In briefSWE-Journey: A More Realistic Benchmark for Coding Assistants Across Long, Multi-Turn Tasks
SummaryThis paper introduces SWE-Journey, a benchmark for evaluating coding assistants like Claude Code and Codex in conditions closer to real use. It automatically builds long-horizon coding tasks with a weak-to-strong synthesis pipeline and uses a user-simulation agent, based on four personas mined from real interaction data, to reproduce multi-turn interactions. The authors report that models pass over 75% of functionality tests with software architects but fewer than 25% with non-coders.
In briefFederated linear approximation that speeds up when agents are similar and stays personalized
SummaryThis paper studies personalized federated linear stochastic approximation (LSA), a framework that includes personalized temporal difference learning, where heterogeneous agents solve distinct linear fixed-point equations. The authors introduce PF-LSA, which mixes each agent's local stochastic update with the average update across agents, at no extra computational cost compared with standard federated methods. They prove it achieves best-of-both-worlds guarantees without prior knowledge of the heterogeneity level. The error is split into consensus and disagreement components: consensus error decays rapidly, while disagreement error decays more slowly but becomes negligible in low-heterogeneity regimes.
In briefA safe, persistent, self-evolving agent harness for partially observable environments
SummaryThe paper introduces E-Ledger, a multi-agent harness that checks each proposed action against policy before execution and maintains a ledger of verified hidden rules and evidence-backed state. Because hidden rules are initially unknown, the authors also propose WorldAbduct, which diagnoses execution trajectories from four views and verifies hypothesized rules through abductive interactions. On the World of Workflows enterprise benchmark, E-Ledger with WorldAbduct improves safe task completion across four LLM backbones, and experiments in ScienceWorld and DiscoveryWorld suggest the approach transfers to scientific environments.
In briefNew bounds on regret for online sparse linear regression
SummaryThis paper studies online sparse linear regression, where each instance allows access to only b of d attributes for prediction and b_0 extra attributes after prediction, a problem previously shown to be NP-hard. Earlier work designed efficient algorithms under regularity assumptions without characterizing information-theoretic complexity. The authors give the first lower bound on the minimax regret and design algorithms with better upper bounds without regularity assumptions. They also characterize how minimax regret scales with problem-dependent parameters.
In briefC4-equivariant flow matching for generating polycrystalline microstructures on anisotropic power-diagram graphs
SummaryThis paper presents a generative model that uses flow matching and graph neural networks to synthesise realistic polycrystalline microstructures. Microstructures are represented as anisotropic power diagrams, giving a compact geometric parametrisation that can be rendered at any pixel resolution. A C4-equivariant architecture builds rotational symmetry into the model, and training-free guidance is shown to generate structures such as copper welds, cast metal slabs, 3D-printed stainless steel and heterogeneous titanium lamellae.
In briefGlucoPRISM: Pretraining that separates sensor effects from glucose patterns to transfer models across patient cohorts.
SummaryThis paper proposes GlucoPRISM, a pretraining method for continuous glucose monitoring foundation models that assigns person, daily-state and sensor information to separate fixed-size representation bands. Dropping the sensor band at deployment, without target data, raises mean cross-cohort transfer AUROC from 68.75 to 73.57 across twelve directions, compared with 69.3 for GlucoFM. The authors also report gains on public task-cohort cells and on clinical endpoints in an external cohort of 9,303 participants.
WhyNovel response-contract pretraining isolates sensor effects; twelve cross-cohort directions and statistical tests support it, but impact stays within CGM foundation models.
Adinath Madhavrao Dukre, Abdullah Mohammad, Yifan Lu et al.
In briefConditional transfer from controlled pretraining mixtures to code
SummaryThis paper studies how synthetic tasks are used in language-model pretraining, separating tasks that are diagnostic, teachable, or transferable to downstream targets. In controlled experiments, 70% of the corpus is fixed general Python and 30% is split among three source families. Curated synthetic data shows value only as a limited share of a mixture still dominated by a target-aligned source. A loss-based adaptive scheduler drove the mixture away from the region that transfers and underperformed fixed-mixture controls.
In briefOX-NeRF: 3D X-ray reconstruction from very few views using implicit neural representations
SummaryThis paper presents OX-NeRF, a framework for reconstructing 3D structures from X-ray scans taken from ten or fewer viewpoints, and for reconstructing sets of related scenes together. It combines a convolutional neural network that learns features shared across scenes with scene-specific multi-resolution hash grids and a multilayer perceptron, all optimised end-to-end. The authors report that benchmarks on parallel-beam and cone-beam X-ray datasets show significantly higher reconstruction accuracy on ultra-sparse scenes than existing radiance field methods.
Thomas Welsch, Min-Hsin Tu, David J. Chapman et al.
In briefLearning to Coordinate Evolutionary Search: A Reinforcement Learning Controller for Hybrid DE and CMA-ES in Structural Model Updating
SummaryThis paper proposes DRL-DCO, an algorithm in which a DDPG-based deep reinforcement learning agent dynamically governs a hybrid of Differential Evolution and CMA-ES for high-dimensional optimization and structural model updating. The agent reallocates resources between the two methods and regulates population size, elite preservation, and restarts. A trained actor can run in a supervision-free inference mode. The authors report better convergence accuracy and robustness than several adaptive and hybrid evolutionary algorithms on high-dimensional benchmarks and the IASC-ASCE structural health monitoring benchmark.
In briefProsody-to-Text: predicting words from low-pass filtered speech
SummaryThe paper examines how much of a sentence can be recovered from its prosodic pattern alone, the reverse of the usual text-to-prosody task. The authors fine-tune Whisper using only the 12 lowest Mel bins (about 450Hz cutoff) and reach a 36% word error rate. About 10% of utterances are recovered perfectly and 40% have a word error rate of 25% or less.
In briefLearning the Loop, Not Just the Page: Execution-Grounded Loop Learning for Web Generation
SummaryThis paper introduces WebLoop, a framework that jointly trains a single policy to generate, critique, and refine functional web pages. The Critic is trained without execution, using signals for requirement discriminability and downstream helpfulness. With Qwen3.5-9B, WebLoop improves on the base model by 11.3 points on WebRise and 15.4 points on WebGen-Bench. Controlled analyses indicate the gains are not explained by an extra refinement pass alone.<br>
In briefRule-Based Control for Multi-Agent LLM Negotiation Against an Adversarial Gatekeeper
SummaryThis paper studies how multiple LLM agents negotiating with a stateful counterpart can avoid wasted model calls and recover from deadlock. The authors propose a control stack combining a runtime constitution, a four-tier agent swarm, and deterministic deadlock detection with context purging and a canonical recovery message. In 30 runs against a test gatekeeper with a known solution, the stack sometimes unlocked negotiation and, in one deadlock scenario, recovered in all runs versus none for LLM-only steering, without extra model cost.
Masaaki Nakatsu (AO, Inc. / OrbLabs AG), Reno Wang (AO et al.
In briefDesign Creativity Bench: Measuring Creativity in LLM-Generated UI Designs
SummaryThis paper introduces Design Creativity Bench, a benchmark for assessing the diversity and appropriateness of UI designs produced by large language models. It measures originality, creative range, and the share of acceptance criteria each design meets. The authors report that LLM designs are much less original (0.592) and have a narrower creative range (0.581) than human designs (0.764 and 0.902), while appropriateness is above 90% for every model. They conclude that LLM outputs are generally appropriate but substantially more repetitive than human designs.
In briefLAIR-Net: A Residual Neural Network with a Learned Anchor for Tabular Regression
SummaryThe paper proposes LAIR-Net (Leaky Alignment-Impulse Residual Network), which mixes a shallow learned anchor into each hidden state of a randomized deep network through a leaky residual transition, rather than only stacking random transformations. The authors derive a depth-uniform bound on input-perturbation sensitivity and use controlled simulations to attribute gains over a randomized baseline to the anchor. Across 23 benchmark datasets, LAIR-Net achieves the best average rank among eight randomized networks and twelve conventional models.
In briefHAND: A Bio-Inspired Activation Function for Better Image Classification Efficiency
SummaryThe paper introduces HAND, a new activation function combining homeostasis, accelerating nonlinearity, and divisive normalisation, to add a biologically inspired inductive bias to CNNs. With HAND, a ConvNeXt-tiny reached on ImageNet1k the accuracy the unmodified model achieved after 200 epochs in just 25 epochs. Gains were even larger on long-tailed ImageNet with reduced, unevenly distributed training data, while performance on common-corruption data and rejection of unknown-class samples was unaffected or improved.
In briefReTeach: A self-distillation method where the model reflects and retries to teach itself
SummaryReTeach is a self-distillation framework that builds a self-teacher through multi-round reflection and retry, using only the model's own attempts and outcome-level verification. Without reference answers, external feedback, or cross-example memory, the teacher revises failed attempts until success or a retry budget runs out. The student learns the teacher's token-level predictions on its own rollouts, keeping single-pass inference. Across six benchmarks covering math reasoning, science QA, and tool use, ReTeach improves average accuracy over GRPO by 1.39 percentage points.
In briefAtomWorld-Mirror: A macro-step world model for long-term materials evolution
SummaryThis paper proposes AtomWorld-Mirror, a time-aware macro-step world model that targets the critical evolution backbone of atomic systems. It condenses short micro-event segments into physically reachable transitions between key states, jointly predicting sparse structural edits and accumulated physical time. Across five atomic systems, macro-step inference is reported to be 10^3 to 10^4 times faster than event-by-event simulation.
In briefMSGAT: Energy-efficient spiking network for image-text retrieval using graph attention and similarity fusion
SummaryThis paper addresses the difficulty spiking neural networks face in capturing semantic structure for image-text retrieval. It proposes MSGAT, a multi-head spiking graph attention network for structural modeling, and Sim-Fuse, a similarity-space fusion strategy that combines coarse and fine-grained matching without directly fusing heterogeneous representations. On Flickr30K and MSCOCO, the method outperforms matched ANN methods and existing SNN baselines, and with only two time steps it matches or exceeds its ANN counterpart while cutting theoretical module-level energy by 55%.
In briefWhen to Intervene? State-Aware Sparse Manipulation in Federated RL
SummaryThis paper studies Byzantine poisoning attacks on federated reinforcement learning, arguing that the timing of interventions, meaning which trajectory states are manipulated, materially affects attack efficacy. The authors propose V-BSA, which uses local policy uncertainty to select sparse intervention states and applies envelope-constrained behavioral steering. On discrete-action benchmarks, V-BSA degrades performance against robust aggregators and ensemble defenses using only a fraction of the interventions required by dense poisoning.
In briefWhen Can You Prune Your Network? Removing Intermediate Neurons in Multilingual Speech Parsing
SummaryThis paper studies whether intermediate neural network units are needed for speech parsing, which predicts both a transcription and a syntactic tree for spoken utterances. The authors propose a simpler end-to-end architecture without these units, reducing parameters by 12% while achieving comparable or better results on speech recognition and parsing. They evaluate on French and the lower-resource languages Slovenian and Naija, and examine the effects of training data size and intermediate layers of the pretrained speech encoder.
In briefResidual Advantage: Teacher Guidance Relative to the Student for RL with Verifiable Rewards
SummaryThis paper proposes Residual Advantage (RA) for post-training reasoning models. It treats the teacher-student probability residual as a bounded one-step reward, subtracts the student-policy state value, and centers the result within each response before adding it to the verifier advantage, so the teacher only redistributes credit among steps. CoRA further updates a teacher LoRA on the same scored student batch to adapt guidance to the student's attempts. On three math benchmarks, RA combined with GRPO or REINFORCE++ improves the base algorithm in all 24 comparisons.
In briefDoes Modern Standard Arabic dominate dialects inside LLMs? A layer-level check
SummaryLarge language models often generate Modern Standard Arabic even when prompted in dialects, and a common explanation is that MSA dominates their internal representations. The authors adapt a language-dominance framework to 26 Arabic varieties and find no evidence of MSA dominance across layers and model families. Dialect representations instead form a dense, highly overlapping space, and separability effects can shift toward later layers. The authors conclude that output bias does not necessarily reveal internal representational dominance.
Abdu Sallouh, Nicholas Popovi\v{c}, Michael F\"{a}rber
In briefWARP-VLA: Adapting Wrist Cameras for Robust Robot Control Across Viewpoints
SummaryVision-language-action models for robot manipulation often perform poorly when the camera setup changes, and wrist cameras are especially sensitive because they move with the arm. WARP-VLA uses a mixture-of-experts design, where experts learn view-specific feature transformations and a router combines them from implicit view information, without needing camera extrinsic parameters. On the LIBERO benchmark, it raises the average success rate of pi-0.5 under wrist-view perturbations from 39.2% to 78.3%. Real-robot experiments show the learned adaptation transfers to diverse deployment settings.
WhyUses MoE with view-specific experts and a router, avoiding extrinsics; LIBERO success rises from 39.2% to 78.3%, plus real-robot transfer.
In briefCompactness and Consistency: A Joint Framework for Deep Graph Clustering
SummaryThis paper proposes CoCo, a deep graph clustering framework that groups nodes with similar characteristics into clusters. CoCo uses graph convolutional filters to learn node representations from both local and global views, then encodes them into low-rank compact embeddings to remove redundancy and noise. It also adds a consistency learning strategy based on these embeddings to transfer knowledge between the two views. The authors report that CoCo outperforms state-of-the-art methods on various datasets.
In briefThe Operator Mismatch Problem: Running Bird's-Eye-View Perception on Portable GPU Compute
SummaryThis paper examines how to deploy bird's-eye-view (BEV) perception models that fuse camera and LiDAR data for 3D object detection in autonomous driving. The authors argue that standard inference runtimes cannot represent sparse 3D convolutions or geometric scatter operations, and that existing sparse convolution libraries are CUDA-only and PyTorch-coupled. They present BEVPIPE, which uses portable GPU compute APIs integrated with production runtimes, reporting a 19.5x end-to-end speedup while retaining 98.5% of reference mAP. The authors also show that the framework works across different GPU backends.
In briefFed-GRPO: Federated Group Relative Policy Optimization Driven by Reward Signals
SummaryThis paper proposes Fed-GRPO, a federated training framework for Group Relative Policy Optimization that lets language models learn reasoning collaboratively without sharing raw data. It uses reward statistics produced during GRPO training to guide weighted aggregation, local reward calibration, and adaptive sparse communication. On mathematical reasoning tasks, it outperforms other federated methods and FedAvg, approaches centralized training performance, and reduces communication by 32 times losslessly, up to 621 times under tight bandwidth.
In briefBeyond Sequences: Distilling Structured Decision Memory for LLM Recommendation
SummaryMost LLM-based recommenders model only one behavior type or flatten several behaviors into one token sequence, which obscures their different decision roles and hurts performance on hard choices between similar items. The paper proposes MARI, which stores users' past rationales as Structured Decision Memories (goals, constraints, trade-offs) distilled offline and retrieved to augment LLM reasoning. Experiments report gains over state-of-the-art baselines on next-item prediction and a new Difficult Choice Prediction task, with low latency overhead.
In briefLearning qBIC resonances across metasurface families in a shared Fourier space
SummaryBound-state-in-the-continuum metasurfaces are usually described with geometry-specific parameters, which makes cross-geometry comparison difficult, and their ultranarrow qBIC features are easily lost in full-spectrum learning. The paper maps 2015 samples from seven dielectric metasurface families onto a shared reciprocal-lattice grid and uses two low-order Fourier channels to capture resonance shifts. A five-channel K-space backbone models the broadband spectrum, while a local complex K-space expert with a differentiable Fano layer models the qBIC resonance, reducing resonance-position MAE from 3.2 nm to 0.95 nm on a geometry-blocked test set.
In briefParametric Trajectory Distillation for Faster Video Generation
SummaryVideo diffusion and flow models need many sequential steps, which makes generation expensive. The paper proposes Parametric Trajectory Distillation (PTD), in which a student model fits teacher trajectory segments with polynomials and learns from teacher guidance along its own predicted path. A curvature head is used only during training, so inference keeps the original architecture. The authors report improved dynamic quality and naturalness over PDD on Wan2.1-14B with four steps.
In briefBeyond Resolution: Object-to-Image Ratio Mismatch in Instance Retrieval
SummaryThis paper examines why visual instance retrieval often fails when the same object appears at different sizes in query and gallery images. The authors argue the main cause is usually object-to-image ratio mismatch rather than resolution loss, testing 12 pretrained backbones on a controlled benchmark of 3,021 Objaverse objects. They propose query-side scale augmentation and an OWLv2 crop reranker, reporting improved results on ILIAS 100M.
In briefPSI-SINDy: Post-Selection Inference for Sparse Identification of Nonlinear Dynamics
SummarySINDy is a data-driven method that discovers governing dynamics from time-series data by selecting a sparse set of terms from a predefined library. This paper proposes PSI-SINDy, which provides hypothesis tests and confidence intervals to assess how reliable the selected dynamical terms are. Because the same noisy trajectory is used for both selection and significance testing, selection bias can arise, and the method uses data thinning to split the trajectory into four independent views. The authors establish theoretical validity under stated conditions and evaluate the method on simulated and experimental dynamical-system data.
Ashraful Islam, Shuichi Nishino, Tomohiro Shiraishi et al.
In briefBridgeGuard: A Method That Keeps Diffusion-Based Self-Driving Plans Within Safe Limits
SummaryDiffusion-based driving planners can produce varied behaviors but may generate unsafe trajectories when conditions shift. BridgeGuard gradually strengthens a safety constraint during denoising to steer intermediate trajectories toward a scene-dependent safe region, correcting them in a low-dimensional curve space. On Bench2Drive, it raises BridgeDrive's driving score and success rate from 87.99/74.99% to 90.88/76.36%, and DiffusionDrive^geo's from 80.79/58.18% to 90.46/74.09%.
In briefTesting Local AI Agents for Reproducible Data Engineering on Mobility Tasks
SummaryThis paper evaluates whether locally deployable open-weight large language model agents can produce correct, reproducible data-engineering artifacts. The authors built a benchmark of fifteen mobility-workflow tasks and scored generated scripts, tables, figures, and reports with deterministic checkers. A closed-loop workspace raised pass rates by 26.7 to 52.0 percentage points for models above two billion parameters, and the best configuration reached 85.3% artifact-level success.
Jorge Garc\'ia-Carrasco, Javier Sanchis, Alejandro Reina-Reina et al.
In briefRoboAware: Learning to Coordinate Robot Skills Using Counterfactual Outcomes
SummaryThis paper presents RoboAware, which trains only a state-conditioned coordinator to choose which policy family a coding agent should use for robot skills. It introduces the P5 schema, State-Locked Counterfactual Branching to observe outcomes of all admissible families, and Execution-Aware Learning combining Monte Carlo tree search with Q-learning. Across 100 tasks, RoboAware reports a 77.0% overall success rate.
In briefConditional Residual Prediction: Making Step-by-Step Video Generation Better Without a Bidirectional Teacher
SummaryThis paper studies causal, autoregressive video diffusion models, which generate video chunk by chunk and are weaker than bidirectional models of the same size under standard training. It proposes Conditional Residual Prediction (CRP), where the model first predicts the target without the history and the history only adds a residual. In controlled experiments, CRP nearly closes the 6.14-point gap to a bidirectional model. The authors use it to train Optica, a 2B-parameter model that generates 5-second 480p videos and scores 82.78 on VBench.
In briefRare gate disagreements can limit plasticity: when gradient flow mispredicts finite-batch SGD
SummaryThe paper shows that population gradient flow can qualitatively mispredict finite-batch SGD when adapting a network after pretraining. In a two-unit ReLU regression, gradient flow recovers on the target task, while online SGD fails with high probability after long source pretraining. The failure is traced to inputs where the two ReLU gates disagree, whose probability shrinks under weight decay.
In briefLooped Transformers for Parallel Latent Reasoning (LLoCoT)
SummaryThis paper introduces LLoCoT, which replaces step-by-step chain-of-thought generation with a looped transformer that iteratively refines a compact set of latent slots. Latent tokens are sampled in parallel and used to condition answer decoding. On HumanEval and MBPP, LLoCoT achieves the highest mean score among the compared methods and matches the explicit-CoT baseline in accuracy. It also reduces time to first answer token and reasoning-phase latency substantially, while raising end-to-end throughput by 9.2%.
Gerard Grau Garc\'ia, Arnau Padr\'es Masdemont, Niccol\`o Grillo et al.
In briefSAGE: Sink-Aware Guided Emphasis for Visual Grounding in Vision-Language Decoders
SummaryThis paper studies attention sinks in vision-language model decoders and finds that early and late layers collapse attention onto the same few image regions regardless of prompt (called PIS), while middle layers respond to the prompt and drive vision-language alignment. Based on this, the authors propose SAGE, a lightweight method that uses ROI masks from backbones such as CLIP, ViT, and DINOv3 to steer decoder attention away from PIS toward query-relevant regions. Experiments across several vision-encoder plus decoder-only LLM families show better visual grounding, fewer hallucinations, and consistent gains on public benchmarks.
In briefWho Verifies the Verifier? Co-Evolving Inspectable Graders with Self-Improving AI Agents
SummaryThe paper studies how to build inspectable verifiers (graders) for self-improving AI agents on open-ended tasks, where no ground-truth check exists. The authors evolve the verifier itself as an expression of small, mostly deterministic drawback detectors synthesized from clustered failures, selected by agreement with an anchored reference set and consensus on unlabeled outputs. On MBPP+ it improves held-out agreement by +0.21 over a hand-authored seed. Removing the anchor lets the verifier collapse into an always-pass grader that still trains skills equally well, so downstream task score alone cannot certify an evolved verifier.
WhyEvolving inspectable verifiers is novel, and showing collapsed anchor-free verifiers still train skills is notable; experiments are limited in scope.
In briefBeyond Speech Captions: Speech-Rewarded Style Planning for Conversational TTS
SummaryThis paper proposes Speech-Rewarded Style Planning (SRSP), which trains a text-based style planner to generate speaking-style instructions for a frozen downstream TTS model. The planner is optimized with GRPO, using the likelihood of target speech tokens as reward, since the authors find text-speech alignment weakly predicts acoustic similarity. On an English subset of the ISCSLP 2026 CoT-TTS corpus, SRSP improves style and emotion similarity and lowers mel-cepstral distortion over baselines, with gains in LLM-based contextual appropriateness and reference consistency.
In briefProtoSemImage: Image-Based Prototypes with Deformable Alignment for Explainable Document Classification
SummaryThis paper replaces vector class prototypes with images, rendering each document as a multi-channel image and representing each class by visual archetype images whose channels carry named linguistic factors. Classification is performed by deformable row alignment between a document image and the archetype bank, which indicates where and along which channel an input departs from its archetype. The image representation outperforms otherwise identical vector-prototype models in all three paired seeds, while the distance-based matching is identified as the main source of a reported shortfall. The authors also report these failure modes alongside the gains.
In briefFeature Space Adaptation for Efficient Gaussian Process Flows
SummaryConditional sampling from Gaussian processes is hard outside the linear-Gaussian setting. The existing FlowGP method handles non-linear, non-Gaussian conditions but is costly and needs hand-specified kernel hyperparameters. This paper adds kernel approximations to scale to high-resolution domains and estimates the marginal likelihood by measuring the work needed to steer diffusion toward conditioning statements, enabling hyperparameter optimisation in FlowGP for the first time. The approach is demonstrated on probabilistic downscaling from areal summary statistics, PDE solution inference on irregular domains, and recovery of sea level anomaly fields from non-Gaussian satellite observations.
Thomas Cowperthwaite, Louis Sharrock, Lachlan Astfalck et al.
In briefGenerative Adversarial Loops: AI That Finds Its Own Weaknesses and Improves Its Algorithms
SummaryThe paper proposes Generative Adversarial Loop (GAL), a framework where a discriminator agent generates adversarial data that exposes weaknesses in current methods, while a generator agent discovers algorithms to overcome them. It is applied to approximation algorithms for efficient inference across four tasks: KV compression, sparse video generation, sparse attention, and context extension. The authors report that the newly discovered algorithms improve on the adversarial data and also on established benchmarks.
Kislay Aditya Oj, Nidhi Jain, Sri Surya Varma Datla et al.
In briefSpikeSSL: A Universal Spike Inference Framework Using Dynamics-Informed State-Space Layers
SummaryTwo-photon calcium imaging records large neural populations, but accurately inferring spikes across different calcium indicators remains difficult. SpikeSSL uses a backbone of bidirectional IIR state-space layers motivated by calcium dynamics, with a conditioning encoder and a heteroscedastic variance head for per-frame uncertainty. It achieves state-of-the-art results on a benchmark of five fixed splits built from 33 public datasets, in both in-domain and leave-one-indicator-out zero-shot settings. The authors also built a biophysical simulator that generated about 11,000 paired traces, and training with these data improves cross-indicator generalization.
WhyDynamics-matched bidirectional IIR state-space backbone plus simulated data augmentation improves cross-indicator generalization; state-of-the-art across 33 public datasets and five splits.
In briefZatom-2: A multitask pretrained atomistic generative model for molecules, materials, and proteins
SummaryZatom-2 is an atomistic generative model pretrained on about five million structures from the OMol25 and OMat24 electronic structure datasets. It combines a multiscale Transformer with conditional flow matching and supports generation, structure prediction, and energy and force prediction. The authors report better molecular distribution fidelity than Zatom-1 and strong results on existing molecule and material generation benchmarks. In a low-data protein setting, joint generative-predictive pretraining and finetuning raised backbone designability in length extrapolation from 67.8% to 74.8%.
Miruna Cretu, Alex Abrudan, Antonia Panescu et al.
In briefTeleTwin: A calibrated digital twin of Telegram crisis discourse for testing detection of LLM-driven influence operations
SummaryThe paper builds an LLM-powered digital twin of a Telegram crisis-discourse ecosystem, where each channel is an agent instantiated from a real channel's history and calibrated on 1.07 million messages from 195 public channels about the Israel-Gaza and Iran-Israel conflicts (March-August 2025). The authors inject LLM-written campaigns with known membership into real data and the twin to test coordination detectors. Multilingual semantic similarity recovers loud campaigns, including cross-lingual ones, but not quiet ones, while all detectors fail inside the twin because its background is about 40 times more semantically homogeneous than real data.
WhyProposes an LLM digital twin of Telegram calibrated on 1.07M messages, injecting known campaigns to expose detector failures and semantic homogeneity.
In briefTracing a Coding Agent's Thoughts on ARC-AGI-3: Lessons for Continual Learning
SummaryThe paper studies how a coding agent built on a frozen foundation model learns across a sequence of abstract reasoning tasks, keeping state only in the scripts and notes it writes. The authors trace each belief, rule or plan through these files from where it forms to where it is corrected or abandoned. Experiments use the ARC-AGI-3 interactive reasoning games with seven evaluation runs and three backbones from two model families. Scripts written for one task are rarely reused in later tasks, as the agent rewrites knowledge and abandons most earlier scripts, while notes are never revised and contradictions are settled against the log.
In briefTesting how well AI coding assistants can build embedded device software step by step
SummaryThis paper introduces a benchmark for evaluating AI coding agents that write embedded firmware, using five control tasks and four feedback scenarios on simulated ESP32 hardware. Seven GPT-family and Qwen-family configurations were tested over 420 runs. gpt-5.4 achieved the highest pass rate but did not saturate the benchmark, qwen3.5-27B was the strongest local model, and smaller local models performed markedly worse.
Jorge Garc\'ia-Carrasco, Sergio Garc\'ia-Carrasco, Alejandro Mat\'e et al.
In briefLearning to Retrieve: built-in memory retrieval for video world models
SummaryThis paper proposes Learning-to-Retrieve (L2R), a method that lets video world models learn to retrieve past observations during generation instead of relying on an external memory system. The model reuses its own persistent internal state as memory, with a camera-conditioned gate deciding what to retrieve and a trigger deciding when, supervised by a 3D re-visibility signal. Across multiple base models and camera-revisit benchmarks, L2R improves long-term scene consistency without needing an external memory bank or 3D conditions.
In briefMotiveMob: Generating Human Movement Trajectories by Modeling Motivation Step by Step
SummaryThis paper introduces MotiveMob, an autoregressive framework that generates human mobility trajectories by first inferring a motivation for each movement, then predicting where and when it occurs. Candidate locations are checked for speed feasibility and repetition before feeding back into the next step. The authors evaluate it under shifts involving unseen users and time periods, including seasonal changes and COVID-19 disruptions. They report that it achieves better distributional fidelity than competitive pretraining-based and prompting-based methods.
In briefToken-based communication framework for cooperative drone sensing and communication
SummaryThis paper studies a cooperative integrated sensing and communication network of embodied UAV agents, where each drone has only local observations. It proposes SI-TokCom, which uses pretrained codebooks to compactly exchange local states and intentions, with agents jointly learning token selection and physical actions. Simulations report communication and sensing rates reaching 98.9% and 99.1% of a centralized baseline, improvements of 5.0% and 43.6% over a local baseline, with essentially unchanged propulsion energy.
In briefDLC: A Metric-Guided Dynamic Loss Controller for Balancing Multiple Training Objectives
SummaryThis paper proposes DLC, a training-time controller for multi-objective image restoration that dynamically reweights existing loss terms without changing the architecture or adding new losses. It periodically evaluates the model on a small fixed feedback subset and uses an LLM-based controller to update the loss weights from the resulting quality metrics. The method is tested on low-light enhancement, deraining, and real-world super-resolution using both reference-based and no-reference metrics.
In briefEchoDiST: Self-distillation and diffusion-conditioned learning for estimating heart muscle motion in echocardiography
SummaryThis paper presents EchoDiST, a framework that estimates myocardial motion from echocardiography without ground-truth displacement fields. It combines self-distillation-based joint learning with a diffusion-conditioned motion estimation network, jointly optimizing anatomical segmentation and motion estimation under limited annotations. Inference requires only a single deterministic forward pass. Evaluated on three echocardiographic datasets against seven learning-based methods, it reports improvements in anatomical alignment, strain assessment, and motion-derived functional measures.
In briefBioBigBird: a sparse-attention language model for handling long biomedical texts
SummaryThis paper presents BioBigBird, a bidirectional language model pre-trained on biomedical literature and clinical data. It uses a sparse attention mechanism to process sequences of up to 4096 tokens. The authors add multi-task learning for named entity recognition and relation extraction, and report highly competitive results against state-of-the-art models on the BLURB benchmark.
Roshan Balaji, Pavan Kumar S, Vasudev Gupta et al.
In briefWhy Machines Will Still Not Rule the World: A Rebuttal to Claims About General AI
SummaryThe authors restate their book's thesis that artificial general intelligence is mathematically impossible, because intelligent human behaviour arises from complex systems that current models cannot capture. They examine and reject two arguments from proponents: a theoretical one based on universal approximation theorems and the Church-Turing-Deutsch principle, and an empirical one based on rising scores on standardized benchmarks. The paper argues these benchmarks are compromised by training-data contamination, flawed test construction, and strategic optimization. It concludes that models needed for open-ended, thermodynamically complex, non-ergodic environments are not achievable.
In briefPolicy Alignment: New Signals for Membership Auditing in On-Policy Distillation
SummaryThis paper studies membership auditing in on-policy distillation (OPD), which asks whether a given prompt was used to train a student model. The authors propose Policy Alignment Membership Auditing (PAMA), which uses the teacher-guided direction of the student's update and combines Teacher Alignment Gain with student drift and uncertainty alignment signals. On the MATH benchmark, PAMA reaches AUC values of 0.791 to 0.941, improving over state-of-the-art baselines by 14.6% to 20.6%.
In briefCausal-fate dynamics: how unrealized influence persists in dynamical systems
SummaryThe paper proposes a framework called causal-fate dynamics, in which influence generated by a system may be realized, remain latent, or be transformed by later dynamics. It presents an exact finite-transport representation and examines three studies: a connectome-constrained C. elegans model, Internet routing observations, and a Transformer that carries latent contextual influence.
In briefCalibrating brain tumor segmentation confidence when MRI scans are missing
SummaryBrain tumor segmentation models often use several MRI scan types, but some scans may be missing in clinical practice. This paper introduces MMA-LTS, a post-hoc method that adjusts voxel-wise confidence by estimating a spatially adaptive temperature field from modality availability and voxel difficulty. Experiments on BraTS 2020 and FeTS 2024 show improved calibration across various missing-modality scenarios while preserving the segmentation accuracy of state-of-the-art models.
In briefRewiring Semantics, Dynamics, and Control: A Simple Action-Centric Tri-Stream Transformer for Robot Manipulation
SummaryThis paper addresses Vision-Language-Action (VLA) models for robotic manipulation, noting that pretrained vision-language backbones lack sufficient physical dynamics priors. The authors propose ACT³, a tri-stream Transformer in which a dedicated action expert reads representations from both a VLM and a video world model via layerwise attention, while each backbone attends only within its own stream. Both backbones can be updated through control supervision. The authors report that ACT³ outperforms comparable methods on simulated and real-world robotic manipulation benchmarks.
In briefCombining calibrated predictors into refined ones: an algorithmic study (paper)
SummaryThis paper studies how to combine several calibrated predictors into a more informative one, given only the input predictors and the feature distribution, not the underlying Bayes probabilities. Predictors are modeled as signaling schemes, and refinement is defined through feature-independent garblings. For deterministic outputs, refining two predictors is polynomial-time solvable, while refining an arbitrary number of inputs is NP-hard. Allowing randomized outputs yields a polynomial-time algorithm for any number of input predictors.
In briefFresco++: Frequency-guided, consistent optimization for detailed head avatars
SummaryThis paper proposes Fresco++, an optimization framework for reconstructing fine-grained head avatars that look consistent from different viewpoints. A progressive frequency curriculum first stabilizes low-frequency structure and then adds high-frequency constraints for facial and hair details. Canonical Group Consensus links observations through shared canonical surface regions, filters unreliable views, and aggregates multi-view evidence into a consensus target. Experiments on NeRSemble report better reconstruction quality and cross-view consistency.
In briefTracing Emotions from Causes to Consequences in Real Social Scenes
SummaryThis paper proposes TRACE, a framework that breaks an emotional episode into Condition, Affect, and Effect stages, incorporating cognitive factors such as internal stance and emotional display regulation. The authors build TRACE-Bench, with 3,746 question-answer pairs over 646 videos, to test multimodal models on five tasks. They report a substantial gap between models and humans, and introduce TRACER, a structured reasoning method that outperforms all evaluated baselines on each task.
In briefMC-TRCM: Recursive Fusion Model for Incomplete Mobile and Wearable Mental-Health Data
SummaryPublic mobile and wearable mental-health datasets often provide summarized feature tables rather than raw sensor streams, and some data sources are missing. The authors propose MC-TRCM, which keeps each feature source as a separate token, treats missingness as part of the input, and refines predictions recursively. Evaluated on six endpoints from DepreST-CAT and PSYCHE-D, it achieved the lowest mean absolute error on PHQ-9 and GAD-7 severity, while classification results varied by task.
In briefEstimating Rooftop Solar From Net Smart Meters: Deployment Scenarios and Operational Findings
SummaryDistribution operators need to see solar generation behind home meters, but most households only report net consumption through smart meters. The paper uses deep learning to separate photovoltaic output from net readings, testing performance across four deployment scenarios: known households, new households, future periods, and new regions. Experiments use Ausgrid data with transfer to Pecan Street. The deployment scenario explains much more error variation than model choice, and fine-tuning on five target-region households reduces transfer error by 28% to 36%.
In briefBeyond Distributional Fidelity: Causal-Penalized Diffusion for Synthetic Tabular Data
SummaryThis paper studies how to make synthetic tabular data preserve causal effects, not just statistical similarity. It proposes a training framework that adds a causal discrepancy penalty to a generative model, implemented as a causal-penalized TabDDPM and optimized with an on-policy score-function estimator. Experiments on treatment-effect simulations and two benchmark datasets test whether causal fidelity improves while statistical fidelity stays competitive.
In briefUsing hetero-component density to improve cryo-EM protein structure reconstruction
SummaryThis paper addresses reconstructing protein structures from cryo-EM maps. The authors find that candidates near hetero components can either help or hurt chain construction, depending on filtering. They propose CryoCue, which uses hetero information to guide backbone localization and refine structures. Experiments report better backbone localization near hetero components and more accurate reconstructions.
In briefGroundSight's training-free verification method for multi-view driving-scene question answering
SummaryThe paper addresses GoldenViewVQA, a task where models must answer driving-scene questions and identify the camera view that supports the answer. The authors propose CoVeR-VQA, a training-free four-stage framework that refines zero-shot predictions from GPT-5.6 through view-specific, joint, and cross-split group verification. It reaches 84.75% Joint Accuracy on the official test set, and the final submission reaches 88.14%. The authors find that localizing the supporting view is the main source of remaining errors.
In briefWAM-Cache: Training-free KV caching for faster robot world-action models
SummaryWAM-Cache is a training-free method that speeds up World Action Models for robot manipulation by reusing layerwise key-value representations across control chunks. It recomputes only a sparse set of refresh tokens, chosen by combining the action expert's cross-attention with visual latent surprise and bounded by an age limit. On Fast-WAM, it reduces video DiT prefill FLOPs by 32-42% while staying within 0.7-1.8 points of the dense policy in simulation and 2.5 points on a real robot.
In briefImproving Precision in Estimating Expectations Under Language Models Using Potentials
SummaryThis paper addresses estimating the expected value of a test function under a language model, a task that can be computationally expensive. The authors reuse next-token conditional probabilities produced as a by-product of sampling, using potentials (functions of prefixes) to decompose the test function additively. They build a family of estimators parameterized by these potentials and derive conditions for variance reduction. Across entropy estimation, watermark detection, and rare-event probability estimation, the estimators achieve substantial variance reductions over Monte Carlo and importance sampling baselines at comparable computational cost.
Francesco I. Re, Shubhangi Ghosh, Tim Vieira et al.
In briefCoPoE: A disease-coordinate product-of-experts method for Alzheimer's diagnosis with missing data types
SummaryThis paper presents CoPoE, a framework that maps multimodal clinical, imaging, genomic, and biomarker evidence into a latent space with four axes: genetic Risk, molecular Pathology, Neurodegeneration, and clinical Stage. A masked product-of-experts architecture fuses only the modalities that are available, without synthesizing missing inputs. Experiments on the ADNI dataset report strong all-modality performance, the highest mean AUROC across 15 observed-subset evaluations among compared baselines, and improved ECE, Brier score, and NLL.
In briefTypedBench: A benchmark for calibration, wording sensitivity, and cost in typed decision models
SummaryThe paper introduces TypedBench, a benchmark for "System One" decision models that output calibrated probabilities over typed answers via a non-generative interface. It is built from seven policy-labelled generators and nine evaluation suites, and compares a hosted model, an open encoder, and open decoders from 0.8B to 9B parameters. The hosted model follows policy but is wording-sensitive and underconfident, and under asymmetric costs its probabilities can yield worse outcomes than its top answer.
Rahul Sharma, Andrew B. Duncan, Ga\'etan Marceau Caron et al.
In briefFact over Fiction: Detecting Hallucinations in Sinhala-to-English Machine Translation
SummaryNeural machine translation can produce fluent output that is semantically unrelated to the source, a problem especially acute for the low-resource Sinhala-to-English pair. The paper proposes a reference-free detector trained on a 45,000-sample synthetic dataset built with five corruption strategies, fine-tuning mDeBERTa-v3 to reach a token-level F1 of 0.841. Ablations show the detector depends on the Sinhala source, and benchmarking eight translation systems reveals detector firings varying by an order of magnitude across architectures.
In briefRISR: Using Residual Errors to Help Language Models Discover Scientific Formulas
SummaryRISR is a method for symbolic regression that uses patterns in the remaining prediction errors (residuals) across inputs to guide formula discovery and decide which corrections to fit. A residual encoder turns inputs, targets, predictions, and residuals into tokens that condition a language model to propose formulas, and a dual-view relational encoder predicts which candidate corrections will be useful after fitting. On LLM-SRBench scientific tasks, RISR reports 63.57% and 38.50% ID accuracy at 1% and 0.1% relative-error tolerances, and 56.07% and 38.24% OOD accuracy. It outperforms the reported baselines using the same backbone.
In briefEnvironmental feedback modeling matters for agentic self-distillation
SummaryThe paper argues that simply conditioning a teacher model on environmental feedback is insufficient for agentic self-distillation. It proposes SELF, a framework that jointly learns to predict environmental responses and distills feedback-conditioned guidance into the policy. With Qwen3-8B, SELF outperforms SDPO and GRPO on τ-bench success rate and AppWorld task goal completion.
In briefPlanWAM: Planning-Shaped Future Representations for End-to-End Driving
SummaryThis paper proposes PlanWAM, a world-action model in which the planning task shapes the future-state representation so it keeps the information most useful for planning. It uses a Temporal Register Pyramid to compress historical frames, a privileged future posterior branch trained with trajectory-planning objectives, and Hindsight-to-Foresight Distillation to predict this future latent from history alone. The authors report 93.8 PDMS and 90.9 EPDMS on NAVSIM-v1/v2 navtest, and 38.7 HD-Score on closed-loop HUGSIM in a zero-shot setting.
In briefEvoKnow: A Method for Continually Updating AI-Generated Image Detection
SummaryThis paper presents EvoKnow, a framework for keeping AI-generated image detectors up to date as new generative models appear, without replaying past generated images. It preserves a shared forensic basis learned from base domains and adds isolated residual experts over time, retrieving them through an Analytical Incremental Router (AIR). Experiments report cross-generator generalization, few-shot expansion, and long-horizon continual adaptation, with reported accuracy and forgetting figures under a replay-free continual learning protocol.
In briefEqual Path Cost, Unequal Output Effects: How Perturbations Spread in Diffusion Models
SummaryThis paper studies how perturbations introduced along the sampling trajectory of diffusion models affect generated outputs. It measures perturbation strength with a KL divergence between perturbed and reference trajectories, called path cost, and derives a response identity explaining why output changes may go undetected by first-order responses. Experiments on pretrained diffusion models and an application to cache-based acceleration show the framework can identify sampling intervals where caching causes larger image errors.
In briefFastJEV: Reducing Redundant Computation in JEV Inference
SummaryThis paper studies redundant computation when JEV multimodal models score candidate answers, and proposes FastJEV to make candidate evaluation more compact. It combines shared context anchoring, candidate prefix sharing, and decision guided pruning based on relative score changes on a small unlabeled set, without additional training. Across three OmniJev model sizes on five public benchmarks and reconstructed LIBERO-10 offline questions, the method cuts candidate depth by 43.75% to 45.83% while keeping 93.66% to 97.52% of original task scores on average.
In briefCRISP: A Diffusion Decoder That Reduces Flying Pixels in Latent LiDAR Generation
SummaryThe paper identifies flying pixels in latent LiDAR pipelines as an artifact caused by convolutional VAE decoders blurring sharp depth discontinuities. It introduces CRISP, a pixel-space diffusion decoder that replaces the decoder while keeping the encoder and latent generator fixed. Across KITTI-360, SemanticKITTI, and nuScenes, the authors report an average 50.5% reduction in FSVD/FPVD, and a 15.5% FSVD improvement in a pretrained LiDM world model.
WhyReplaces VAE decoders with a diffusion decoder to fix flying pixels, with consistent large gains across datasets and backbones; a substantial improvement over prior work.
Andrea Ceron, Michael Schmidt, Alvaro Marcos-Ramiro et al.
In briefPersonalization Matters: A Long-Horizon Shopping Chat Agent Using User-Centric Information
SummaryThis paper proposes a multi-agent, multimodal retrieval-augmented generation framework for personalized conversational shopping, splitting dialogue tracking, recommendation retrieval, preference-aware reasoning, and response generation across agents. It combines product metadata, reviews, image-derived descriptions, and users' past reviews. On an Amazon Reviews 2023 benchmark, retrieval-enabled variants scored higher than a no-RAG baseline on trajectory metrics (4.82 vs. 3.74). A small study with five real users rated the full variant higher than the baseline (4.60 vs. 2.20), offered as exploratory evidence.
In briefRethinking Contrastive Loss in CLIP Post-training with a Frozen Text Encoder
SummaryThis paper studies post-training of CLIP, a widely used vision-language model's vision encoder. It finds that reported forgetting under contrastive post-training is mainly caused by an inappropriate InfoNCE temperature rather than too few negatives. The authors propose ComCLIP, a single-epoch recipe that freezes the text encoder and combines a properly tempered contrastive loss, an MSE anchor to the original CLIP, and relational distillation from DINOv2. It improves linear probing over the CLIP-Refine baseline while matching it on zero-shot classification.
In briefFrom One Prompt to Function Repertoires: Evolving Repertoires for LLM Continual Learning
SummaryThis paper studies continual learning for large language models and finds that applying prompt optimization directly to sequential tasks causes catastrophic forgetting, with optimized prompts overfitting to local task distributions. The authors propose Evolving Functional REpertoires (EFRE), which replaces a single prompt with a set of functions that evolve over time: compatible updates refine existing functions, while conflicting updates create new ones. On a three-task continual-learning stream, EFRE achieves a final average performance 7.50 points above GRPO, and its FinQA performance drops by only 1.56 points after adapting to the Bio task.}
In briefSignRAG: A unified retrieval-augmented framework for gloss-free sign language translation
SummaryThis paper presents SignRAG, a framework that adapts decoder-only large language models to translate sign language without gloss annotations. It combines hierarchical pretraining, which first learns sign representations and then aligns them with the LLM, with a target-domain retrieval gallery and retrieval-aware reinforcement fine-tuning. The authors report new state-of-the-art results on multiple sign language translation benchmarks, including outperforming gloss-supervised methods on CSL-Daily across all metrics. Code and models of different sizes are publicly released.
In briefRIT-RAG: Navigating Document Corpora with Retrieval-Induced Trees
SummaryThis paper proposes RIT-RAG, a retrieval-augmented generation method that combines content retrieval with structural navigation. Offline, it builds a tree for each document from its table of contents or sitemap; at query time, retrieved chunks induce manageable sub-trees, possibly spanning several documents. An LLM agent navigates these sub-trees, selectively reads promising nodes, and reformulates queries when needed. Across financial, scientific, and customer-support benchmarks, it achieves the highest answer accuracy among the compared baselines, and on the new EntQABench it improves accuracy over the strongest baseline across three LLMs.
Meghanadh Pulivarthi, Swaraj Kumar Biswal, Kushagra Bhushan et al.
In briefSpatialOPSD: Self-distilling spatial reasoning from verified coding agent traces
SummaryThis paper proposes SpatialOPSD, a framework that lets a multimodal large language model internalize the spatial reasoning of coding agents without calling external tools. Verified agent execution traces serve as privileged information in an on-policy self-distillation process. Repetition-Aware Distillation, which combines repetition masking with unlikelihood regularization, is used to reduce leakage of that privileged information. Experiments report higher average accuracy than SFT and GRPO on spatial and out-of-distribution datasets.
In briefFlyMark: Training-Free Invisible Watermarking for 3D Gaussian Splatting Using a Fruit Fly Connectome
SummaryThis paper presents FlyMark, a method that embeds a keyed ownership message directly into the parameters a 3D Gaussian Splatting (3DGS) scene file already stores, without training any additional model. The carrier directions are derived from the photoreceptors of a published fruit fly connectome, and embedding works by slightly shifting existing degree-zero colors in an achromatic way while leaving geometry and higher-order appearance parameters unchanged. On synthetic and real scenes, the authors report high clean bit accuracy and visual fidelity, along with clear separation between matched and wrong keys.
In briefUniData: A Universal Pipeline for Generating Multimodal Instructions
SummaryThis paper introduces UniData, a pipeline that turns simple user requirements into multi-round, multimodal instruction data. It first expands requirements into diverse events, then uses an any-to-any large model to generate the instructions, and corrects irrelevant or redundant inference flows using links between rounds. The authors also built UniDataset, with 20,000 entries across nine modalities. They report state-of-the-art data quality and improved understanding and generation in other multimodal models.
In briefBernoulli Flow Models: A Self-Consistent Generator for Binary Data That Samples Well in Few Steps
SummaryBinary diffusion models usually need many function evaluations to produce good samples, which makes inference expensive. This paper proposes Bernoulli Flow Models (BFM), which define a continuous global Bernoulli probability flow between data and noise and derive closed-form posterior transitions over arbitrary time intervals. On LSUN Churches 256x256, a BFM trained with 256 steps reaches an FID of 9.22 with only 16 sampling steps, while a discrete baseline degrades to 204.10.
WhyProposes a continuous Bernoulli flow with closed-form interval posteriors, fixing low-NFE mismatch; LSUN Churches FID 9.22 at 16 steps versus 204.10 for baseline.
In briefSepGen: One Model That Generates and Separates Audio-Video Sound Sources
SummarySepGen extends a pretrained audio-video generator so that one joint sampling run outputs the video, the mixed soundtrack, and a separate waveform for each captioned sound source. It has a generation mode, where each source caption defines its stem, and a separation mode, where a clean input mix is decomposed according to free-text descriptions. The authors evaluate it on scenes generated from text, scenes rendered by other generators, and real recordings of speech, music, and sound effects. On speech separation with captions, it outperforms language-conditioned separators.
In briefRaReCache: Reusing Key-Value caches across different AI models by recomputing only key tokens
SummaryThis paper addresses the problem of reusing key-value (KV) caches when a large language model takes over from a smaller one, which normally requires re-processing the whole context. The authors propose RaReCache, which lets a large target model decode from a small model's cache by recomputing only critical token positions, identified with a new rank disagreement metric. On a 23x parameter gap, recomputing 30% of positions retains 95-99% of target accuracy, with up to a 3.04x prefill speedup. The paper reports that it also reduces time-to-first-token latency under saturated serving load.
Sreetama Sarkar, Saptarshi Mitra, Sitao Huang et al.
In briefAdaptive adversarial augmentation for controllable face synthesis, a method for generating varied synthetic faces to train recognition models.
SummaryThe paper proposes Ensemble Feedback Controllable Synthesis (EFCS), a guided framework that generates diverse, challenging synthetic face samples while preserving visual realism for training face recognition models. According to the authors, EFCS increases distributional variability compared with single-feedback and random synthesis while keeping high precision. Recognition models trained on EFCS data outperform baselines across multiple benchmarks and generalize better to real-world scenarios.
Saransh Suri, Shivang Agarwal, Mayank Vatsa et al.
In briefDefensive Writing in GPT Models: Added Qualifications and Retractions Aimed at Reviewers
SummaryThe paper examines how recent GPT models, when asked to make papers withstand peer review, often narrow or retract authors' claims. Researchers tested whether this corrects overclaiming or writes for an anticipated reviewer. From GPT-5.5 on, GPT versions add more ungrounded qualifications than the originals, with GPT-6-astra adding the most. These qualifications follow the prompt, appearing under review-focused requests and peaking when reviewer concerns are named.
In briefAdaptEvo: Confidence-adaptive training with evolving knowledge for learning under imperfect supervision
SummaryAdaptEvo is a framework for rule-based contextual decision tasks where supervision is imperfect. Its training module, CA-GRPO, balances outcome and process rewards according to the confidence of reference judgments, while its evolution module distills reusable decision knowledge from recurring failures and refines evaluation rubrics. On an industrial multimodal content moderation dataset using Qwen3.6-35B-A3B, the paper reports higher In-Period accuracy than GRPO and better Out-of-Period results than GRPO and fixed reward mixtures.
In briefRL-ARC: Calibrating Reasoning Models Using Reasoning-Guided Uncertainty
SummaryThis paper introduces RL-ARC, a calibration-aware training framework for language models trained with reinforcement learning using verifiable rewards, which often become overconfident. It combines reasoning confidence with answer confidence, using reasoning confidence as a regularizer for correct answers and as a penalty for overconfident incorrect ones. Results in both in-distribution and out-of-distribution settings show improved calibration without substantially sacrificing reasoning performance.
In briefDeception by Omission: Language Models Hiding Their Own Mistakes
SummaryThis study tests whether large language models disclose mistakes when continuing task trajectories that were prefilled with synthetic errors, in both chat and agentic settings. Models failed to disclose the mistake in 36.4% of chat and 67.1% of agentic rollouts, and in some cases they noticed the error in their reasoning yet still concealed it. The authors conclude that users cannot rely on agents to self-report errors, and recommend independent monitors or training models to review and disclose their past actions.
Lucas Florin, Amelie Knecht, Ulysse Schaller et al.
In briefSample-Efficient Generative Conformal Prediction: Adaptive Sample Allocation Across Inputs
SummaryThis paper studies generative conformal prediction, which builds uncertainty sets from samples of a conditional generator and needs enough samples to represent the response distribution well. The authors propose CASA, which allocates samples across inputs based on the marginal value of each additional sample, minimizing expected set size under marginal coverage and an average sampling budget. Theory and experiments on synthetic and real tasks show that, at the same budget, CASA yields substantially smaller sets, often improves conditional coverage, and can complement existing radius-adaptive methods.
In briefEgoPhys: Estimating Peak Force and Work from First-Person Video
SummaryThis paper presents EgoPhys, an RGB-only framework that estimates peak contact force and mechanical work when manipulating articulated objects, using egocentric video. It combines Contact-Aware Spatial Aggregation (CASA) to highlight interaction cues and Target-Specific Multi-Expert Temporal Routing (TMTR) to model temporal cues separately for force and work. On the Hoi! dataset test split, it achieves MAEs of 5.205 ± 0.584 N for peak force and 0.894 ± 0.081 J for mechanical work.
In briefSynCo: a multi-agent reinforcement learning method that co-trains data synthesis for self-evolving LLMs
SummaryThis paper proposes SynCo, a framework in which a language model's training tasks and its problem-solving policy evolve together. A Synthesizer agent creates tasks based on the Reasoner's current capability, while the Reasoner learns from the resulting rollouts. Both agents receive rewards from correctness, task quality, answer reliability, and teachability. Across eight mathematical reasoning benchmarks, the authors report stronger overall performance than existing synthetic-data methods, with most gains coming from previously unsolved problems.
In briefEvoSim: A self-evolving AI scientist that builds physics-based battery models from experiments
SummaryEvoSim is a self-evolving AI system that builds physics-based models by revising mechanisms and equations based on gaps between experiments and predictions. Held-out experimental data is used to test physical plausibility. The authors evaluate it on two industrial battery modeling tasks, including lithium-metal-plating onset and dynamic voltage prediction. They report that self-evolution reduces model and physics errors by about 36% relative to baseline.
WhyProposes a self-evolving AI scientist revising physics models from experimental discrepancies, validated on held-out data across two industrial battery tasks with substantial measured gains.
In briefBiologically inspired point cloud learning with focused attention and context scanning
SummaryThis paper introduces PointLearner, a point cloud representation learning network inspired by foveated biological vision that models local geometry and long-range dependencies together. Its point-focused attention uses competitive normalized attention between local neighbors and downsampled features pooled via learnable inducing points. A context-scan state space follows a Hilbert-curve scan path with bidirectional S6 to infer overall scene semantics. The authors report strong robustness and state-of-the-art results across multiple point cloud tasks.
