In briefSimplex Diffusion Models: a generative model that denoises discrete data on the probability simplex
SummaryThis paper proposes Simplex Diffusion Models (SDMs), which lift the diffusion process onto the probability simplex so uncertainty over categories is kept across denoising steps, reducing information collapse in discrete diffusion. The models train with a cross-entropy loss and use a DDIM-like sampler with tunable stochasticity. The authors report results competitive with or better than discrete diffusion baselines on OpenWebText, TinyGSM, and GSM8K.
Justin Deschenaux, Alexandre Galashov, Andrew Campbell et al.
In briefBaRe-Mem: A Bayesian memory that tracks advisor reliability for robust agent consultation
SummaryIn multi-agent systems, advisors' abilities vary by task and misleading advice can make consultation worse than reasoning alone. BaRe-Mem is an online Bayesian reliability memory that estimates advisor reliability from the central model's internal beliefs and updates it from past interactions, using these estimates to weight advice and decide between consulting and reasoning autonomously. Across nine benchmarks and six central models it was more robust to misleading advisors than debate and majority voting, and on harder tasks it stayed above autonomous reasoning at all tested misleading levels. The authors also extend the mechanism to worker allocation, where on MuSiQue it improved task completion over routing by historical success counts and identified capable workers earlier.
In briefRareDx: Controlled knowledge integration and knowledge-graph-based policy optimization for rare-disease diagnosis
SummaryThis paper presents RareDx, a system for rare-disease diagnosis, a long-tail reasoning problem where language models often favor common conditions. It converts heterogeneous patient records into a ranked-diagnosis task and trains models with RareDx-KGPO, a knowledge-graph-grounded policy optimization method whose reward combines graded relevance, ontology proximity, biomedical similarity, and phenotype consistency. The authors report that the RareDx system centered on Qwen3.5-9B reaches 38.34 macro Hit@10 across eight benchmarks, 1.60 points above GPT-5.5 under the archived protocol.
WhyCombines knowledge-graph-grounded policy optimization with controlled routing for rare-disease diagnosis, with eight-benchmark comparisons and ablations.
In briefGraph world models for planning constrained epidemic policies
SummaryThis paper presents EpiMind, a framework for planning epidemic policies across regions under resource constraints. It uses a graph-structured recurrent state-space model to generate joint, policy-conditioned rollouts, and graph-temporal ADMM to optimize regional interventions while enforcing shared-resource feasibility. The authors report a 29% reduction in admission RMSE compared with graph-free dynamics modeling, and plans within 1-5% of the best feasible constant policy.
In briefLess sycophancy, stronger refusal? AI safety lessons from interpretability research
SummaryThis paper tests whether reducing language model sycophancy, meaning excessive eagerness to please users, can strengthen refusal of harmful requests. The authors use compensatory feature injection during fine-tuning, with sparse autoencoders identifying a sycophancy feature across three Qwen3.5 base models. Positive injection reduced learned sycophancy, but did not consistently improve direct refusals. Under user pressure, checkpoints trained with positive injection recovered part of the refusal loss caused by sycophantic fine-tuning.
In briefQuestion-specific knowledge graphs for more efficient visual reasoning
SummaryThe paper proposes VisKG, a reinforcement learning framework that trains models to convert images into question-specific knowledge graphs, filtering out visual details unrelated to the task while keeping the entity-relation structure needed for reasoning. Experiments on science, mathematics, and general visual understanding benchmarks report performance comparable to or better than baselines, using fewer tokens than caption-based representations. Training with GDPO improves average accuracy by 2% over GRPO-trained models.
Ting-Chih Chen, Emile van Krieken, Shujian Yu et al.
In briefLearning Robustness Settings via Bilevel Optimization Instead of Manual Tuning
SummaryThis paper proposes a distributionally robust learning framework that learns the parameters of the robustness mechanism from held-out data rather than extensive tuning. It uses bilevel optimization with upper- and lower-level minimax problems, yielding two versions for settings with and without group labels. The authors report that the approach matches the generalization guarantees of exhaustive grid search while being more computationally efficient. Experiments under simultaneous intra-group and inter-group distribution shifts show its efficacy and scalability.
In briefContinuous Context Management: Compressing Agent Memory at Every Turn
SummaryThis paper studies Continuous Context Management (CCM), in which language model agents update a compressed memory at every turn instead of keeping the full interaction history until a threshold is reached. Without fine-tuning, CCM substantially reduces input usage and prompt size, though it lowers task success for most models on TerminalBench-2. The authors apply GRPO with privileged full-history distillation to open-weight models, reporting gains on WebShop and a modest improvement over GRPO on Endless Terminals.
In briefLearning to reason about persistent object states so video segmentation keeps track of identities
SummaryVideo segmentation models track objects by carrying instance information across frames, but unreliable updates during long occlusions, reappearances, or similar-object interactions can cause identities to drift. This paper presents POSReasoner, a trainable, plug-and-play framework that decides when an observation should change an object's state, using Propose-Verify reasoning and a learned gate. It reports consistent gains over strong baselines on long-term video object segmentation and video instance segmentation benchmarks, with the largest improvements under occlusion and reappearance.
In briefOptimal Networks for Agentic Information Aggregation
SummaryThis paper studies how a network of learning agents can combine information, building on the Kearns, Roth, and Ryu (SODA 2026) model. Each agent sees some features and its parents' predictions, fits a linear predictor, and passes its prediction forward. The authors analyze two designer settings and show that exact aggregation needs two parents per agent, with depth and agent count bounds that are optimal up to constant factors.
WhyGives matching tight bounds on depth and agent count for exact aggregation in the networked learning model, separating adaptive from oblivious designers; rigorous but mainly theoretical impact.
MohammadHossein Bateni, Zahra Hadizadeh, MohammadTaghi Hajiaghayi et al.
In briefLook Before You Judge: Training-Free Region Mining for Explainable Deepfake Detection
SummaryMultimodal large language models can explain deepfake verdicts in words, but these explanations may not rely on actual visual evidence. The paper proposes Look Before You Judge, a training-free framework that finds candidate evidence regions by contrasting attention on an original image and its blurred version, inspects each region, and combines local findings with global context. It requires no masks, external forensic models, or parameter updates. On TriDF and MMTD-Set across five open-source MLLMs, it improves accuracy by up to 12.8%, reduces CHAIR by up to 33.4%, and lowers hallucination rate by up to 21.3%.
Chia-Ling Chen, Yu-Ting Ta, Jian-Yu Jiang-Lin et al.
In briefARISE: Adapting Agent Reinforcement Learning to Shifting Capability Gaps
SummaryThis paper introduces ARISE, a reinforcement learning framework for long-horizon agents. It uses rollout results to continually adjust evaluation criteria, exploration guidance, and training priorities as the agent's weaknesses change. The authors report experiments on the SkillsBench and Terminal-Bench benchmarks showing improved overall task performance and training efficiency.
In briefAutoRef: Automatically Optimizing the Control Program for Multi-Reference Image Generation Agents
SummaryThis paper proposes AutoRef, which uses a coding agent to iteratively rewrite a harness, the executable program that controls how an image generation agent handles reference images, while keeping the generation and reasoning models frozen. On four-reference held-out tasks of the MultiBanana benchmark, the open-weight FLUX.2 [klein] 4B model improves from 5.72 to 7.37, which the authors report matches or exceeds proprietary models such as Nano Banana Pro and GPT-Image-1.5. The optimized harness also improves results when the generator, reference count, benchmark, evaluator, or reasoning model changes, without re-optimization.
In briefLet the Neurons Die: Attacks Exploiting ReLU Dying-Neuron Degradation
SummaryThe paper examines how ReLU networks can suffer dying neurons, whose zero outputs block gradients. It presents three training-time availability attacks based on data ordering and poisoning that push ReLU units toward negative pre-activations without changing training labels or parameters directly. On an MNIST fully connected network, reordering 100 of 60,000 examples lowers five-epoch test accuracy from 96% to 95%, while adding 200 poisoned single-class samples lowers it to about 86-88%.
In briefGeoGAE: Scalable Graph Autoencoder Using Hyperball Cloud Representations
SummaryThe paper represents a graph as a cloud of hyperballs, which yields a specific, typically unique node ordering. Building on this, it proposes GeoGAE, an autoencoder in which a Transformer encoder maps the hyperball cloud to a graph-level embedding and a Transformer decoder reconstructs the graph from that embedding. The approach aims to capture both global graph structure and local relational patterns. Experiments on multiple graph datasets from various domains show it can encode and reconstruct graphs from their embeddings.
Rados{\l}aw Nowak, Anna Bielawska, Bogusz Stefa\'nczyk et al.
In briefDEVOTE: Deep Epistemic Value Functions for Optimistic Exploration in Reinforcement Learning
SummaryThe paper empirically studies how epistemic uncertainty is represented, propagated, and optimized in deep epistemic value functions, identifying distinct failure modes along each axis. Based on these findings, it proposes DEVOTE, a model-free reinforcement learning algorithm that controls how uncertainty generalizes beyond observed data, stabilizes its temporal propagation, and adapts to the resulting non-stationary exploration objective. The authors report that DEVOTE reaches novel states more effectively and achieves higher task return than strong model-free and model-based exploration baselines on reward-free exploration and continuous-control tasks.
Leander Diaz-Bone, Marco Bagatella, Jonas H\"ubotter et al.
In briefChunk-level sparse autoencoders learn more reliable semantic features for language models
SummarySparse autoencoders help interpret and steer language models, but faithful reconstruction alone does not ensure meaningful concepts. The paper introduces chunk-level SAEs that encode mean-pooled activations over contiguous token spans, with Mean-Chunk, Cross-Chunk, and Joint-Chunk variants. The authors report that these models learn more reliable high-level semantic features and improve downstream tasks such as retrieval, reasoning detection, and steering.
In briefReward-Aligned Reweighting for On-Policy Distillation: Aligning Teacher Guidance with Task Outcomes
SummaryThe paper notes that standard on-policy distillation weights every token of a student's output equally, ignoring how much each decision contributes to the final task result. It proposes R²-OPD, which continuously reallocates teacher supervision based on outcome agreement and the size of teacher-student disagreement, giving more influence to corrections aligned with reward while keeping dense feedback. Across seven mathematical reasoning benchmarks, it achieves the highest average accuracy among compared methods and outperforms standard OPD on all seven, with a 1.6-point average gain in a code generation extension.
In briefMechBench: Can AI Scientific Agents Discover Mechanisms Beyond Phenomenal Laws?
SummaryMechBench is a benchmark that separates recovering observable phenomenal laws from discovering the underlying mechanisms that produce them. Agents receive only observational data and scientific context, and are tested with mechanism probes on internal consequences not inferable from the phenomenal law alone. Results show a large gap: Codex with GPT-5.6-sol reaches 35.00% phenomenal-law accuracy but only 13.75% mechanism accuracy on the Core-set.
In briefOne Proposal for All Margins: Zero-Shot Learned Sampling for Binary Matrices with Fixed Row and Column Sums
SummaryThe paper studies binary matrices with fixed row and column sums, focusing on counting and uniformly sampling them via sequential importance sampling. The authors show that the ideal proposal distribution corresponds to a GFlowNet policy and introduce MarginFlow, which learns a proposal shared across margins. Evaluated zero-shot on 1190 held-out margins, it matches or beats the best of 31 hand-designed configurations on 1187 of them.
WhyLinks ideal SIS proposals to GFlowNet policies and amortizes them across margins via a set transformer; 1190 zero-shot evaluations are thorough, but impact is confined to a niche area.
In briefImproving Generative Model Self-Training with Geometrically Modified Outputs
SummaryThis paper studies self-training of generative models, where a model is improved using its own outputs, which can cause degradation such as model collapse. The authors propose Geometrically Modified Outputs (GMOs), which reweight the singular values of the generator's input-output Jacobian to strengthen its leading singular directions. This creates a stronger negative signal for self-training. Experiments on several one-step generative models show GMOs improve negative-guidance methods such as Neon and SIMS compared with standard outputs.
In briefReVA: A scene-centric benchmark for question answering on drone videos
SummaryThis paper introduces ReVA, a dataset for question answering on remote sensing drone videos, designed to test multimodal large language models on spatiotemporal, scene-centric, and reasoning abilities. It contains 2,438 videos from 18 cities worldwide (580K frames) and 22K question-answer pairs across 11 tasks. The pairs were created with a semi-automatic pipeline using language models plus human verification. Evaluating 23 proprietary and open-source video LLMs on ReVA reveals fundamental limitations of current models.
In briefAn RL View of On-Policy Distillation: Least-Square Policy Distillation for Efficient LLM Reasoning
SummaryThis paper studies on-policy distillation (OPD) through reinforcement learning, linking its reverse-KL objective to KL-regularized policy optimization. It introduces Least-Square Policy Distillation (LSPD), which adds optimistic exploration and off-policy data reuse from value-based RL. LSPD outperforms existing distillation baselines across six math reasoning benchmarks, with an average gain of +1.59 points in Avg@16. Its fully off-policy variant matches vanilla OPD using only the first 25% of rollout batches.
In briefSolveEdit: A Benchmark for Goal-Driven Visual Problem Solving in Generative Models
SummaryThe paper introduces SolveEdit, a benchmark that tests whether generative models can solve visual problems through scene transformation, given an image and a goal. It contains 2,728 cases and scores results with atomic transition contracts that measure completion and unintended changes. The strongest evaluated model reaches only 57.0% SolveScore. The authors also propose SolveEdit-PLAN, a two-stage visual planner that improves SolveScore by 9.1 points on average across three generators without modifying the editor.
In briefStructured latent modeling to separate shared and unique information in multimodal prediction
SummaryThis paper proposes a framework for separating shared task-relevant information, modality-specific predictive information, and task-irrelevant dependence within learned continuous multimodal representations. It combines intermediate-layer contrastive or masked objectives with per-source invertible normalizing flows and a supervised low-rank latent variable model. The authors report improved predictive performance across diverse multimodal benchmarks.
In briefSRHarness: A runtime framework for agentic symbolic regression
SummaryThe paper introduces SRHarness, a domain-specific runtime framework for agentic symbolic regression that organizes long search trajectories through composable scientific actions, persistent scientific state, and trajectory lifecycle management. On LLM-SRBench, under matched LLM backbones, it improves both numerical generalization and symbolic recovery. For example, with DeepSeek-v4-flash-0731 it reaches 93.69% symbolic accuracy on LSR-Transform, versus 62.16% for SR-Scientist. The authors argue that effective agentic symbolic regression depends on structured runtime support, not only on models or tools.
In briefPhysics-guided diffusion model for generating and diagnosing rare failures in hydrogen production
SummaryThis paper addresses the scarcity of process data for rare abnormal conditions, such as fouling, catalyst deterioration, and thermal runaway, in the water-gas shift reaction used for hydrogen production. It proposes a conditional diffusion model constrained by reaction governing laws to generate physically consistent rare-event trajectories, which balance the training dataset. A hazard score is also introduced to assess risk severity from operating trajectories. Simulations show the method outperforms purely data-driven models in synthetic data quality and rare-event diagnosis.
Md Abrar Rafid Siddique, Bibek Aryal, and Qiugang Lu
In briefSprout: a video-understanding agent that builds memory as it answers questions
SummaryThis paper introduces Sprout, an agentic framework for long video understanding that builds a temporal memory tree while answering questions rather than in a separate offline stage. The agent watches the video segment by segment at low frame rate, records each segment as a text node, and revisits key intervals at higher frame rate to add detail. The memory persists across questions, and the authors report that replacing video inputs with text memory lowers context usage while keeping accuracy competitive with representative offline memory methods on benchmarks across three models.
In briefFrom Scores to Samples: Elastic Forcing for Faster Autoregressive Video Generation
SummaryThis paper proposes a training method for few-step autoregressive video generation that avoids the bidirectional diffusion teacher and online fake-score model used in Distribution Matching Distillation. Instead, it learns the rollout distribution directly from reference videos by minimizing maximum mean discrepancy in frozen self-supervised video representation spaces. Using the same architecture as Self-Forcing, the 1.3B model raises the VBench Total score from 83.80 to 84.64 while retaining 17 FPS. Removing auxiliary score models also enables 14B post-training on eight H200 GPUs.
In briefAHMAD: Adaptive Hybrid Multi-task Vision Learning with Assisted Distillation for Keypoint Detection
SummaryThis paper proposes AHMAD, a generalist multitask vision framework that uses a shared encoder-decoder with lightweight task-specific projectors for five tasks: semantic segmentation, instance segmentation, depth estimation, keypoint detection, and object detection. The authors report state-of-the-art results of 53.1 PQ for COCO-val panoptic segmentation and 66.5 mIoU for semantic segmentation. To reduce the cost of top-down keypoint detection, which needs multiple forward passes, they introduce a knowledge distillation method that processes the whole image in a single forward pass.
Mohammad Mahdi, Nedyalko Prisadnikov, Yuqian Fu et al.
In briefTracking Spatial Evidence and Role Binding in Vision-Language Models
SummaryThis paper studies how vision-language models internally reason about the relative positions of objects. It shows that high accuracy can hide inconsistencies when objects swap positions or their query roles are reversed. Using activation patching, the authors find a staged progression from early-layer location representations to intermediate query-object representations and late-layer answers. They also identify a role-related direction that, when steered, improves accuracy and consistency in most settings without retraining.
In briefUniversal approximation of distribution-to-distribution maps via pushforwards
SummaryThis paper studies whether continuous operators between spaces of probability measures can be approximated by pushforward models whose transformation depends on the input distribution. The authors show that universal approximation fails when atomic inputs are allowed, because some operators that split or redistribute atomic mass cannot be approximated. They introduce a uniform level set condition, which holds for compact families of absolutely continuous measures, and prove that under this condition any continuous operator with finite p-th moment outputs can be uniformly approximated in p-Wasserstein distance. The results extend to measure-theoretic transformers and to continuously varying source measures.
WhyIntroduces a uniform level set condition and proves uniform approximation of measure operators by measure-dependent pushforwards, with a counterexample for atomic inputs.
In briefTopoEP: topology-aware load balancing so GPUs hosting busy experts stop slowing MoE training
SummaryDynamic routing in large expert-parallel Mixture-of-Experts training causes heavy load imbalance, so GPUs hosting hot experts become stragglers. The paper introduces TopoEP, a GPU-native, topology-aware load-balancing system that decides expert replication and token rerouting at each MoE layer and microbatch without host synchronization. On a 32-GPU NVIDIA H800 cluster integrated with Megatron-LM, it improves end-to-end training throughput by 6.2% to 11.4% across three representative MoE models.
In briefAnalog computing revisited: a fully analog, minimal ultrasonic damage detector for material-integrated structural health monitoring
SummaryThis paper proposes a fully analog damage detection system for ultrasonic testing (UT) of structures. It approximates the Hilbert transform with all-pass filter networks to extract the signal envelope, applies simple analog arithmetic for feature extraction, and uses an analog neural network for damage classification and regression. The authors expect the full system to use fewer than 100 transistors and test it on PZT sensor signals from steel plates with circular defects.
In briefHow language models recover meaning from corrupted inputs
SummaryThis paper studies why language models often still give correct answers when their inputs contain deletions, replacements, or misspellings, a behavior the authors call context restoration. They analyze attention-only transformers and five pretrained LLMs from 1B to 32B parameters on arithmetic, reading comprehension, and multiple-choice tasks. The restoration follows a two-phase internal process, and failures can be predicted from hidden states with a mean ROC-AUC of 0.78.
In briefHandwriting recognition relies on high-pixel-variance patterns
SummaryThis paper studies why pixel reconstruction outperforms contrastive methods for self-supervised pretraining in handwritten text recognition, reversing the trend seen in natural-image classification. The authors argue that discriminative signal concentrates in high-variance pixel directions, so objectives preserving that content transfer best. Six self-supervised methods were compared under matched settings on six handwriting benchmarks across five languages, and pixel-grounded methods achieved the lowest error overall.|
In briefWhy PRM-guided denoising underperforms in discrete diffusion reasoning
SummaryThis paper studies whether process reward models (PRMs) can guide step-by-step denoising in discrete diffusion language models to improve reasoning at test time. Under equal forward-pass budgets, deterministic PRM guidance loses to independent sampling with an outcome reward model (ORM) reranker, for example 65.18% versus 75.13% on GSM8K. The authors trace this to weak PRM signals that prune good candidates and to the PRM being a poor final judge. They suggest keeping correct partial solutions alive and using a verifier trained on final states for selection.
In briefRethinking Causal Action Tokenization with Conditional Annealing in Flow Matching
SummaryThe paper argues that existing action tokenizers treat tokenization as compression, producing representations misaligned with autoregressive backbones. The authors propose CATok, which extracts action tokens by progressively annealing a flow-matching process, yielding a coarse-to-fine causal token space. Evaluations on three simulation benchmarks and real-world manipulation tasks report gains in reconstruction-compression tradeoff, inference efficiency, task success rate, and training efficiency over existing methods.
In briefConvergence analysis of mirror-descent Soft Actor-Critic for continuous-action reinforcement learning
SummaryThis paper studies convergence guarantees for Soft Actor-Critic (SAC), an entropy-regularised reinforcement learning method for continuous actions. It considers the case where the target policy comes from policy mirror descent and compares this with the classical Gibbs target. The authors give conditions for strong convexity and smoothness of the actor objective and prove an O(N^{-1/5}) best-iterate finite-time rate up to approximation errors. They also show that the mirror-descent step size controls target drift, whereas the Gibbs bound has a non-vanishing tracking term.
In briefTetra: Serving Leech-Lattice Quantized LLMs at 2.7 Bits per Parameter
SummaryThis paper presents Tetra, a new codebook for quantizing large language model weights on the Leech lattice, designed to avoid a lookup table too large to store. It uses a 64-state Golay-code trellis and one shared table, letting the kernel read about 2.148 bits per weight. Tetra’s Qwen3 4B, 8B and 14B models average roughly 2.7 bits per parameter, with MMLU scores below 4-bit AWQ and above llama.cpp’s IQ2_XXS. All measurements were taken on one NVIDIA L40S GPU.
In briefW2Rep: Learning visual features by watching how scenes change
SummaryW2Rep is a masked feature-prediction framework that lets an independently encoded source image take part in predicting content at the same or a different moment in a video. The predictor is conditioned on visible video context, the queried location, and the signed time gap between source and target. The authors report that it improves frozen and fine-tuned recognition across model scales and downstream tasks under their comparison protocol, with joint video encoding giving further gains.
In briefA.D.A.M.O.: A visual-symbolic framework letting language drive virtual humans' perception, reasoning, and actions
SummaryThis paper presents A.D.A.M.O., a framework that uses a pretrained vision-language model with tool calling to unify perception, reasoning, and action in one control loop for virtual humans. It maintains a dual world model combining egocentric visual input with synchronized symbolic state to carry out tasks from natural language prompts. The authors also introduce a controlled task suite, organized by a complexity taxonomy, for diagnostic evaluation. Experiments in controlled scenes show semantic labeling strongly affects task completion and failure modes, while reasoning errors remain comparatively rare.
Alessandro Emmanuel Pecora, Stefano Calzolari, Francesco Strada et al.
In briefCLIMB: A multi-turn benchmark for diagnosing co-occurring clinical conditions
SummaryThe paper introduces CLIMB, a benchmark where a doctor model interviews a simulated patient to identify a set of co-occurring clinical conditions. Cases are synthesized from clinical decision algorithms and diagnostic datasets. Across six frontier and open models, none recovered the exact condition set in more than 10% of interactive cases, and performance declined when conditions co-occurred.
Yusuf Kesmen, Aniruddha Mukherjee, Yena Chang et al.
In briefAraDynFact: A dynamic framework for evaluating factual knowledge of Arabic in LLMs
SummaryThis paper introduces AraDynFact, a dynamic evaluation framework for measuring the factual Arabic knowledge of large language models. It automatically extracts facts from Arabic Wikipedia and generates answerable questions, without relying on heavy human intervention. The authors apply it to several models and report a high degree of correlation with existing hand-crafted Arabic benchmarks.
Ignacio Iacobacci, Faroq Altam, Zhaozhi Qian et al.
In briefHow Far Are We from Removing the Visual Encoder? Scaling Laws for Encoder-Free Multimodal Pretraining
SummaryThis paper compares scaling laws for multimodal language models with and without a pretrained visual encoder. Removing the encoder shifts the compute-optimal model size larger for the multimodal objective, while the text objective is nearly unchanged. Encoder-free models underperform at small scales but are predicted to catch up at around 10^22 FLOPs, with the language model learning to handle visual processing through vision-specific adaptation.
In briefAutoBCI: Forecast-guided automatic design of EEG decoding networks for brain-computer interfaces
SummaryThis paper presents AutoBCI, an agent-based framework in which a Designer Agent and a Forecaster Agent work together to discover EEG decoding architectures across tasks such as emotion recognition, motor imagery, and sleep staging. The Forecaster predicts full-budget validation performance from early training curves to select promising candidates. Evaluated on 14 EEG datasets with six LLMs against ten baselines, the architecture found with Claude Opus 5.5 reaches 64.16% average test balanced accuracy, versus 63.87% for the strongest baseline, REVE.
In briefManifold-stable flow matching: a method keeping generated samples on valid geometric structures
SummaryThe paper introduces manifold-stable flow matching (MSFM), which can start from an arbitrary ambient prior, not necessarily supported on the data manifold. It uses contraction theory to combine learned tangential transport with prescribed normal contraction, using analytical projectors for known manifolds and PCA-estimated local affine proxies for unknown ones. Reported results include a terminal off-manifold error of order 10^-6 on an ellipse experiment, and higher success rates and lower rotation-manifold deviation in Push-T and Robomimic Square robotic tasks.
WhyCombines contraction theory with flow matching to enforce manifold invariance; gains on Push-T and Robomimic show value, though experimental scope is limited.
In briefChoosing which images to label by linking a model's internal features to its prediction errors
SummaryThis paper studies how to pick images to label when annotation budgets are limited. The authors link the target model's own internal features to its prediction errors and their expected effect on performance, then select images for labeling and retraining without using labels for candidate images. Evaluated with an object detector on two datasets, the method outperformed selection based only on feature rarity in all four evaluation settings.
In briefDiMoP: diffusion-based motion learning with frame-level pseudo-labels for skeleton action recognition
SummaryThis paper presents DiMoP, a method for recognizing actions from skeleton data. It masks joints, adds progressive noise, and denoises them conditioned on visible joints, learning the distribution of joint motion rather than deterministic coordinates. A pseudo-frame classifier adds discriminative learning using sequence-consistent pseudo-labels without manual annotation. The authors report state-of-the-art results on NTU RGB+D 60/120 and PKUMMD, including a 1.1 point gain on NTU RGB+D 120 cross-subject.
Shanaka Ramesh Gunasekara, Wanqing Li, Nikalal Kaldera et al.
In briefJust Initialize: A Training-Free Starting-Point Method for Large Routing Problems
SummaryThis paper introduces Just Initialize, a training-free, solver-agnostic component that generates good starting solutions for large-scale routing problems. It compresses a large routing instance into a compact surrogate space, optimizes the global routing structure there, and recovers it as a starting point for the original problem. The authors report experiments on TSP, CVRP, VRPTW and PCTSP with 1K to 100K nodes, finding solutions comparable to or better than state-of-the-art methods at much lower cost, with an average speedup of about 70 times.
In briefSOLO: Training Billion-Parameter Language Models Locally with a Shared Output Readout
SummaryThis paper proposes SOLO (Shared-Output LOcal learning), which replaces each module's private readout with a shared, read-only copy of the final module's readout, so that every module receives information from the whole network without passing gradients between modules. On Transformers with 340M to 2B parameters pretrained on 15B tokens, SOLO stays within one point of backpropagation in average zero-shot accuracy. Because it avoids update locking, it allows larger micro-batches, reaching up to 1.44x the best throughput of pipeline backpropagation on the same partition.
WhyReplacing private readouts with a shared output copy approaches backprop at 340M–2B scale, with up to 1.44x throughput and ablations supporting the claim.
In briefRevision, Not Restart: Revisable Visual Plans for Robot Control
SummaryThis paper proposes Revisable Temporal Planning (RTP), a method for robots that keeps a predicted visual future and revises it after execution feedback rather than replanning from scratch. A learned revision bridge resumes from an intermediate state saved during visual generation and adapts the continuation to current observations. On RoboMME and RMBench, RTP reports task-averaged success rates of 48.6% and 84.8%, respectively.
In briefA geometric framework for analyzing shape of axillary lymph nodes in breast cancer
SummaryThis paper presents a framework that combines topology-aware surface processing with multi-resolution spherical harmonic analysis of axillary lymph nodes segmented from CT scans. The processing yields valid genus-0 surfaces, described by 20 predefined geometric feature families. A family-specific mixed-resolution model reached an AUC of 0.918, versus 0.884 for the PyRadiomics Shape14 baseline. Independent validation was performed in a multicenter lymph node cohort and on the LIDC-IDRI lung-nodule dataset.
In briefBuilding Transformation Layers for Riemannian Neural Networks
SummaryThis paper proposes a framework for constructing fully connected and convolutional layers over computationally tractable Riemannian spaces, generalizing these Euclidean layers to non-Euclidean geometries. The framework includes several previous fully connected layers across different geometries as special cases. It is instantiated on ten representative manifolds, including three hyperbolic models, five symmetric positive definite geometries, and two Grassmannian perspectives. Experiments on different manifolds demonstrate the effectiveness and applicability of the approach.
In briefReSPO: Reshaped Sequence Policy Optimization to Fix Gradient Starvation in Off-Policy Learning
SummaryThe paper studies gradient starvation in off-policy reinforcement learning, where reusing rollouts across updates lets clipped policy optimization suppress under-generated positive responses and over-weight severely over-generated negative ones. It proposes ReSPO, which replaces clipping with a smooth two-branch sequence-level kernel derived from an alpha-divergence variational objective and an exponential variance-control tilt. Experiments on dense and MoE Qwen3 models report faster early optimization, better final training scores, and higher held-out benchmark performance.
In briefMulti-task learning of conditional mean operators for dynamics and uncertainty quantification
SummaryThis paper studies learning conditional mean operators (CMOs) jointly across related conditional distributions. It proposes MTL-CMO, which learns shared finite-dimensional function spaces and task-specific linear operators, and T-CMO, which reuses those shared spaces to estimate a new distribution's operator in closed form. The authors report that shared spaces improve uncertainty quantification and yield compact, physically meaningful representations of Langevin and plasma dynamics that support parameter identification.
Sami Chemlal, Thibaut Germain, R\'emi Flamary et al.
In briefCan LLMs act as general asynchronous agents that keep taking in new input while thinking?
SummaryMost LLM agents work in a strict sequence of reading, thinking, and replying or calling tools. Real applications such as voice assistants, robots, and monitoring systems must also receive new inputs while busy with other work. The paper proposes an asynchronous LLM framework in which users or agents can define inference coroutines with overlapping memory states. The authors report that Qwen 3.x models can run asynchronously for streaming video understanding, video games, and monitoring without task-specific training.
George Yakushev, Denis Mazur, Vladimir Bartenev et al.
In briefFrontier Learning: Training Language Model Reasoning at the Edge of Ability
SummaryThis paper proposes frontier learning, an open-ended post-training method for improving the reasoning of large language models. It uses procedural generators online during training to continually produce informative problems, and a regret signal to prioritize and explore frontier difficulty levels. The authors report that across several reasoning tasks and model families, the approach achieves higher relative gains than fixed-pool baselines.
Robin Faro, Shyam Sundhar Ramesh, Ilija Bogunovic et al.
In briefSemantic Prefix Oracles for LLM Decoding: Contracts and Differential Validation
SummaryThe paper introduces semantic grammar specifications, a declarative way to add scope, typing, and declaration constraints to context-free decoding for program generation. It prunes only those output prefixes whose semantic contradictions cannot be repaired, and gives sufficient conditions guaranteeing that every remaining branch has a realizable witness. Differential tests against ocamlc and cc show zero false prunes on all prefixes of 65 valid programs, and locate 25 of 30 invalid programs mid-stream versus 0 for a syntax-only oracle.
In briefInJePA: Intention-first latent planning with linked JEPA world models for visual navigation
SummaryThis paper proposes InJePA, a visual navigation method that proposes latent intentions before generating actions, rather than evaluating complete action candidates afterward. A goal-conditioned JEPA proposes latent intentions, while an action-conditioned JEPA predicts future latent states, both aligned in a shared latent space. The authors evaluate closed-loop image-goal navigation in Habitat MP3D and report effective multistep navigation, with applicability across different encoders.
In briefAutomated Wildlife Species Identification in Camera Trap Images
SummaryThis thesis proposes an end-to-end framework to identify species in low-contrast camera trap images, especially small animals. It combines a Swin-BiFPN backbone within a Faster RCNN detection network with a visual semantic extraction module based on the LLaVA v1.5 (13B) multimodal large language model. The module enables zero-shot detection and offers cues about animal behavior. Five multimodal LLMs were evaluated using traditional NLP metrics and LLM-based judges.
Nowshin Amin, Nafisa Tabassum Oyshi, Tahmid Abrar Zidan et al.
In briefConvex optimization gets cheaper when accuracy is costly: a randomized multilevel method
SummaryThe paper studies convex optimization when gradients can only be approximated, with computational cost growing as δ^(-γ) in accuracy δ. It introduces a randomized multilevel oracle that gives an unbiased estimator of the approximation, lowering the cost of one call to δ^(2-γ)σ^(-2). In the convex case, expected compute to reach loss ε is Θ(ε^(-γ)), better than Θ(ε^(-(γ+1))) at fixed accuracy, and under μ-strong convexity the exponent halves to γ/2.
WhyA randomized multilevel oracle improves convex optimization complexity from ε^{-(γ+1)} to ε^{-γ}, with strong-convexity bounds; rigorous theoretical gains.
In briefWhen Should Video Counts Change? Learning State Maintenance for Causal Counting
SummaryThe paper introduces StaMina (State Maintenance), a method that maintains video counting state through state-conditioned updates, distinguishing new observations, new objects, and completed events. It is evaluated on SVCBench under prefix replay (Full) and persistent streaming (Stream) settings. The 4B and 8B models reach 41.9/36.4 and 44.9/38.2 Gaussian Precision Accuracy, with the 8B model improving over Counting-SFT on the same queries.
In briefAdaptive Safety Filtering for Frozen Cruise Control Policies Using Conformal Residual Calibration
SummaryThe paper addresses frozen adaptive cruise control (ACC) policies that can violate constraints when deployment dynamics differ from training conditions. It proposes residual-aware conformal action filtering (RACF), which calibrates residuals of a fixed nominal predictor and converts their quantile into an operating margin for action projection, without retraining the policy. In a comparison over 2,400 controller-trial units, Adaptive RACF reached 94.3% episode safety, 19.9 percentage points above the nominal CBF-QP baseline, while reducing projection frequency.
In briefTraining-free framework that moves the best reasoning strategy among multi-agent language models
SummaryThis paper introduces SAGE (Self-Adapting Group of Experts), a training-free framework for multi-agent reasoning. It uses answer agreement, prefix consistency, and reciprocal peer review to pick a strategy donor, then transfers that donor's reasoning strategy to other agents while keeping their roles. Experiments across several agent backbones and reasoning benchmarks report higher average accuracy than the evaluated baselines.
Mohammad Atif Quamar, Nurbek Tastan, Karthik Nandakumar et al.
In briefGLAD: A Global-Local Adaptive Detector for Robust Speech Deepfake Detection
SummaryThis paper addresses weaknesses of self-supervised learning (SSL)-based speech deepfake detectors, which generalize poorly to unseen domains and often miss fine-grained signal artifacts. The authors propose GLAD, which uses a hierarchical global-local backbone to fuse global linguistic and acoustic features with local signal details, and a hierarchical adaptive gating mechanism that adjusts layer-wise focus per sample. A composite augmentation strategy, SaniBoost, is also introduced to reduce shortcut learning from environmental biases. The authors report that GLAD significantly outperforms state-of-the-art methods, especially on unseen domain cases.
Zelin Zhao, Guanjie Huang, Danny Hin Kwok Tsang et al.
In briefSpectral super-resolution of satellite images using a neural operator model
SummaryThis paper frames spectral super-resolution, which turns multispectral satellite images into hyperspectral-like images, as an operator learning problem. It proposes SSRON, a Deep Operator Network that learns mappings from downsampled spectra to continuous spectra. Trained on Sentinel-2A-like imagery to EMIT images, it outperforms baseline models on all metrics and predicts bands unseen during training. The authors also suggest it can estimate spectra at finer wavelength intervals than the native sensor.
In briefAwarenessBench: A Benchmark for Assessing Cognitive Abilities of Language Models
SummaryThis paper introduces AwarenessBench, a benchmark for assessing language models' cognitive abilities across four dimensions: metacognition, self-awareness, social awareness, and situational awareness. It covers 15 cognitive functions with 14,381 samples. Evaluating 18 state-of-the-art models, the authors find that all outperform random baselines, with more advanced models scoring higher. The best model surpasses human averages overall, but most fall short in metacognition and self-awareness, and awareness appears to be distinct from general language or reasoning progress.
In briefBiMoGen: Bidirectional Motion-Text Generation with Masked Discrete Diffusion
SummaryThis paper proposes BiMoGen, a unified masked discrete diffusion framework that handles both text-to-motion generation and motion-to-text captioning. It uses a two-stage training scheme that first learns cross-modal correspondence and then specializes for bidirectional generation, along with a self-correction mechanism to reduce errors from inference-time self-generated context. Experiments on HumanML3D and KIT-ML are reported to show competitive performance on both tasks.
In briefReduce, then encode: a volume-reduction method letting 2D brain MRI models process 3D scans
SummaryThe paper addresses how to use pretrained 2D foundation models on 3D brain structural MRI volumes. The authors propose Multiscale Volumetric Reduction (MVR), which compresses each anatomical view's slices into a few 2D components, including a PCA-based base component and multiscale residual detail components, before encoding with a frozen 2D model. Under the frozen-encoder setting, MVR performs strongly overall on ADNI, OASIS, and ABIDE and generalizes from ADNI to AIBL.
In briefLasting Partners Push Up Prices Among Learning Pricing Agents
SummaryThis paper studies pricing agents that repeatedly meet on a platform, asking whether the platform's choice of who faces whom changes the prices the agents learn. Using the Bertrand duopoly model of Calvano et al., prices are set by tabular Q-learning, with randomized assignment of keeping the same partner, seeing rival prices, and messaging. Keeping the same partner raised average training-period profit levels by 0.27 of the gap between competitive and monopoly profit (95% CI 0.20 to 0.35).
WhyA preregistered randomized experiment shows fixed partners raise learned prices, separating punishment from static best response with reasonably solid evidence for algorithmic collusion research.
In briefKernelOnet: An Interpretable Neural Operator Built on Kernel Functions
SummaryThis paper proposes KernelOnet, an interpretable neural operator that explicitly builds kernel functions into its architecture to match boundary-type kernel-expansion methods. It offers three kernels: a learnable data-driven kernel, a physics-informed kernel that embeds analytic fundamental solutions, and a hybrid kernel. On three benchmarks and one shallow-water waveguide engineering problem, the authors report high accuracy.
In briefStructural Alignment for Reliable Industrial AI: Linking Physical Reality, Data, Models and Human Intent
SummaryThis paper proposes a framework that treats reliability of industrial AI as a problem of aligning four interacting domains: physical, representational, machine, and human cognitive. The domains are connected by two interfaces, digitalization and goal encoding. The authors characterize the space of admissible solutions by existence, non-uniqueness, robustness, and interpretability, and show how mismatches at the interfaces lead to reliability failures. Case examples in healthcare, energy grids, and subsurface exploration indicate that dominant failure modes differ across domains but share a common structural origin.
In briefRethinking Visual Token Compression for Video LLMs: A Simple Baseline
SummaryVideo LLMs face slow inference because long videos produce many visual tokens. This paper introduces SimpleCluster, a training-free baseline that clusters visual features across frames with position awareness and represents each cluster by its mean feature. Experiments on four video benchmarks and three Video LLMs show it matches or beats recent compression methods across many retention ratios, and stays robust at very low retention such as 1%. The authors find that better downstream performance tracks better preservation of the original feature distribution, especially global coverage.
In briefThe Hidden Ratio in Adam: A Stable Structure for Compression and Sign Dynamics
SummaryThis paper studies the Adam optimizer in the tied-β regime, where its two moment EMA decay rates are equal. The authors show its adaptive behavior can be described by a transformed ratio that is approximately scale-stable and heavy-tailed across tasks, model scales, and training stages. This allows a reparameterization that stores the second moment as a compressible state, which the authors report can be held in a fixed 4-bit codebook with performance competitive with full-precision Adam. The transformed ratio also links Adam to sign-based methods such as Signum.
In briefEmbodied Semantic Communication for Collective Autonomous Agents: A Tutorial on Representation, Wireless Delivery, and Closed-Loop Coordination
SummaryThis paper proposes embodied semantic communication (ESC), a paradigm that turns information transmission into action-oriented semantic interaction so heterogeneous agents can coordinate actions from shared information. It clarifies ESC's conceptual boundary, system characteristics, and technical pathways, and maps relevant mathematical tools such as semantic information theory, world models, and multi-agent decision theory. It also summarizes open challenges, including measurable semantic reliability, ambiguity-triggered interaction, and bandwidth-adaptive transmission.
In briefInductive Feedback for Mixed-Policy Distillation: a new way to train language models with verbal feedback
SummaryThis paper studies how to use verbal feedback, often produced by a capable model, to post-train language models. The authors argue that standard on-policy distillation can transfer teacher preferences the feedback did not motivate, while leaving much of the feedback's guidance unused. They treat feedback as evidence for or against a token coming next, build a target distribution within a trust region of the student, and derive a shared-rollout estimator of a symmetric divergence. Experiments report gains over the common recipe and a contrastive variant on knowledge-based and agentic benchmarks.
In briefTRACE: Localizing Decoding-Time Risk for Better LLM Confidence Calibration
SummaryThis paper introduces TRACE, a single-pass confidence estimator that keeps the decoded answer and tracks token-level surprisal and predictive entropy during generation to localize uncertainty spikes. It converts this localized trace risk into answer-level confidence, and its calibrated variant TRACE+ maps trace-only features to probabilities using a held-out split without extra generations or external verifiers. Across four tasks and seven LLMs, TRACE+ improves Brier score and AUROC over the strongest baselines.
In briefGeoForward: Geometry-guided cross-view alignment for fast 3D scene reconstruction
SummaryThis paper proposes GeoForward, a feedforward 3D reconstruction method that first uses sparse LiDAR points and calibrated cameras to establish cross-view correspondence, then refines it with visual features. The geometry-conditioned representation is fused with multi-view context from VGGT to predict depth and Gaussian attributes. On nuScenes, it achieves the lowest cross-camera depth-consistency error and Chamfer distance among compared methods while keeping competitive reconstruction quality.
In briefDeveloper-Written MCP Error Messages Hurt the Most Capable Agents
SummaryThis paper examines 3,001 error messages from 150 widely used MCP servers, finding that 949 tell the caller what to do next, and about half of those steps depend on information the server cannot see. Tests with five OpenAI models on Berkeley Function Calling Leaderboard tasks show that following these steps, such as running a terminal command, lowered recovery rates. The authors propose two remedies: MCP developers can name a specific server tool in the step, and agent developers can delete the step with a one-sentence prompt.
In briefRecovering brain source dynamics from unknown local changes in EEG
SummaryEEG records mixtures of brain-source activity, and experiments that excite only part of the source-state space leave the dynamics unidentified. The authors show that unknown local changes to a single source's update rule leave a rank-one, source-specific signature in later EEG that can be isolated and used to recover source interactions. They give sufficient recovery conditions and an estimator with a noise bound. In simulations on MRI-derived anatomy, eight unknown changes recovered all dynamics in 32 of 32 systems when baseline excitation covered only four of twelve sources. Nonlocal changes and forward-model error limit accuracy.
Ayana Mussabayeva, Jiaqi Sun, Anuar Aimoldin et al.
In briefMultilinguality in Hybrid Attention Language Models
SummaryThis paper presents the first study of how hybrid attention, which combines full attention with recurrence-based alternatives, affects the multilingual ability of large language models. The interpretability analysis finds that cross-lingual representations follow the ordering of recurrent and full-attention layers, with a notable spike in cross-lingual alignment around the first full-attention layer. In distillation experiments on multilingual data, all alternative layer orderings outperformed the standard ordering throughout training, learning up to 2.5 times faster. The authors suggest that multilingual models may benefit from starting with a full-attention layer rather than recurrent layers.
In briefFirst learn, then memorize: why diffusion models generalize before they memorize
SummaryThis paper studies why diffusion models trained on a finite dataset first generate novel samples and only later collapse onto their training set. The authors show that training dynamics are governed by the Neural Tangent Kernel Gram matrix on noisy data, and that repeated noising splits its spectrum into a large-eigenvalue part linked to generalization and a small-eigenvalue part linked to memorization. Analytical results in high-dimensional limits and experiments on CelebA and U-Nets support this picture, and truncating the Gram matrix or penalizing the second bulk is shown to tune the transition.
WhyExplains diffusion generalization-then-memorization via two spectral bulks of the NTK Gram matrix, with analysis plus CelebA and U-Net experiments.
Rapha\"el Urfin, Tony Bonnaire, Giulio Biroli et al.
In briefHow Well Can LLMs Mimic Real Students' Ratings of Educational Feedback?
SummaryThis paper examines whether large language models can simulate real learners' subjective evaluations of feedback on high-school biology questions, at both group and individual levels. Six models were compared with and without learner-specific information, such as personality traits and example evaluations. The results show LLMs have limited ability to simulate learner evaluations. Adding learner profiles and examples improves score calibration and individual-level simulation, but more often fails to improve group-level consistency.
Momoka Furuhashi, Kouta Nakayama, Takashi Kodama et al.
In briefFrom Pixels to Poses: Learning Tool Use from Human Demonstrations Centered on Objects
SummaryScaling robotic manipulation is limited by the scarcity of real-world robot data, and existing methods that use human videos are computationally expensive and still need paired human-robot data. This paper introduces P2P-T, an object-centric framework that learns tool use directly from human demonstrations in two stages: pretraining an object-centric world model to extract stable pose priors, then integrating these into an efficient pose-aware low-level policy. The approach bypasses human-robot aligned data and reports a 73% improvement in execution performance over the previous state of the art on complex real-world tool manipulation tasks.
In briefDeep learning in neuroscience: from molecular mechanism modeling to classifying consciousness states
SummaryThis review critically examines current methods for studying conscious states, including classification, clustering, modeling of brain states under anesthesia, and identifying measurable neurobiological markers. It compares neural network detection of consciousness states from EEG and fMRI data, modeling of brain structural-functional dynamics under anesthetics, and neurophysiological indicators correlated with consciousness level. Deep neural models show growing effectiveness, but limitations remain, including limited model interpretability, lack of standardized metrics, and questions about the specificity of consciousness markers.
Elena Benderskaya, Anastasiia Alifanova, Svetlana Batalova et al.
In briefA decision-support system for legal reasoning with explainable outcomes
SummaryThis paper presents a framework that applies formal rules from regulations to support decisions, focusing on compliance for Law Enforcement Agencies processing data. It shows how symbolic AI and SPARQL can support legal reasoning. It also describes an algorithm that generates justifications for reasoning results, and a decision-tree method for determining what additional information to request when reasoning is inconclusive.
Jeremy Bouche-Pillon (IRIT, IRIT-MELODI, IRIT-ADRIA et al.
In briefEgo-Forge: Generating first-person video from third-person footage without text or geometry attention
SummaryThis paper addresses generating first-person (egocentric) video from third-person (exocentric) footage and a target head trajectory. The proposed Ego-Forge framework replaces external text captions with conditioning tokens derived from the model's hidden states and removes geometry-guided attention, using only a lightweight depth prior. The authors report state-of-the-art results on Ego-Exo4D, faster end-to-end inference, no external annotation at inference, and generalization to in-the-wild scenes.
In briefFrom Input to Output: A Flexible Agent for Interpreting Sparse Autoencoder Features
SummaryThis paper proposes functional interpretation, which treats a sparse autoencoder (SAE) feature as a mapping from the input semantics that activate it to its effects on output under intervention. It introduces DAFI, an agent that gathers evidence and refines input-side, output-side, and functional interpretations using component-specific feedback. On GemmaScope, DAFI outperforms SAGE on input score and Token Change on output score while using fewer tokens than a general coding agent. The paper also reports gains from distilled skills and improved steering-feature selection on AxBench.
In briefDo Temporal Link Predictors Need Learned Memory? A Simple Count-Based Baseline
SummaryMany temporal link predictors learn node representations to summarize past interactions. This paper proposes a predictor based on statistical language modelling that pools transition and co-occurrence counts across sources, smoothed with destination frequencies or Kneser-Ney continuation counts, and combined with popularity, history, and recency without node embeddings. It achieves the highest MRR on 7 of 16 TGB and TGB-Seq datasets and outperforms several baselines broadly. With only 9 to 13 learned parameters, it serves as a simple, competitive baseline for future neural temporal link predictors.
In briefd-OPD: Future-Aware On-Policy Distillation for Block Diffusion Language Models
SummaryThe paper addresses a supervision mismatch when distilling autoregressive language models into block diffusion language models, where the student sees visible future context within a block but the teacher conditions only on the causal prefix. It proposes d-OPD, which corrects the teacher distribution by incorporating visible future information within each block so supervision better matches what the student uses. The authors report that on Qwen3 models from 0.6B to 8B, d-OPD improves the six-benchmark average by up to 4.0 points over OPDLM and cuts training time by 1.35 to 1.58 times.
In briefNeural encoders that turn unevenly sampled signal sets into fixed-size vectors
SummarySensor networks and radars often deliver signals as organized sets, where each signal may have different sampling parameters. This paper studies neural network encoders that map such heterogeneously sampled signal sets into a fixed-size vector space, so they can be processed without sampling differences. The representations are evaluated mainly by how well they discriminate waveforms. All experiments use synthetic complex-valued radiofrequency signals.
Martin Bauw, Santiago Velasco-Forero (CMM), Jesus Angulo (CMA)
In briefDo coding agents reuse existing code or reinvent the wheel?
SummaryThis paper introduces RepoReuse, a multi-turn benchmark that audits whether coding agents reuse existing code in real repositories, with requirements revealed turn by turn and the workspace accumulating across turns. An audit over 3,000 turns finds that agents progressively explore relevant repository code less, reuse their own history less, and leave duplicated logic in 50.8% of task chains by turn 5, while pass rates barely change. The authors argue that such redundancy is invisible to pass rates, so code generation should be evaluated beyond functional correctness.
In briefStratified inoculation prompting: limits unwanted behaviour in language model fine-tuning while keeping desired traits
SummaryThis paper studies how to fine-tune language models so they learn desired behaviour without also learning undesired behaviour that appears in the same training data. The authors propose stratified inoculation prompting (SIP), which oversamples a small clean subset under diverse non-eliciting prompts while inoculating the remaining examples. Compared with standard inoculation prompting, SIP substantially reduces undesired behaviour and preserves more desired behaviour, and it also yields lower emergent misalignment rates in the harmful-advice setups tested.
Kajetan Dymkiewicz, Tim Farrelly, Adam Prada et al.
In briefFiona: Faster Encrypted Neural Network Inference Using Packing-Aware Ternary Weights
SummaryFully homomorphic encryption lets neural networks run on encrypted inputs, but it is still much slower than plaintext inference. This paper presents FIONA, an offline optimizer that selectively converts weights to ternary values within a given packing layout, keeping full-precision weights for sensitive groups. It also replaces some polynomials with lower-degree versions to reduce multiplicative depth and bootstrapping. On VGG11, ViT, and BERT, FIONA cuts plaintext-ciphertext multiplications by 53.4% to 79.5% and speeds up encrypted inference by 2.38x, 1.68x, and 1.84x, with under 1% accuracy loss.
In briefInterference in concept extraction that goes beyond feature geometry
SummaryThe paper introduces effective interference, a measure combining feature geometry and code statistics to capture realized interactions between features, distinguishing constructive from destructive effects. Under local fixed-support assumptions, it characterizes how four architectural mechanisms (feature orthogonalization, bias compensation, gain adaptation, and encoder-decoder separation) shape interference. Experiments with sparse autoencoders show that constrained architectures selectively reduce overlap among co-active features, while bias, gain, and encoder freedom allow constructive cross-contributions to remain.
In briefJailbreak-inspired method (J4U) for black-box uncertainty estimation in reasoning models
SummaryReasoning models aligned with reinforcement learning often become overconfident, and black-box settings lack access to logits. The authors find that existing black-box methods, such as paraphrase-based self-consistency and confidence verbalization, barely improve on simple repeated sampling. They propose prompt-level relaxation operators to widen the output distribution, and introduce J4U, a jailbreak-derived uncertainty quantification technique. Across 3 datasets and 4 models, including a closed-source one, J4U achieves statistically significant gains in more settings than the strongest baseline, with larger average ECE reductions.
Lucas Biechy, C\'edric Eichler, Adrien Boiret et al.
In briefQuasi-linear kernel attention with infinite capacity: an attention design that keeps expressivity while scaling nearly linearly with sequence length
SummarySoftmax attention in transformers has computational cost that grows quadratically with sequence length. The paper defines a kernel's capacity as the maximum sequence length for which its attention matrix can approximate the identity, as a measure of expressivity. It reports that softmax, Gauss, and Laplace kernels have infinite capacity, while common quasi-linear kernels have finite capacity. The authors propose additive kernels built from univariate spline and polynomial exponential kernels, prove they keep infinite capacity, and show they can be computed quasi-linearly via sorting, with benchmarks against softmax backends for long sequences.
Nicolaj Rux, Johannes Hertrich, Sebastian Neumayer
In briefFrom Data to Program: Fast, Direct Inference of Generative Programs from Empirical Data
SummaryThe paper introduces PRODiGI, a pretrained model that infers an explicit, executable generative program from a finite set of samples in a single forward pass. The inferred programs support direct sampling, density and score evaluation, and inspection without relying on the pretrained model. The authors also propose program-space fine-tuning, which adjusts differentiable program parameters to match generated and empirical samples while keeping model parameters fixed.
