trails stronger LLMs on ZS-CSD, especially in separating favor from against.
In briefSeven forms of LLM-integrated applications, from chatbots to autonomous agent systems
SummaryThis survey examines how large language models are embedded in software and whether labels such as copilot, agent, and RAG reflect real architectural differences. It describes seven recurring forms in a shared vocabulary of agents and tools. Each form is characterized by architectural pattern, execution control and user intervention, number of agent calls per task, and tool use. An illustrative corpus of 22 systems shows where each form reaches its limits.
Irene Weber (University of Applied Sciences Kempten, Germany)
In briefBenchmark testing whether models can predict a program's exact output and internal state from code alone
SummaryThe paper introduces a benchmark for predicting a program's final output and checkpoint state from its source code and input, without tools or code execution. It uses 400 cases from 371 Python and C++ programs, with paired shorter and longer execution traces, and evaluates seven settings from four model families. Reasoning-enabled settings outperform non-reasoning ones by 33.1 to 55.2 percentage points, and the strongest setting scores 77.0% on longer-trace final output. The authors note that changed inputs and checkpoint tasks alter several factors together, so the gaps do not isolate trace length alone.
In briefSenseWeather: Multimodal Sensing for Robust Robot Vision in Extreme Weather
SummaryThis paper introduces SenseWeather, a benchmark with aligned RGB, depth, and infrared images captured under rain, snow, dust, fog, and darkness, along with clean references and perception tasks. It also presents Fusion-to-RGB, a parameter-efficient method that fuses these sensors to recover an RGB image compatible with existing downstream models. Results show RGB+D+IR fusion improves detection performance over RGB-only restoration across all compared detectors, though these gains do not consistently track standard image-quality metrics.
In briefHow automated research is evaluated: a survey of benchmarks and practices
SummaryThis survey reviews how automated research systems are evaluated across six targets: literature synthesis, ideation, executable workflows, scholarly writing, automatic peer review, and end-to-end research. It compares task construction, evidence sources, evaluators, and scoring procedures. The authors highlight three lessons and translate them into reporting and audit recommendations. The survey helps readers choose benchmarks and design future studies.
In briefRelative Patch Response Learning for detecting AI-generated images that generalize to new generators
SummaryThis paper addresses detecting AI-generated images, which is increasingly important as generative models produce realistic images that can spread misinformation. The authors observe that self-attention causes real and generated image patches to influence each other, making per-patch source labels imprecise. They propose Relative Patch Response Learning (PRL), which learns from how each patch's score changes between two mixed views of an aligned image pair. PRL outperforms the best prior methods by 4.3% on eight standard benchmarks and 5.9% on three in-the-wild benchmarks in average balanced accuracy.
In briefLearning structured dynamical systems from partial, missing data
SummaryThis paper studies learning structured linear dynamical systems constrained to convex sets when only a small subset of observations is available at each time point. It proposes an estimator minimizing a bias-corrected, potentially non-convex objective, with non-asymptotic statistical error bounds and convergence guarantees for projected gradient descent. The theory is applied to subspaces, bi-isotonic matrices, and matrices formed by sampling Lipschitz functions.
Aravinda Kanchana Ruwanpathirana, Hemant Tyagi, Sunny G.W. Wang
In briefUnderstanding Latent-Dimension Scaling in Dynamical-System Learning via Spectral Reliability
SummaryThe paper studies whether larger latent dimensions help dynamics learning through autoregressive prediction, analyzing learned evolution via Koopman operator eigenstructure. It uses relative residuals to detect spurious eigenpairs and compares latent-prediction loss with spectral-residual loss across six chaotic systems. The spectral-residual model achieved lower median rollout error in all systems and dimensions, with error decreasing more consistently as dimension increased.
In briefConditional Kernel Stein Discrepancy: Testing Whether Conditional Distributions Match a Target Model
SummaryThis paper studies conditional goodness-of-fit testing: given a possibly non-normalized conditional target model and samples from a joint distribution, it assesses how well the samples' conditional distribution matches the target. The authors extend unconditional kernel Stein discrepancies to the conditional setting using an operator-valued kernel on the covariate space. They show the statistic vanishes if and only if the conditional model and the true conditional distribution agree for almost all covariates. Experiments on smooth manifolds and discrete spaces evaluate level, power, and runtime, demonstrating the approach's viability.
In briefLEVER: a proof search method that optimizes proof quality during search
SummaryLEVER is a proof search algorithm for LLM-powered theorem provers that optimizes a programmable objective over correct proofs during the search rather than after a proof is found. It scores partial proofs on an AND/OR proof graph, combining realized values with predictions for open subgoals, while the Lean kernel guarantees correctness. On PutnamBench in Lean 4 under matched budgets, it cost 34% less than a strong single-conversation agent and raised the solve rate from 80% to 96%. Adjusting objective weights traces a quality-cost trade-off curve.
WhyProposes optimizing a programmable proof objective during AND/OR graph search, cutting cost and raising solve rate on PutnamBench with matched-budget comparisons.
In briefTrajectory-Guided Fault Localization for Improving Agent Skills
SummaryThis paper proposes SkillMorph, a method for evolving reusable skills used by code agents. It compares failure and success execution trajectories across repeated runs and tasks to identify suspicious actions, then uses them to locate skill edit sites and generate revisions. The authors report that on SWE-Skills-Bench and CannBot, evolved skills achieve higher trial-level accuracy and execution consistency than original skills and four existing methods. The method was also applied in an AI operator-development team, which accepted 6 skill-revision pull requests.
In briefPose-Free Feed-Forward 3D Scene Inpainting with Learnable Mask Attention and Support Token Refinement
SummaryThis paper presents FreeInpaint, a feed-forward framework that generates complete, 3D-consistent scenes directly from unposed multi-view images containing masked regions, without requiring camera poses. It extends a 3D foundation model to propagate masked regions from a reference view to other views, using a Learnable Mask Attention mechanism to handle masks that disrupt cross-view correspondence. A Support Token Refinement strategy injects diffusion-generated, confidence-weighted tokens to improve under-observed regions under severe occlusion. The authors report superior inpainting quality across diverse datasets with fast inference speed.</summary>
In briefGRPODropout: Selectively Dropping Rollouts to Keep RL Language Models Diverse
SummaryThis paper addresses policy entropy collapse in GRPO-style reinforcement learning for large language model reasoning, where loss of sampling diversity limits further improvement. The authors propose GRPODropout, which before the standard update removes a small number of high-probability positive-advantage rollouts and recenters the remaining advantages. Experiments report higher accuracy and higher actor entropy than original GRPO while using fewer rollout samples for updates.
In briefDADP: A Brain-Inspired Dynamic Method for Pruning Neural Networks During Training
SummaryThe paper proposes Dynamic Activity-Dependent Pruning (DADP), which scores connections during training by the accumulated product of pre-synaptic activations and post-synaptic error gradients. It uses a single global threshold instead of fixed per-layer budgets, so sparsity is allocated across depth automatically. Tested on MLP, VGG-16, ResNet-18, BiLSTM-CRF and MiniBERT, it reports 73.67% accuracy at 99% sparsity on ResNet-18 versus a 76.06% dense baseline, and entropy and effective-rank measurements suggest feature diversity is preserved.
In briefStructural and representational agency in LLM-driven timbre generation
SummaryThis measurement study examines a working timbre-generation interface driven by a large language model. The set of sounds a user can reach is precisely characterised, but the displayed position differs from measured position by roughly three times that region's width, with errors arising from model-predicted coordinates and from parameter interpolation. Of 36 refinement instructions, 26 produced no perceptible change, and a weaker model reported convergence while performing no action.
In briefVEDJE: A Video-Efficient Joint Encoder for Scalable Video-Text Search
SummaryThis paper presents VEDJE, a method for matching videos with text queries. It compresses features within sampled frames and stores them in a reusable cache, and uses feature-change prediction as an auxiliary training signal to improve retrieval without extra query-time work. On MSR-VTT, MSVD, DiDeMo, and ActivityNet, it improves R@1 over matched first-stage retrievers in both directions. A compact 12 KiB-per-video cache in one configuration keeps text-to-video recall within 0.2 points.
Shahaf Wagner, Gabriele Serussi, Dan Ben Ami et al.
In briefRecognizing unseen audio-visual events by fusing complex-valued similarity scores
SummaryThis paper addresses open-vocabulary audio-visual event localization, which labels video segments with event classes including ones unseen during training. The method computes complex-valued visual-text and audio-text similarities using a frozen multimodal foundation model and learns their fusion with a complex-valued neural network instead of a fixed scalar rule. It reports state-of-the-art results on OV-AVEBench's unseen-class split at 66.5/59.1/54.1% Acc/Seg-F1/Event-F1, and on a modified AVE dataset at 60.7/51.9/50.4%.
Anirudh Praveen, Koteswar Rao Jerripothula, Pratik Joshi et al.
In briefUsing Large Language Models to Detect Spin in Clinical Trial Reporting
SummaryThe paper notes that spin, which distorts how trial results are presented, appears in over 50% of randomized controlled trials that fail to reach statistical significance. The authors built a system to automatically detect outcome switching, using 300 labeled pairs of primary and reported outcomes, with prompt engineering, token-probability classification, and majority voting. On a test set of 2,496 examples, it achieved an F1 score of 0.78 and accuracy of 0.90, outperforming baseline text similarity models but trailing fine-tuned BERT.
Tja\v{s} Ajdovec, Marko Robnik-\v{S}ikonja, Simon \v{S}uster
In briefConventionPlay: Capability-Limited Training for Robust Ad-Hoc Collaboration
SummaryThis paper proposes ConventionPlay, a reinforcement learning approach for ad-hoc collaboration in which agents learn to identify their partner's best convention. Agents are trained against a learned population of partners with varying adaptability, some following one fixed convention and others adapting to only a subset of conventions. Experiments show that ConventionPlay outperforms existing ad-hoc collaboration methods against test partners that can adapt to multiple conventions.
In briefGuarantees for accurately recovering signals from one-bit measurements using posterior sampling
SummaryThe paper studies how many one-bit noisy measurements are needed to recover signals drawn from a prior distribution. It measures the prior's complexity using its approximate covering number and shows posterior sampling succeeds with high probability when the measurement count scales with the logarithm of that number. The guarantee remains reliable if the learned prior is close to the true distribution in Wasserstein distance, and a matching-type lower bound is given. Experiments with diffusion priors on FFHQ and ImageNet are reported.
WhyCharacterizes prior complexity via approximate covering number with matching-order bounds, validated with diffusion priors on FFHQ and ImageNet; impact mainly within compressed sensing.
In briefProbability-Signature Dynamics: How Two-Layer Networks Learn Modular Addition
SummaryNeural networks trained on modular addition often form Fourier-structured representations that generalize exactly, but how gradient training selects these structures remained unclear. The paper introduces probability signatures, which describe leading gradient interactions via conditional statistics of the training distribution. For modular addition these signatures are cyclic shift operators diagonalized by the discrete Fourier transform, explaining Fourier sparsity, frequency matching, and phase alignment. The framework also accounts for corrupted examples under label noise decreasing loss faster early on, and the authors report observing the predicted frequency on XOR.
In briefFrom Suppression to Repair: Reducing Object Hallucination in Vision-Language Models by Aligning Distributions Locally
SummaryThis paper addresses object hallucination, where large vision-language models generate content not grounded in the image. It proposes ResOT, a training-free method that repairs hidden representations at inference time by projecting hallucinated directions out of a faithful subspace and aligning the hallucinated distribution to the faithful one using Gaussian optimal transport. Experiments on three representative models report substantially reduced object hallucination along with improved caption quality and multimodal benchmark performance.
In briefDetecting speech impairment during brain surgery using self-supervised speech features
SummaryThis paper studies how to automatically tell dysarthric speech from normal speech in recordings made during awake craniotomy, using the DATABRASE corpus. The pipeline isolates patient speech with speaker diarization, combines handcrafted acoustic features with multilayer wav2vec 2.0 embeddings, and uses a cascaded classifier. Under leave-one-speaker-out evaluation, the speech representation mattered more than classifier choice.
Kanthila Chinmayi (IRDL, LaTIM), Abdallah Nassib (LARIS) et al.
In briefLightweight vision-language models that turn scanned documents into structured JSON
SummaryThis paper compares eight open-source vision-language models with up to 7B parameters on converting document images into text and schema-compliant JSON, using three university heritage collections. Models are evaluated in zero-shot, few-shot, and fine-tuning settings with CER, ANLS*, and mAP-F1, and the effects of hyperparameter tuning, image preprocessing, and multi-stage training are tested. The authors conclude that carefully adapted models of this size can offer a private, institution-controlled alternative to manual transcription or commercial APIs.
Uddipan Basu Bir, Vincent Christlein, Andreas Maier et al.
In briefFast Pose Tracking of Rigid Objects with Compact Pose Graph Optimization
SummaryThis paper addresses long-horizon 6D pose tracking of novel objects, which current methods handle only with costly onboarding or by maintaining a reconstruction throughout the sequence. The authors propose a lightweight tracking module that can sit on top of many correspondence estimation methods. It optimizes only relative pose constraints in a pose graph, weighted by uncertainty derived from geometric alignment, instead of point-based optimization. On four real-world benchmarks, it reportedly matches reconstruction-based trackers in accuracy at a fraction of the optimization cost.
In briefSeek-and-View Reasoning: Finding the Right Viewpoint to Understand Space
SummaryExisting multi-view spatial reasoning relies mainly on sparse input views, which makes cross-view alignment fragile for vision-language models. The paper proposes Seek-and-View reasoning, which locates a question-relevant view to supply spatial evidence, and introduces Vantage, a training-free, model-agnostic framework pairing a VLM with a 3D foundation model. Experiments on six VLMs across five benchmarks report consistent improvements without fine-tuning.
In briefPRAXIS: Learning dynamics of self-improving models with symbolic archives
SummaryThis paper introduces PRAXIS, a framework that treats generators, learners, and symbolic archives as interacting dynamical processes in self-improving learning systems. It proves that KL-constrained generator updates and controlled archive-weight changes bound objective drift, and gives convergence-type guarantees for stochastic gradient descent. Experiments on visual robustness, relational graph reasoning, and algorithmic graph reasoning show generator stabilization, decreasing learner loss, and archive concentration consistent with the theory.
In briefSoftmax attention on Gaussian mixtures: linear when possible, selective when needed
SummaryThis paper studies the infinite-prompt limit of softmax attention, the core mechanism of Transformers, on Gaussian mixture data. Unlike prior theory using only Gaussian data, Gaussian mixtures add latent structure, multimodality, and nonlinear dependencies. The authors show that softmax attention can represent and learn, via gradient-based methods, optimal solutions to tasks such as supervised classification and denoising. It matches linear attention on linear tasks and uses query-dependent context selection to solve nonlinear tasks linear attention cannot.
Simon Gabet (LMO), Etienne Boursier (LMO, CELESTE) et al.
In briefMemento 3: Model-Based Recursive Self-Improvement via Reflective Rulebooks
SummaryThis paper presents Memento 3, which enables frozen LLM agents to continually learn explicit world models through external memory. The agent keeps a natural-language rulebook of revisable hypotheses about environment dynamics, compiles it into executable code for prediction and planning, and revises both using prediction errors, accepting updates only when they are faithful to the rulebook and reproduce observed transitions. On ARC-AGI-3, the single-model agent clears all levels of all 25 public games with a mean RHAE of 100.0 using 44% of human actions. In an Atari Pong case study, a learned feedback controller wins 21:0 in each of three evaluated episodes without further LLM calls.and
In briefAuraLuxMuse: Adaptive Fusion Modeling for Aesthetic Stage Lighting Design Using Music and Expert Guidance
SummaryThis paper presents AuraLuxMuse, a system that assists stage lighting design by mapping music and professional lighting cue sequences into a shared retrieval space. It retrieves and adapts editable lighting cues according to designers' preferences, rather than replacing the designer with an unconstrained generator. The authors also introduce Musilux, a dataset of paired musical audio and lighting cues. The system is evaluated in virtual simulation and professional laboratory settings using objective and subjective measures.
In briefPhase-aware video generation for realistic solid-gas physics
SummaryThis paper presents PAVG, a video generator that produces videos of solids and gases and how they interact. It uses a dual-branch design to model each phase's distinct motion, with spatiotemporal cross-attention capturing their physical coupling. The authors also built a simulation corpus of over 700K physical trajectories covering solid, gas, and solid-gas interaction scenarios. The paper reports that PAVG improves motion adherence, physical plausibility, and visual quality over existing approaches.
