In briefA³PreCore: Robust Rigid Registration for Brain-Mask Transfer in Native FLAIR Space
SummaryThis paper presents A³PreCore, a six-degree-of-freedom rigid registration method that aligns bias-corrected T1-weighted images to native FLAIR images so brain masks transfer reliably. It also introduces extended quantitative quality control (E-QC) to characterize masks along several geometric and intensity dimensions. In 78 participants, gross misregistration occurred in 21 fsl_anat outputs and in none of the A³PreCore outputs.
Elmira Mirzabeigi, Zahra Khalilzad, Taylor Snowden-Richardson et al.
In briefPhySGA: A physics-grounded agent framework for interactive 3D object assembly
SummaryThis paper presents PhySGA, a framework that assembles 3D objects through agent exploration and physical simulation, handling rigid, articulated, and deformable objects. It uses an Assembly Agent for global planning and coarse adjustment, a Critic Agent for simulator state diagnosis and fine-grained correction, Action Routers that map plans to physical primitives, and a self-evolving skill library. The authors report that experiments in Isaac Sim show improvements over existing methods in relation fidelity, physical plausibility, and visual quality.
In briefSpeculative Sampling: Large bidirectional windows for low-latency interactive video generation
SummaryInteractive video generation usually uses small bidirectional chunks to reduce latency, but larger chunks can improve quality and use GPU parallelism better. The paper proposes Speculative Sampling, which expands the modeling window by appending noise latents as padding and caches predictions for reuse when the condition is unchanged. Predictions are delivered incrementally, while uncommitted future content stays replaceable. The authors' latency analysis indicates this removes the waiting-time lower bound set by chunk boundaries, shifting optimization toward faster per-chunk sampling.
In briefOpen Benchmark of AI Impact on Humans (ImpactBench): an open, expert-guided platform for assessing how AI affects people's wellbeing
SummaryThis paper introduces ImpactBench, an iterable, expert-guided open platform for measuring how generative AI models affect users' psychological, physical, and social wellbeing. It delivers 809 metrics drawn from 26 benchmarks and tests ten frontier LLMs in simulated multi-turn conversations with matched teen and adult scenarios, totaling 48,540 conversations. An LLM judge scored the conversations, checked by reliability audits and human expert review. Models promoted desirable behaviors in 68.7% of evaluations but avoided harms in only 53.1%.
WhyTranslates expert constructs and existing benchmarks into behavior-grounded metrics, evaluating ten frontier models across 48,540 conversations with human audits.
Chayapatr Archiwaranguprok, Krystal Kirkland, Rachel Poonsiriwong et al.
In briefPerSo: Testing LLMs' social intelligence in real-world persuasion dialogues
SummaryThis paper introduces PerSo, a benchmark that evaluates large language models' social intelligence through both cognitive inference of others' mental states and behavioral choice of social actions. It uses real persuasion dialogues from Reddit's ChangeMyView community, with tasks built through an LLM-in-the-loop pipeline and validated by human annotation. The results show performance generally rises with model size but varies across dimensions, with the strongest model slightly exceeding the best human on the behavioral task while the best human stays ahead on the cognitive task.
Ali Zamani, Soroush Esfahanian, Erfan Shahabi et al.
In briefDetecting potato leaf diseases from photos with a compact neural network
SummaryThis study develops a lightweight convolutional neural network to classify potato leaves as healthy, early blight, or late blight. It was trained on 2,215 PlantVillage images and tested on 247 independent field images from Ethiopian farms, reaching 90.28% accuracy in the field. The paper also compares the model with VGG16 and MobileNetV2 in classification performance and computational efficiency.
In briefSentiment Prediction When Parts of Multimodal Input Are Missing
SummaryThis paper studies multimodal sentiment prediction when contiguous spans of text, audio, or video are missing, using aligned CMU-MOSEI features. The authors propose RM-Net, which marks unobserved positions with a learned token and is fused with a LoRA-adapted BERT text model. On the held-out split, the fused model reaches macro-F1 of 0.655 on clean inputs and 0.634 under 40% mixed missingness, though the authors note this split was inspected during development and is not a blind test.
In briefPrismaCoder: Advancing open multimodal code generation with decomposed rubric synthesis
SummaryThis paper introduces PrismaForge, a data pipeline that uses a synthesized rubric to break down program correctness, code quality, and visual correctness after rendering into checkable criteria. The rubric is used both to filter training data and as a reward for reinforcement learning on visual coding tasks. The authors built PrismaCoder-2M, an open dataset for multimodal code intelligence, and trained and released the PrismaCoder-9B and PrismaCoder-35B models. The paper reports that these models lead open models by a wide margin and rival leading proprietary systems.
In briefSimple baselines for relational deep learning, revisited
SummaryThis paper examines whether simple methods can match complex models for predicting outcomes from relational databases, represented as temporal heterogeneous graphs. The authors apply fixed low-pass feature propagation to relational entity graphs and train an MLP, then refine predictions with Correct-and-Smooth, a label propagation method that uses historical labels. On RelBench classification and regression tasks, the MLP with neighborhood-filtered features outperforms target-entity-only features on most tasks and recovers much of the performance of heterogeneous GNNs. The authors also note that Correct-and-Smooth requires a separately constructed meta-graph, which they identify as a key challenge.
Viet The Nguyen, Vijay Prakash Dwivedi, Gunther Gust
In briefGenerative representation learning for scalable latent tokenization
SummaryThe paper proposes a self-supervised representation learning method meant to serve as a tokenizer for generative modeling. It supervises a representation loss by factorizing the data's pointwise mutual information, estimated from differences between conditional predictions of a learned generative model. The authors report that it avoids dimensional collapse and feature suppression where DINO and LeJEPA fail, and that its quality improves predictably with more compute. Scaling results are reported on ImageNet and GPIC.
WhyUses factorized pointwise mutual information from a generative model as a representation objective, with strong FID and compute-efficiency experiments against DINOv2 and WebSSL.
In briefActiveEvidence: Decision-focused evidence graphs that certify when to stop acquiring visual evidence
SummaryThis paper studies how a multimodal agent decides whether to acquire visual evidence, which candidate or resolution to query, and when further interaction can no longer change its decision. It proposes a Counterfactual Evidence Interaction Graph and a gap-directed refinement method that stops once an epsilon-decision certificate guarantees the selected action is unchanged within a tolerance. On 640 cases, the controller reached a task score of 0.842, evidence recall of 0.873, 1.74 accepted calls, and 742ms p95 latency.
In briefPhysics as Network: Training-Free Fast Neural Simulation
SummaryThis paper proposes PolyNet, which couples a continuous-time neural representation of polynomial dynamics with a simple recurrent network for discrete-time integration. When the coefficients are known, it requires no training, preserves the prescribed numerical scheme, and runs efficiently on hardware. Tested on the one-dimensional Korteweg-de Vries equation against finite-difference Runge-Kutta solvers and three trained neural models, it reaches accuracy comparable to the numerical solvers and achieves high GPU throughput.
In briefThe Wealth of Agents: Scaling Agentic AI Across Developers
SummaryThis paper proposes multi-developer agentic systems (MDAs), in which several developers combine their own agents into one collaboratively built system instead of a single developer controlling everything. On the GAIA general assistant benchmark, specialized developers whose individual agents reach at most 35% accuracy reach 70% when combined in an MDA. This is 12 points above the median general-purpose single-developer system built with several times each specialist's budget. The authors also present theoretical models showing how MDAs can encourage specialization and more efficient use of developer resources.
Austin W. Hanjie, Carlos E. Jimenez, Zirui Wang et al.
In briefVASPO: Self-Distillation Policy Optimization with Privileged Value Estimation (a training method that lets reasoning models use both teacher guidance and value estimates)
SummaryThis paper proposes VASPO, a method for training reasoning models that combines token-level guidance from a privileged self-teacher with a critic's predicted returns inside the PPO framework. The authors observe that privileged information improves the critic's value prediction accuracy toward the end of responses, while self-teacher guidance is strong early, weakens mid-generation, and rebounds at the end. Across six mathematical reasoning benchmarks, VASPO improves mean accuracy over PPO by 2.59 and 1.81 percentage points on Qwen3-4B and OLMo3-7B, respectively, and achieves the highest mean accuracy among evaluated methods on Knights and Knaves.
In briefBayesian Bandit Method for Choosing Fair Classifiers Online
SummaryThis paper studies how to choose, online and as data arrives, which classifier (AI model, large language model, or human labeler) to query, balancing predictive accuracy with subgroup fairness over a sensitive attribute. It allows true class labels to be observed with delay. The authors frame the task as a Bayesian multi-armed bandit problem solved with Thompson Sampling. Experiments on real datasets involving tabular classifiers, language models, and human labelers show substantially lower regret than baselines, and the paper provides theoretical guarantees covering both accuracy and fairness.
In briefTwo-level framework for ranking clinical novelty of disease trajectories from longitudinal EHRs
SummaryThis paper presents a two-level framework for prioritizing disease trajectories discovered from graph-based analysis of longitudinal electronic health records. It measures path-level and node-to-endpoint semantic sparsity using sentence embeddings and calibrates it against PubMed references, then combines it with reliability-adjusted clinical priority to rank 99,527 branching motifs. In a high-semantic subgroup, branch-enhanced forecasting reduced conditional event-time error, while a monotonic link between semantic score and forecasting benefit was not supported.
In briefOn-policy residual learning (ORL) for diffusion language models: a training method that lets the model refine its own predictions
SummaryThis paper proposes ORL, a framework for on-policy residual learning in diffusion language models (DLMs), which carries predictive distributions across denoising steps to improve the speed-quality trade-off. It learns from the model's own residual trajectories and uses an entropy-contraction objective to sharpen predictions, alongside systems work such as asynchronous residual capture and block-diffusion attention. The authors report up to 6.2x higher tokens per forward and about 6.9x higher end-to-end throughput on a 4B model at comparable or better quality.
In briefVascuSR: Anatomy-aware diffusion model that improves vessel detail in CT super-resolution
SummaryVascuSR is a diffusion framework that improves CT image super-resolution by adding explicit 3D vascular guidance and iterative feedback, aiming to avoid anatomically unsupported vessel details. It was tested on re-annotated MSD Task08 and HVM datasets. It outperforms interpolation and learning-based baselines on segmentation accuracy and 3D vascular structure metrics, while keeping image quality comparable to its backbone.
In briefReaDy: Using Recurrent States as Dynamic Context for Faster Speculative Drafting in Hybrid Linear Attention LLMs
SummaryThis paper introduces ReaDy, a speculative decoding method that uses the recurrent states of hybrid linear attention (HLA) models to retrieve supplementary context for each draft position. A lightweight state reader integrates this context into draft representations through residual updates, with a two-stage adaptation strategy. Across seven benchmarks covering mathematics, code generation, and dialogue, ReaDy achieves an average 3.16 times decoding speedup over autoregressive decoding and outperforms DSpark on every benchmark.
In briefGraph-enhanced disease trajectory discovery and time-to-event forecasting from EHRs
SummaryThis paper proposes a graph-enhanced framework that discovers disease-progression trajectories from longitudinal MIMIC-IV electronic health records and forecasts endpoint events. Patients are split before graph construction, transitions are learned only from training data, and a Transformer retrieves up to 24 branching motifs per endpoint. Event-time error decreased across five seeds, but the gains were heterogeneous over time, and superiority in survival prediction was not established.
In briefOLiVES: An outdoor low-light video benchmark for enhancing and segmenting dark footage
SummaryThis paper introduces OLiVES, a benchmark dataset of outdoor low-light videos with dynamic scenes, built to support both low-light video enhancement and video object segmentation. It was captured with professional cinema-grade cameras and contains over 25,000 aligned normal- and low-light frame pairs plus over 46,000 segmentation annotations. The authors benchmark state-of-the-art methods on both tasks and report that training on OLiVES improves performance.
In briefHindsight-Privileged RLVR: Per-token credit from a single rollout, no group sampling
SummaryThis paper proposes HP-RLVR, a method that uses each completed trajectory's realized future and verifier feedback to set token-level update weights from a single rollout. It needs no reference solution or learned critic and uses one eighth the rollouts per prompt of GRPO. In paired Qwen3-8B experiments, HP-RLVR trains 1.21x, 1.46x, and 2.30x faster than GRPO on mathematics, code generation, and ALFWorld, with overall accuracy comparable.
In briefPrivacy risks in brain MRI foundation model embeddings
SummaryThis paper examines whether learned embeddings of brain MRI scans still retain patient-specific information. The authors test five frozen pretrained encoders on two longitudinal T1-weighted datasets, OASIS-2 and MIRIAD, under an attacker who has the embeddings, the public encoder, and auxiliary scans. The highest-risk representation reaches up to 99% rank-1 patient linkage on MIRIAD and over 90% on OASIS-2, and recorded-sex prediction reaches 90.9% and 95.5% balanced accuracy. The results indicate that replacing raw MRI with embeddings alone does not establish anonymity.
Imen Bajar, H\'eber Hwang Arcolezi, Antoine Boutet et al.
In briefLearning the best policy from one continuous trajectory of an agent's experience
SummaryThis paper studies reinforcement learning when an agent can learn only from a single continuing trajectory, rather than freely sampling any state-action pair. For average-reward settings, the authors develop an algorithm that decides where to collect information while ensuring the needed samples can be obtained efficiently. They give an upper bound on the interactions needed to identify a near-optimal policy and prove a matching information-theoretic lower bound. The results show that, for a broad class of environments at sufficiently high accuracy, single-trajectory learning is no harder in the worst case than unrestricted sampling.
WhyMatching upper and lower bounds show single-trajectory average-reward learning is no harder than free sampling at high accuracy.
In briefDecoupled Marginal Sharpening: A Training-Free Method for Better Reasoning at Inference Time
SummaryThis paper studies how to improve large language model reasoning at inference time without updating weights or using external reward models. It proposes Decoupled Marginal Sharpening (DMS), which separately controls how strongly the model favors particular reasoning paths and how strongly it favors final answers, paired with an annealed Sequential Monte Carlo sampler. The authors report that DMS achieves the highest mean pass@1 among compared methods on MATH500, GPQA-Diamond, HumanEval, and AIME24/25 across three base models.
In briefLost in Translation: Why Lower Bias Scores Don't Mean Fairer Models — a study on whether multilingual bias audit averages are reliable
SummaryThe paper audits ten open models in eight languages and finds that stereotype scores fall below English levels in all 70 model-language pairs, with drops for Hindi, Korean and Swahili clearly exceeding those caused by rewording English. The authors argue that a lower average does not show a model is fairer, since item-level scores correlate much less between English and translated items. They also document three common pipeline errors that erase bias signal and release their code.
WhyItem-level audits across ten models and eight languages show lower average bias scores do not reliably indicate fairness, with careful controls for noise.
In briefAdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents
SummaryPretrained vision-language-action models often fail midway through unseen long-horizon tasks. AdaRSI lets a coding agent first improve its skills and harness, then retry failing sub-steps from saved states to find persistent failures, and fine-tunes only the unreliable sub-steps using self-collected demonstrations and carried-over experience. On 50 target tasks in RoboCasa365 with GR00T N1.5 as executor and no human demonstrations, it improves all three suites and raises success on unseen composite tasks from 4.6% to 13.8%.
In briefSafeAdapt-Poker: Budget-Constrained, Risk-Calibrated Adaptation to Changing Opponents
SummaryThis paper addresses poker agents whose opponent histories can mislead them after the opponent's behavior shifts. SafeAdapt-Poker is a sequential controller that, at each decision, weighs the strategic risk of deviating from a frozen anchor strategy against the future adaptation value lost by spending the remaining deviation budget. A cumulative-KL ledger bounds total deviation from the anchor. Results are reported on exact Leduc poker, abstract HUNL cross-play, and public-history prediction benchmarks.
In briefChronoCulture: Testing whether AI models track how cultural attitudes change over time
SummaryThis paper introduces ChronoCulture, a benchmark of 10,360 temporal pairs built from longitudinal survey data across 72 countries. Each pair asks the same survey question in the same country in an older and a newer year, using full population response distributions as ground truth. Testing six large language models, the authors find they can often predict attitudes at one point in time but do not reliably capture how those attitudes change, and specifying the year in the prompt gives little overall improvement.
In briefRoLaQ: Joint rotation and range calibration for 2-bit KV caches in latent attention
SummaryThe paper introduces RoLaQ, an offline method that calibrates a shared rotation and quantization ranges to store multi-head latent attention key-value caches at two bits. It refines the rotation within subspaces set by the frozen key and value projections, and learns thresholds and reconstruction gains jointly. Across three MLA models and six benchmarks, it reports the highest average accuracy among compared 2-bit methods. On DeepSeek-V2-Lite with SGLang, it provides 3.96 times the BF16 KV capacity and 2.51 times the decode throughput at 256K input tokens.
In briefSelective Growing Q-Networks: Adaptive action grids in decoupled Q-learning
SummaryThis paper studies value-based Q-learning for continuous control, where discretising actions trades early exploration against final precision. The authors propose SGQN, which replaces uniform grid growth with an adaptive grid that adds bins near high-value actions and prunes low-value ones, and compare it on eight DMControl tasks. They find the adaptive grid itself does not explain the results; starting coarse and epsilon-greedy selection matter more, and SGQN also differs from baselines in other respects, so no single component is credited.
In briefLearning by Examples: A Pretrained Tabular Model for Power Grid Steady-State Analysis
SummaryThis paper tests whether TabPFN, a pretrained tabular foundation model, can act as a general neural solver for power grid steady-state analysis. It is evaluated on power flow prediction, N-k contingency screening, and voltage stability margin prediction on 118- and 2000-bus systems. Without retraining or fine-tuning, it matches or outperforms task-specific baselines in several settings, suggesting potential for a range of grid analysis tasks.
Olayiwola Arowolo, Oluwadara Adedeji, Peter Palensky et al.
In briefEqual CER, Unequal Damage: A Severity-Weighted Error Rate for OCR
SummaryThe paper proposes semCER, a modified character error rate for OCR and vision-language models that keeps CER's alignment and denominator but weights each character edit by severity assigned by an LLM judge, without word segmentation. In English, Vietnamese, and Chinese, the severity charges track readers' ratings edit by edit. The authors release 61 million judged edits, aggregated reader ratings, and a distilled scorer.
WhyProposes severity-weighted CER per grapheme edit, validated against reader ratings across languages and 61M judged edits, a substantive improvement for OCR evaluation.
In briefMolLean: A Process-Verifiable Language for Chemical Reasoning
SummaryThis paper introduces MolLean, a language that grounds chemical reasoning in an explicit molecular state to reduce hallucinated structural and property claims. A Lean-based kernel checks state transitions and assertions, and MolLean-MCP provides stepwise verification feedback during generation. On 8,800 matched examples across 12 tasks, whole-trace refinement outperformed direct chain-of-thought with Qwen3.6-27B on all tasks, and MolLean-MCP performed better on reaction tasks.
In briefSame Claim, Different Interpretations: How LLM Agents Resolve Ambiguity
SummaryThe paper argues that scientific claims are often underspecified, so verification agents may silently pick one of several defensible readings. The authors introduce Claim Cards, which record the verifier's interpretation in nine fields, each marked as stated, inferred, or default, before evidence gathering. In 900 runs of a GPT-5.4 agent on 60 expert-labeled claims, cards did not measurably change agreement with expert labels, but editing one field changed verdicts on about a third of claims.
In briefAutomatically determining causes of driving actions in road scenes using qualitative reasoning and four-valued logic
SummaryThis paper addresses road scene understanding for automated driving by determining the causes behind an ego car's actions, such as stopping, slowing down, or turning. The authors propose combining spatiotemporal reasoning with Belnap's four-valued logic to handle complex, possibly contradictory causes. The method analyses a Qualitative eXplainable Graph (QXG), an abstract representation of spatiotemporal relations between road entities. The approach is demonstrated on several scenes from the NuScenes dataset.
Nassim Belmecheri, Arnaud Gotlieb, Nadjib Lazaar et al.
In briefL2G-SR: Guiding LLM-based symbolic regression with a learned structural verifier
SummarySymbolic regression searches for equations that explain observed data. The paper argues that numerical error alone can favor accurate but structurally wrong formulas. L2G-SR uses a learned verifier, trained with tree edit distance supervision, to score structural similarity and complement numerical error during parent selection. On LLM-SRBench, it reports state-of-the-art numerical fit, equation-level accuracy, and tree edit distance across the evaluated LLM backbones and domains.
In briefTopology-aware learning of representations for mechanical TPMS metamaterials
SummaryThis paper proposes a representation learning framework that integrates persistent homology into a DMAE backbone to encode global structural connectivity of TPMS metamaterials. A persistence-guided masking strategy and an auxiliary persistent-homology decoder keep topological feature counts in the latent representation. Evaluated on two TPMS datasets, the method improves family classification and property clustering and retains strong regression performance.
In briefCOUNT: Scaling Language Models with Host-Memory N-gram Counts and Training-Free Adaptation
SummaryThis paper introduces COUNT (Counting Offline Unlocks N-gram Transformers), which builds n-gram embeddings without learned parameters by storing precomputed n-gram counts in host memory. At a fixed compute budget, adding COUNT as an extra input embedding for pretraining 1B-parameter language models improves validation loss as much as a model using 57% more compute. The authors also report that COUNT can switch to an n-gram count table from a domain-specific corpus at inference time to adapt to new tasks without training.
In briefWhat Disagreement Means: An Identifiability Limit for Pairwise Preference Learning
SummaryThis paper asks whether disagreement in pairwise comparison data reflects stable heterogeneity, random noise, or non-binary answers hidden by forced-choice interfaces. The authors propose a latent semantic model with three quantities: aggregate preference, persistent disagreement, and non-binary mass. They prove that aggregate preference is uniquely identified from binary feedback, while the other two vary over a feasible region. Repeated same-annotator labels and abstention options identify them, and theory is checked on synthetic data.
Manoj Saravanan, Rohit Kumar Salla, Shrikar Reddy Kota
In briefCausal syntax circuits in code models: budget and probe shape whether a preference looks localised
SummaryThis paper asks whether code models' habit of writing Python when no language is specified comes from a small, findable set of internal components. The authors run one pipeline, including mean ablation, activation patching, greedy circuit extraction, random-control tests and activation steering, across 55 model-language settings. They find the preference is large, grows with model size and persists after instruction tuning, but its measured size depends mostly on the probe, and whether it looks localised depends on the component budget.
WhySystematic grid of 55 settings shows localisation claims hinge on probe and budget choices, a substantive critique of circuit methodology.
Harsh Kumar, Tanmay Shailesh Joshi, Rahul Maity et al.
In briefTurboMamba: a language model design for faster inference and training with better next-token prediction
SummaryThis paper introduces TurboMamba, a language-model architecture in which a global Mamba-3 model runs once per completed token block and a small Transformer predicts the next token using that global context and earlier tokens in the block. Three lightweight additions improve local predictions. On the public FineWeb-Edu dataset, compared with twelve small-scale models including Transformer, Mamba-3, Jamba and Zamba, TurboMamba reports the lowest mean test negative log-likelihood, along with higher training throughput and faster batch latency.
In briefIneqAgent: an agent that proves inequalities in Lean by combining language models with symbolic computation
SummaryThe paper presents IneqAgent, an agentic framework that pairs LLM reasoning with symbolic computation to prove inequalities in Lean4. Lemma-level Monte Carlo Tree Search guides lemma construction and goal reformulation, while sum-of-squares, Bernstein, and difference-substitution backends produce exact proof certificates. On four benchmarks, it solves 175 of 186 problems (94.09%), and it also yields Lean-verified strengthenings of Dittert's conjecture in dimensions four and five.
WhyCombines lemma-level MCTS with symbolic backends for Lean inequality proofs, verifying Dittert and Merris-related results across 186 benchmark problems.
In briefThe Accuracy Well: Accuracy That Rises, Falls, and Recovers During Training
SummaryThis paper studies how classification accuracy changes during training when networks start from very small initial weights. It finds that accuracy rises quickly, then drops, and only much later recovers, a pattern the authors call the "accuracy well." The phenomenon is observed in deep linear networks, ReLU networks on MNIST, convolutional networks on CIFAR-10, and synthetic tasks. The authors give a three-stage explanation and a theoretical analysis for linear networks, showing that accuracy can reveal learning dynamics that training loss alone does not.
In briefInfant Fingerprints: How Well Can Their Quality Be Judged and Recognized?
SummaryThis paper examines how to assess the quality of infant fingerprints and use them for recognition, since small ridges and collection issues such as movement and uneven pressure make adult quality standards unsuitable. The authors tested ten fingerprint quality assessment methods on a longitudinal dataset covering six age groups from newborns to three months. They also tuned each method's parameters to fit age-specific characteristics. Results indicate that tuned settings outperform default ones for most methods and age groups, improving the reliability of infant fingerprint recognition.
Jo\~ao L. H. Dall Agnol, Dalcimar Casanova, Rodrigo Minetto et al.
In briefSwarmRSI: Scientific agent swarms that recursively improve themselves through experiments
SummaryThis paper presents SwarmRSI, a framework in which swarms of scientific agents improve their research abilities using their own experimental experience. It combines feedback-guided mutation, which revises hypotheses and implementations using other agents' results, with complementary recombination, which reuses retained hypotheses and program components across agents, in a recursive feedback loop. Experiments on OpenProblems, NatureBench and MLE-Bench Lite report state-of-the-art performance, and accumulated experience improves results on held-out tasks.
Qianru Meng, Xi Zhang, Konstantinos Papakostas et al.
In briefContinual learning for reliable in-sequence molecular communication
SummaryThis paper studies in-sequence molecular communication, where bio-nanomachines exchange molecule-encoded information in a specified order, and proposes a simulation-based method that evaluates communication protocols by gradually increasing experimental complexity. The method, Affine Memory Replay (AMR), uses a continual learning model to estimate communication performance and synthesizes replay samples by interpolating and extrapolating stored past simulation samples. Experiments show that AMR significantly outperforms existing continual learning methods in maintaining overall estimation accuracy across sequentially learned tasks, while balancing memory stability and learning plasticity at moderate training cost.
In briefEvaluating RAG Models for Multilingual South Asian Cultural Question Answering
SummaryThis paper evaluates retrieval-augmented generation (RAG) models on question answering in South Asian languages, where cultural and historical context matters. The authors compare answers produced without context, with context, and with RAG, and examine differences between culturally grounded and non-cultural questions. They also test whether RAG helps popular or less popular cultural questions more, aiming to identify factors for designing better hybrid RAG systems.
Fazle Mohammed Tasfiq, Sabab Attin, Sheikh Md. Tahmid Hossain et al.
In briefOrionMoE-v1.0: A Mixture-of-Experts Tabular Foundation Model for In-Context Learning
SummaryThis paper introduces OrionMoE-Small-v1.0, a compact tabular foundation model that combines distribution-aware representations with dataset-conditioned sparse mixture-of-experts computation. Built on a dual-axis transformer, it is pretrained only on synthetic classification tasks and makes predictions without gradient-based adaptation. On multiple tabular benchmarks it competes with substantially larger foundation models, with especially strong results on small-context and high-dimensional datasets.
Mohamed Bouadi, Nassim Bouarour, Aditya Tanna et al.
In briefFrom Unreliable to Usable: Decomposing LLM Entity Extraction Failures and Confidence Signals in Cyber Threat Intelligence
SummaryThis paper examines why large language models give unreliable entity extraction from full-length cyber threat intelligence reports and whether confidence signals can make their output usable. Using a preregistered, frozen evaluation on 354 incident campaigns, it compares two low-cost models with a flagship model and finds low precision across tiers. Most failures are background mentions and scope gaps rather than hallucinations, while confidence tiering ranks outputs but cannot verify them, since single-sentence perturbations often flip high-confidence judgments. The authors conclude that confidence tiering may set review priorities but should never be treated as evidence of truth.
In briefTracking a small set of experts in online convex optimization, showing a computational cost
SummaryThis paper studies a continuous version of prediction from expert advice, where a learner must compete with a comparator that switches k times among only m experts. The discrete case was already solved with an efficient optimal algorithm, but this work examines online convex optimization with a continuous action space. A discretization-based inefficient method achieves regret Õ(md+k), while the best known efficient algorithm has regret Õ(dk). Under the hypergraphic planted-clique conjecture, the authors prove that every polynomial-time algorithm incurs regret at least (k√d)^(1−o(1)) on some instances, showing this computational gap is inherent.
In briefCurvature-Aware Adaptive Perturbation for SAM: Varying Perturbation Strength Across the Loss Landscape
SummaryThe paper notes that SAM's fixed perturbation radius ignores local curvature, so perturbations can be too small or too large. The authors propose CAP-SAM, which adjusts the perturbation radius based on a sharpness ratio to better search for robust flat minima. They prove it keeps the same convergence rate as standard SAM in non-convex stochastic settings. Experiments on CIFAR-10, CIFAR-100, and ImageNet across several architectures report consistent gains over SAM and its main variants.
Alessia Rossi, Felipe Campelo, Gabriel L. Oliveira
In briefDiminishing data returns for sparse autoencoders: good function doesn't ensure reproducible features
SummaryThis paper examines how much activation data sparse autoencoders (SAEs) need, training four architectures on three language models with budgets from 5M to 500M tokens. Functional metrics such as reconstruction often approach the full-data reference at much smaller budgets, while one-to-one feature reproducibility keeps improving with more data. The authors also compare random and diversity-based data selection and provide a theoretical analysis separating statistical consistency from feature identifiability.
Sambit Ghosh, Madelyn Mathai, R. Venkatesh Babu et al.
In briefDoes momentum help reinforcement learning with verifiable rewards? A study of when it works.
SummaryThis paper studies whether first-moment momentum, implemented as an exponential moving average, helps reinforcement learning with verifiable rewards (RLVR). Using a bandit formulation, the authors prove that with exact gradients, increasing the EMA coefficient within a range that avoids overshoot speeds up optimization. With sampled gradients, they show that keeping the learning rate divided by (1 minus beta) in a safe range preserves a nonvanishing gain. Bandit simulations and Qwen experiments support choosing larger coefficients in practice.
In briefHybrid instance segmentation that learns to suppress label noise from cheap box labels
SummaryThis paper proposes a hybrid supervised method for instance segmentation that combines inexpensive bounding box labels with a smaller set of accurate pixel-wise labels. Two auxiliary models are designed to suppress mislabeling between foreground and background and noise from instance mislabeling, producing more accurate pseudo pixel-wise labels. Experiments on PASCAL VOC 2012 and Cityscapes reportedly show competitive performance with much lower labeling cost.
In briefHyperspherical Contrastive Representations for Planning and Exploration: A Self-Supervised RL Method (C-Slerp) That Learns to Stack Cubes From Scratch
SummaryThe paper introduces C-Slerp, a self-supervised reinforcement learning algorithm that learns to build towers of up to eight cubes from scratch, without prior data, rewards, demonstrations, or pretrained models. It reuses contrastive representations to infer intermediate waypoints in closed form via spherical linear interpolation, without a separate high-level policy or graph search, and adjusts the number of waypoints based on representation geometry. Experiments report success on long-horizon navigation and robotic manipulation tasks, including settings where strong baselines fail entirely. Ablations show that a fixed number of waypoints substantially degrades performance.
In briefCorrects a memorization lower bound: posterior frequencies and the cost of not fitting
SummaryThe paper shows that Feldman's memorization lower bound, whose coefficient depends on how often a point appears in the training sample, fails in both its conditional and averaged forms. It gives an exact three-point counterexample in which, under certain sampling priors, the ratio of the proposed penalty to actual excess risk grows exponentially, ruling out a universal constant-factor fix. Keeping the original model, the authors provide an exact replacement based on full-sample posterior probabilities.
In briefTraining language models to fail gracefully: switching to a safe fallback when success is unreliable
SummaryThe paper notes that reinforcement learning can encourage models to guess, since only success earns reward and a safe fallback always scores lower. The authors propose GracefulRL, which aims to preserve task success while teaching models to execute a safe fallback when acting would be unreliable. This behavior emerges from outcome-level reward alone, without fallback annotations or a verifier, and is reported across a range of tasks and scales.
Daman Arora, Xuecheng Liu, Himanshu Gaurav Singh et al.
In briefNeural operator method for upsampling and personalizing HRTFs
SummaryThe paper introduces SPARO, a neural operator that reconstructs a continuous head-related transfer function (HRTF) field, queryable at arbitrary directions, from sparse measurements. It projects measurements onto learned spherical latent modes, processes them with Transformer blocks, and decodes at query directions. On SONICOM with the LAP24 configuration, SPARO achieves the best LSD and ILD error at all sparsity levels from 3 to 100 measurements, while remaining competitive in ITD error. It reconstructs dense HRTFs 15 to 373 times faster than RANF and is also validated on HUTUBS.
Meenakshi Krishnan, You Zhang, Ramani Duraiswami et al.
In briefCortex: Storing Context Gains in Model Parameters
SummaryThis paper introduces Memory Consolidation, which aims to internalize the performance gains that context provides to a language model while keeping its other abilities intact. The authors propose Cortex, which generates prospective queries and distills context-conditioned outputs into mergeable low-rank or structural feed-forward updates. Across 4B to 120B dense and mixture-of-experts models, Cortex lies on the consolidation-retention Pareto frontier in most settings. After 50 sequential factual edits on a 14B model, Cortex-Structural recovers 88.9% of the context-induced gain and retains 99.8% of original non-target performance.
WhyUnifies context-based editing, skill learning, and steering as parametric consolidation; evaluated across many backbones and task families, with strong sequential-edit retention results.
In briefWhere to Place Gaussians: A Reusable Label-Aware Allocation Method for Budget-Constrained Gaussian Splatting of CT
SummaryThis paper studies where to place a fixed number of 3D Gaussians when rendering CT scans with Gaussian splatting so that vessels are preserved. The authors propose a reusable allocation front end that uses a patient's CT, anatomical labels and transfer function to choose Gaussian positions without per-patient optimisation, implemented as a neural prior or a seven-parameter closed form. On 20 held-out patients, both versions improve vessel-masked PSNR over uniform allocation and pruning.
In briefSpectral Indistinguishability and Learning-Curve Identifiability in Quadratic Networks
SummaryThis paper studies spectral summaries and scalar learning curves in centered quadratic networks, a setting motivated by information-organization sculpting. The authors build analytic input densities with Gaussian marginals whose spectral measures match, yet whose loss curves for certain teachers differ. They also quantify the observation counts needed under different query types. For Gaussian inputs with a rationally independent teacher spectrum, one exact loss curve identifies the initial predictor and, under additional conditions, the teacher spectrum and active dimension.
In briefTurning Human Corrections into Better Segmentation: A Mathematical Review
SummaryThis critical review examines when human corrections become useful learning for image segmentation and how to achieve better results at matched human effort. It compares active learning, machine teaching, interactive segmentation and sequential predictive evaluation within one mathematical framework. The authors separate refining the current mask from reusable learning, and simulated interaction from human evidence. No new algorithm or empirical superiority is claimed.
In briefFOLD: A Training-Free Method for Faster Long-Context Speculative Decoding
SummaryThis paper proposes FOLD, a training-free method that speeds up LLM generation through speculative decoding on long inputs. It keeps exact positions for nearby tokens and treats all distant tokens as far memory at one shared position, with attention sharpened to counter dilution. Tested on six target-drafter pairs and benchmarks up to 256K tokens, it mostly achieves the highest speedup, with the best case rising from 2.08× to 4.22× over autoregressive decoding.
In briefSTP-GUI: Spatially-Aware Token Pruning for Efficient Web GUI Agents
SummaryThis paper addresses the high inference cost of vision-language model agents that operate web interfaces from high-resolution screenshots and multi-turn histories. It proposes STP-GUI, a query-guided token pruning method that selects tokens in spatially coherent windows so small elements like buttons and icons keep their local structure. At 21% token retention on Online-Mind2Web, it keeps 77.7% of full-token performance while cutting latency by 31% and FLOPs by 57%. Results on Android in the Wild suggest the method also transfers to mobile GUI environments.
In briefFrom Relevance to Validity: Zero-shot tool selection with functional and contextual checks
SummaryLLM agents often face unfamiliar tasks and must pick tools without prior examples. The paper notes that tools judged semantically relevant may still fail functionally or contextually after transfer. It proposes a progressive validation framework that adds endpoint-grounded functional compatibility and task-conditioned contextual evidence to relevance ranking. Experiments on cross-domain and cross-dataset transfer show consistent ranking improvements, with larger gains under ambiguity, and ablations confirm both validation components contribute.|
In briefPredicting Tool Outcomes Before Image Restoration Agents Act
SummaryThis paper presents PreActIR, an agentic image restoration framework that forecasts the effects of each candidate tool before running it, including residual degradation, image quality, fidelity, harm risk, and prediction uncertainty. The agent executes the top-ranked tool, and an independent verifier commits or rolls back the output before replanning. In a controlled evaluation with a fixed candidate pool, verifier, and budget, it used 42.5% fewer real tool calls and 63.4% fewer rejected calls than execute-then-evaluate trial and error, while improving Y-PSNR by 0.92 dB.
In briefWeighted-median method for choosing, from known candidate distributions, the one closest to an unknown target
SummaryThe paper studies selecting, from n known candidate distributions, the one closest in total variation distance to an unknown target distribution, using only target samples. The authors give an algorithm that uses O(ε⁻²log(n/δ)) samples and runs in O(nε⁻²log(n/δ)) time, returning with probability at least 1−δ a hypothesis within 3 times the best distance plus ε. Compared with prior work, the running time's dependence on the failure probability δ improves from polynomial to logarithmic.
In briefFrom Triage to Impact: A Multi-Hop Reasoning Benchmark for Explainable Persian Hate Speech
SummaryThis paper introduces a fine-grained, multi-hop benchmark for explainable Persian hate speech analysis, covering content triage, target identification, tactic and moral foundation analysis, evidence extraction, narrative inference, and severity assessment. Evaluating large language models, the authors find performance generally declines as tasks require finer semantic and inferential reasoning, and adding intermediate reasoning steps substantially improves severity prediction. The benchmark is publicly available on GitHub.
Ghazal Kalhor, Sima Kamandloo, Mehrsa Shamsian et al.
In briefReAl: a plug-in that refines neural PDE surrogates by aligning their long-run dynamics
SummaryNeural PDE surrogates trained with mean squared error often miss high-frequency, fine-scale structures. The paper introduces ReAl, a model-agnostic add-on that uses a conditional flow-matching model to learn the residual details around the base prediction, then aligns multi-trajectory rollouts with trajectory-level rewards. Experiments across several surrogate architectures and PDE tasks report gains in high-frequency fidelity, forecasting accuracy, and long-horizon stability.
In briefSCOUT: A cascaded retrieval method that refines visual queries with semantic signals for faster visual document search
SummaryThis paper addresses retrieving information from visually rich documents with tables, figures, and layout-dependent content. The proposed SCOUT framework has two tracks: a visual track that filters candidates with a lightweight retriever and reranks them with a multi-vector model, and a semantic track that searches offline-generated page and document summaries. Results from both tracks are fused through cross-modal query refinement. On 4 benchmarks against 6 baselines, the authors report query-time latency up to 80.7x faster (28.8x on average) than the best baseline and NDCG up to 7.8% higher.
Diana Turmakhan, Aseem Srivastava, Vivek Gupta et al.
In briefWord persistence without lasting meaning change in AI agent communities
SummaryThe paper compares MoltBook, an AI agent community over 8 weeks, with the Chinese BBS PTT (19 years) and the English forum Reddit over matched 8-week windows. AI agents keep popular terms in circulation, but their semantic drift is about twice that of human platforms and their neighborhood stability is 20-30% lower, with weaker long-range temporal dependence. The authors conclude that lexical diffusion alone does not guarantee diachronic semantic consolidation.
Chih-Han Lin, Yu-Yun Chang, Po-Ya Angela Wang et al.
In briefFractional Is Better: Learnable Derivative Orders for Neural Operators
SummaryThis paper studies how to improve neural operators, which learn mappings between function spaces to solve PDEs, by adding derivative features aligned with the PDE's structure to the input. The authors prove that such derivative features improve approximation bounds and find that the optimal derivative order is strictly less than the PDE order, due to a bias-variance tradeoff in spectral space. They introduce ∂-NO, an architecture-agnostic augmentation that learns fractional derivative features, and report consistent accuracy gains across benchmarks and architectures.
In briefDo AI Agents Gossip? Evaluative Talk in Human and AI-Agent Online Communities
SummaryThis study compares gossip, meaning evaluative talk about absent third parties, on Moltbook, a forum populated by AI agents, and PTT, a large human-operated bulletin board in Taiwan. Gossip appears on both, but it is 2.6 to 6 times less prevalent on Moltbook, depending on the denominator used. Challenges to gossip occur in 65.1% of human threads but only 21.3% of agent threads, suggesting AI-agent communities reproduce the informational side of gossip while its relational functions remain unresolved.
In briefBeyond Expert Rankings: Contrast Calibration for Mixed-Precision MoE Quantization
SummaryThis paper addresses storage costs in mixture-of-experts language models by proposing NatScale, a framework that assigns higher bit-widths to sensitive experts and lower ones to resilient experts. It calibrates gradient-based expert weights with a power transformation toward an information-theoretic contrast target while preserving their rankings, and combines output error estimates with a prior based on weight spectra, variance and precision decay. Across three models and nine settings, NatScale achieves the lowest KL divergence and highest top-token agreement against five baselines, and the lowest perplexity in seven settings.
In briefReferFlow: Grounding Multi-Step Visual References in Complex Expressions
SummaryThe paper argues that referring expression comprehension benchmarks mostly test only the final target, missing the intermediate visual references in complex queries. It introduces trajectory-level REC, an ordered sequence of grounding steps, and builds ReferFlow-Bench with four trajectory types: Refinement, Shift, Reset, and Mixture. It proposes ReferFlow, which uses stateful region latent reasoning, trajectory-aware parallel coordinate decoding, and Grounding-Aware GRPO training. The authors report that ReferFlow achieves leading performance on general REC benchmarks and ReferFlow-Bench.
In briefHuman Demonstrations as Post-Training Data for Any Robot: Retargeting Objects Instead of Hands
SummaryThis paper proposes using phone-recorded human demonstrations as post-training data by keeping only the object's pose, contact intervals and a provenance-tagged 3D reconstruction, not the hand's motion. Any robot, described by a URDF and a short file, is solved against this record and scored in simulation, giving each robot-episode pair a pass, a named-mode failure, or a refusal with its reason before motion. The analysed sample covers 42 captures, 8 robots and 157 pairs, and the authors note that no policy has yet been trained on the record.
In briefLifted State Policy Optimization: Self-Correction Built Into the Generator
SummaryThis paper proposes Lifted State Policy Optimization (LSPO), which attaches a learned continuous coordinate to each textual answer to form a lifted state that an energy function scores and orders. The method trains on energy descent net of edit and step costs and internalizes selected low-energy answers into the direct generator. On Qwen3.5-27B, the single-pass policy reaches 71.8% aggregate accuracy across math, science, code, and logic benchmarks, matching test-time reflection (71.9%) with 28% lower latency and 25.8% fewer tokens.
In briefNeuroDETR: A Multiscale Token-Based Method for Detecting Brain-Signal Events in Continuous EEG
SummaryThis paper proposes NeuroDETR, a framework that detects event classes and their onsets and offsets directly from continuous EEG recordings without predefined event boundaries. It combines multiscale temporal codebooks, a spatial-pyramid vector quantization encoder, and an event-aware sparse attention module called FOCAL. The authors report state-of-the-art results on eight EEG datasets covering diverse paradigms.
In briefNon-asymptotic Convergence of Average-reward Q-learning with Options
SummaryThis paper studies the finite-sample convergence of Q-learning for average-reward reinforcement learning with options, which are temporally extended hierarchical actions. The authors show that both inter- and intra-option Bellman operators are multi-step span contractions under unichain assumptions. They propose two option Q-learning variants and prove mean-square convergence rates of roughly 1/iterations.
In briefDoc2Harness: Testing How Well LLMs Build the Support Systems for Coding Agents
SummaryThis paper introduces Doc2Harness, a benchmark that tests how well large language models can build the harness, the surrounding framework that runs a coding agent, rather than just completing tasks. Models receive an interface skeleton and a development document, then construct a complete harness in an isolated container, which is later run against test cases driven by a scripted mock model. Across nineteen frontier models and 328 test cases, the best pass rate was 53.05% and the weakest was 27.44%, with failures clustering around how cross-cutting functionality is split across modules.
In briefWhen Task-Aware Imputation Matters: Robustness Under Train-Test Mismatch
SummaryThe paper treats missing entries as variables in a task-conditioned optimization problem, offering a unified framework for handling missing data. It presents two variants: an amortized approach for data missing at both training and test time, and a latent optimization approach for data missing only in training. Evaluated on 98 tabular datasets, with robustness tested on 36, the study finds reconstruction error is a weak proxy for predictive utility. Task-aware imputation achieved win rates of up to 80% in challenging train-test mismatch scenarios.
Maria V. Salvador, Ana F. Sequeira, Jaime S. Cardoso
In briefInfiniSplat v2: Improving Feed-Forward 3D Gaussian Scene Reconstruction from Any Views
SummaryThe paper introduces InfiniSplat v2, a method that builds 3D scenes from input images in a single feed-forward pass so that novel views can be rendered, including under large viewpoint changes. It separates geometry and attribute prediction using a frozen geometry branch, aligns predicted Gaussians with ground-truth camera poses during training, and uses large-angle target view supervision. Across six datasets, the same trained model reaches state-of-the-art mean LPIPS and DISTS among evaluated methods in both pose-conditioned and pose-free settings. In the pose-conditioned setting, the paper reports reductions of 25.8% in LPIPS and 32.9% in DISTS relative to the strongest baselines.
In briefHow Transformer Training Order Affects Learning on Unbounded Inputs
SummaryThis paper studies how a shallow transformer trains on variable-length inputs, and whether its attention layer and output layer can be learned separately. The authors show that a one-step attention update followed by output-layer training fails on a star-free language recognition task. They prove that a three-phase gradient descent, updating the output layer, then attention, then fine-tuning the output layer, can learn the language. They also report that attention learning lagging output learning helps on other language recognition, regression, and in-context learning tasks.
In briefUnified Parsing and Figure Description for Documents: The Read-Parse-Describe Dataset and Method
SummaryThe paper introduces Read-Parse-Describe, a dataset in which models take a page image and output an ordered list of layout elements with bounding boxes, labels, reading order, text, and alternative descriptions for visual elements. It contains 2,402 documents, 18,163 pages, 125,218 layout elements, and 11,338 visual instances. The authors also propose a plug-and-play space-awareness module that adds grid-based spatial guidance to vision-language models, and experiments show existing systems struggle to jointly recover reading order and visual semantics while the enhanced model performs strongly.
In briefTreeFlex: Agents That Navigate Their Own Search Trees
SummaryTreeFlex is a framework where an agent stores each intermediate state as a summarized node in a search tree, and a single language model decides which node to resume from. The model is trained with reinforcement learning using a tree-aware variant of GRPO and process rewards. The paper reports 48.8% accuracy on a deep-research benchmark and 54.9% resolved on SWE-bench Verified, above the compared methods.
WhyLetting the agent's own policy choose tree navigation with joint tree-aware RL is novel, backed by comparisons on two benchmarks.
In briefDo one-step generators need such complex training? A simpler two-stage method.
SummaryDiffusion and flow matching models usually need several network evaluations at inference, which has motivated one-step generators that often rely on complex training. This paper proposes Priming Models, a two-stage framework that first trains on a reconstruction objective and then performs feature-space distributional post-training. The authors report that the reconstruction stage alone is enough to remove the difficulty of distributional training. They report improved compute-quality trade-offs on class-conditional image generation, text-to-image generation, and world modeling, including an FID of 1.19 on ImageNet 256×256.
WhyA two-stage reconstruction-then-distributional training framework shows useful novelty; ImageNet results at one-tenth compute reach near-SOTA FID across several generation tasks.
In briefArgument Graph Uncertainty: Measuring Model Confidence from Reasoning Logic
SummaryThis paper proposes a post-hoc method that estimates a language model's confidence on multiple-choice questions using only the logical structure of one chain-of-thought reasoning trace. The reasoning is split into propositional units, and an NLI model builds two argument graphs whose edge scores are aggregated into a Dirichlet posterior over answer options. On GPQA, MMLU-Pro and TruthfulQA, the method outperforms token-based and consistency-based baselines and complements existing approaches, without extra calls to the evaluated model.
In briefAnuvani-Diarization: Zero-Shot Speaker Diarization Across 22 Indian Languages
SummaryThis paper studies speaker diarization, which identifies who spoke when, across 22 Indian scheduled languages. It introduces Anuvani-Diarization, a self-supervised end-to-end segmentation model adapted to Indian conversation using uncertainty-aware silver supervision. On 1,164 recordings it achieves a duration-weighted diarization error rate of 16.78%, the lowest among systems tested and second only to the best published result. The authors also show that the benchmark's annotation convention shifts results and that conventions learned from other Indian transcriptions do not transfer.
In briefComparing deep-learning direction estimators for real-time ISAC beamforming.
SummaryThis paper proposes a unified framework to compare classical and deep-learning direction-of-arrival (DoA) estimators, including an ensemble design, within an ISAC beamforming pipeline. The estimators are paired with SVD-based beamforming and evaluated under identical conditions for accuracy, latency, weight-generation time, and effective spectral efficiency. The authors report that the Regression Convolutional Neural Network (RCNN) offers a favorable balance of accuracy, efficiency, and beamforming performance, making it a promising candidate for real-time deployment.
In briefDynamic detection of backdoors in multimodal LLMs by tracking how internal signals change during decoding
SummaryThis paper proposes EVO-Detect, a framework for detecting backdoor attacks in Multimodal Large Language Models. Instead of relying on static snapshots, it monitors how internal signals evolve during early decoding steps through an Attention Flow channel and a Logits Flow channel. Extensive evaluations reportedly show it substantially outperforms baselines such as BYE and TCAP across diverse attack families.
In briefTest-time adaptation method (TEMPO) that helps robot walking policies adjust to changing conditions during deployment.
SummaryThe paper addresses reinforcement-learning locomotion policies for robots that degrade when conditions shift, such as rough terrain, payload changes, or sensing problems, and where retraining is impractical. It studies test-time adaptation, updating a pretrained policy online in a self-supervised way without rewards. The proposed TEMPO method adds an entropy regularizer on the critic and a score-based memory with importance sampling to stabilize updates under severe shift. Experiments on Go1, Go2, and G1 in MuJoCo Playground and on a physical Unitree Go2 show consistent gains, up to 80% in a single-device setting.
Motasem Alfarra, Pietro Mazzaglia, Markus Peschl et al.
