In briefA network that enhances the quality of compressed screen content videos, which have abrupt motion and sharp text and graphics.
SummaryScreen content videos have sudden motion, scene changes, and fine details like text and graphics, which weaken methods that rely on temporal continuity. The paper proposes STM-Net, a spatial-temporal multi-scale network for enhancing compressed screen content videos, built from three modules: a prior-guided spatio-temporal dispatcher, bidirectional temporal feature extraction, and cascaded multi-scale feature distillation. The authors report that STM-Net outperforms state-of-the-art methods in both objective metrics and subjective visual quality.
In briefShared weights, selected computations: how looped Transformers route each loop's operation
SummaryLooped Transformers reuse the same layers across repeated loops, and this study asks whether each loop actually performs the same computation. Using graph walks as a test case, the authors show that recurrent progress varies and that a learned linear layer J can steer a frozen loop toward different transitions by changing its entering hidden state. Activation patching indicates that attention routing is a causal pathway for this steering.
In briefSolving multi-agent box-pushing puzzles with LaCAM
SummarySokoban is a classic planning puzzle, but its multi-agent version has been little studied. This paper uses recent advances in multi-agent pathfinding (MAPF) to design Sokoban-LaCAM, a planner that handles task assignment and collision-free paths together. It reports efficiently solving instances with tens of agents and boxes while keeping completeness and eventual optimality guarantees. The authors argue MAPF can serve as a useful basis for broader collective automation problems.
In briefMaDE: A post-hoc correction that keeps learned trajectory predictions physically feasible
SummaryNeural trajectory predictors can achieve low error while violating dynamics, actuator limits, or state constraints. The paper introduces MaDE, a frozen post-hoc operator that maps predicted state transitions onto a learned feasible dynamics manifold, inferring controls and correcting them by gradient-based inequality reduction. On simulated and recorded vehicle data, MaDE drives the one-step dynamics residual far below that of raw predictors, at the cost of a higher average displacement error.
In briefOPSRD: On-Policy Self-Role Distillation, a method that uses a fixed expert role to guide a language model's math reasoning
SummaryThe paper introduces OPSRD, which uses the same base model with a fixed expert role as a teacher that provides next-token distributions on the prefixes of a role-free student's own generated trajectories, without reference solutions. Training supervises only the highest-entropy half of student positions, using teacher-weighted forward KL with clipping. Experiments on three competition-math benchmarks with Qwen3-1.7B, 4B, and 8B show gains over the base models without role prompts at inference, and forward KL gives the highest macro-averaged accuracy among the three divergences tested at every scale.
In briefConfidence-Scored Video Temporal Grounding: Controllable Generative Localization
SummaryThis paper studies video temporal grounding, which locates events in a video from natural-language descriptions. The authors add a lightweight confidence head that scores each generated time interval, enabling ranking, threshold-based selection and rejection without an external verifier at inference. On a fixed OMTG-Bench candidate pool, confidence ranking raises query-macro Recall@0.5 at 10% and 25% return budgets.
In briefLLM Persona Unlearning: Making a Chosen Persona Hard for a Model to Reproduce
SummaryThe paper argues that pre-training gives large language models many role and style behaviors, and post-training makes a helpful Assistant the default without removing other modes from the weights, so explicit prompts can still elicit them. The authors define persona unlearning as a weight-level edit and introduce PersonaUnlearnBench, a paired benchmark across six LLMs and five personas. They propose PaCE, which locates an internal behavior direction from target and desirable responses and trains the model away from the target persona, reporting consistent suppression with high response quality at a moderate utility cost.
In briefAlgorithmic Recourse Under Competition: Studies how improvement advice from machine learning models can become invalid when many people compete for limited resources and acceptance thresholds shift.
SummaryAlgorithmic recourse suggests minimum-cost changes so that individuals rejected by a model can obtain a desired outcome. This paper argues that when people compete for limited resources, widespread recourse implementation can shift the acceptance threshold even with the scoring model unchanged, potentially invalidating the original recommendations. The authors propose a framework that jointly optimizes recipients and their required score targets to balance recourse cost and post-shift validity. They develop an algorithm based on the Implicit Function Theorem and evaluate it on synthetic and real datasets.
In briefA PyTorch Library for Hyperspectral Image Models: A Unified Framework and Benchmark
SummaryThis report presents a modular framework that unifies 55 representative hyperspectral image models across six deep learning paradigms under a common registry and standardized tensor handling. It integrates 24 benchmark scenes and uses spatially disjoint regional blocking with guard bands to avoid train-test pixel overlap. Across 6,600 seeded runs, the authors find that scene difficulty matters more than architecture, and that some models under 1 M parameters match much larger ones. The code is publicly available on GitHub.
Tanishq Rachamalla, Aryan Das, Srishti Kaushik et al.
In briefGrammarRL: Reinforcement learning to make grammar-constrained decoding work better
SummaryThis paper proposes GrammarRL, a label-free reinforcement learning method that adapts language models to grammar-constrained generation using two self-supervised rewards from the model's own likelihoods. The rewards measure how likely the constrained output is given the input and how well the input can be reconstructed from the output. Evaluated on sign language gloss translation, hierarchical text classification, and named entity recognition with Llama models from 1B to 8B parameters, it outperforms constrained greedy decoding by 9.8 points on average and matches or beats beam search on two of three tasks at greedy-decoding inference cost.
Gabriele Tuccio, Antonino Furnari, Aldo Gangemi et al.
In briefCompletion-aware cross-fidelity reinforcement learning for bus holding across multiple lines
SummaryExploratory reinforcement learning on a live bus fleet is impractical, and policies trained only on historical data cannot gain new experience. The paper studies hybrid offline-and-online (H2O) RL, which combines fixed-target replay with simulator interaction, where the simulator's transition and event-duration dynamics can differ from the target. It addresses a failure mode in which generalized passenger time decreases while passenger journeys remain incomplete.
In briefBlocking forbidden skills in LLMs before release via gradient sealing
SummaryOpen-weight LLMs can be fine-tuned downstream, which may let users instill illicit knowledge. The paper studies preemptive unlearning, which aims to stop such capabilities from being acquired later, and finds that existing retrospective methods leave gradient pathways open even when outputs are suppressed. The authors propose gradient sealing, which pushes relevant pre-activations into the negative region where ReLU-family derivatives are zero or near zero. Experiments across several LLM families show stronger resistance to downstream acquisition than retrospective baselines.
In briefFIGS: A study testing whether AI stays truthful over long chats without becoming cold
SummaryThis paper introduces FIGS, an evaluation framework that tests whether language models keep factual integrity while offering grounded support across extended dialogues. It uses an adaptive 10-turn conversational simulator that challenges the model as users repeat requests or push back. The authors release 500 multi-turn scenarios and an automated judge, separating sycophancy from calibrated empathetic validation. Results show leading models either drift toward sycophancy over time or over-correct into detached rigidity, so balancing honesty with support remains unsolved.
Sidharth Pulipaka, Ruta Binkyte, Ivaxi Sheth et al.
In briefFork-dLLM: Avoiding the Flexibility Trap in Diffusion Language Models
SummaryThis paper studies masked diffusion language models, which generate tokens in parallel. It finds that confidence-based sampling often delays unmasking at high-entropy fork positions, reducing generation diversity. The authors propose Fork-dLLM, a hybrid sampler that uses autoregressive-style ordering only at uncertain fallback steps, and ForkGRPO, a reinforcement learning post-training method built on it. The reported experiments show pass@k scaling comparable to autoregressive sampling at 2-3x higher efficiency.
Stipe Frkovi\'c, Metod Jazbec, Christian A. Naesseth
In briefDimension-free rank lifting via random hyperplane arrangements
SummaryThis paper studies how wide a randomly initialized hidden layer of a neural network must be so that its random features have full row rank for a set of input vectors separated by an angle. For positively homogeneous nonpolynomial activations such as sign and ReLU, the authors show that roughly (1/θ)·max{m, log(1/δ)} neurons suffice, a bound independent of input dimension that is essentially tight. They also give a width guarantee for stable rank lifting, a quantitative version concerning the smallest eigenvalue of the feature Gram matrix.
WhyGives a dimension-free rank lifting bound exponentially improving prior results, with rigorous proofs unifying stable rank guarantees.
In briefCross-modal jailbreak attack that splits editing intent across text and image inputs
SummaryThe paper proposes the Split-Intent Jailbreak Attack (SIJA), which places the editing target in text and hides the safety-critical action in a visual cue. The authors introduce the SI2ESBench benchmark for testing image editing models with split-intent inputs. Reported attack success rates are 68.28% on GPT Image 2 and 91.38% on Qwen-Image-3.0-Pro. The paper also proposes editing-intent reconstruction as a defense that evaluates the jointly inferred transformation.
In briefDoes Role-Play Prompting Really Help LLMs? A Study of When It Works
SummaryThis paper tests whether role-play prompting consistently improves large language model performance across models, domains, and languages, using MMLU and MMLU-Redux. It finds the gains depend on model capacity, knowledge domain, and prompt language, and proposes a persona-related cognitive alignment hypothesis to explain when role-play works. The authors also introduce MLCP, a training-free multilingual prompt concatenation method, which they report outperforms vanilla role-play prompting on all tested LLMs.
In briefPredicting Multi-View Rashomon Representation: Can We Learn Where Models Disagree?
SummaryThis paper examines how different foundation models can encode the same input very differently, which the authors call Rashomon Representation. Such disagreement may signal inputs that a model handles inconsistently, making it a potential reliability indicator. Instead of comparing multiple models, the authors train a lightweight predictor that estimates disagreement from a single model's representation. Experiments across several models and datasets report that this disagreement is input-dependent, predictable, and generalizable.
In briefSEAR: A Benchmark for Evidence-Grounded Audio Reasoning in Deepfake Detection
SummaryAudio language models are increasingly used to detect audio deepfakes, but existing benchmarks only check whether their verdicts or explanations sound plausible, not whether they rest on real acoustic evidence. The paper introduces SEAR, a four-task benchmark that tests acoustic evidence identification and quantification, deepfake detection, and forensic rationale generation. It also proposes BAEA, an agent that gives a frozen model controlled acoustic tools. Experiments with six models show a clear gap between plausible rationales and verifiable evidence reasoning.
In briefMeasuring how role-prompted AI models leak capabilities beyond their assigned persona
SummaryThis paper studies role-capability leakage (RCL), where AI reasoning models given a role, such as a kindergarten student, write convincing in-role text yet still perform above that role's level on benchmarks. The authors introduce RoleCapBench, covering six educational roles and four levels from elementary school to A-level, and test three open-weight models. Naive role prompting gives strong role-voice scores but leaves above-role accuracy at 0.811 to 0.898. The authors also propose Injection, an inference-time method that adds role-specific capability guidelines and a prefilled response prefix, which reduces above-role accuracy for most models.
Pakhapoom Sarapat, Saksorn Ruangtanusak, Kunat Pipatanakul et al.
In briefBayesNDE: Bayesian Generative Modeling for Neural Density Estimation
SummaryThis paper introduces BayesNDE, a neural density estimator built on Bayesian generative modeling. It avoids invertible networks and Jacobian-determinant computation by inferring a sample-specific latent posterior to form an adaptive proposal, then applying bridge sampling to estimate density. On nonlinear and multimodal synthetic datasets, it improves density value estimation and structure recovery over state-of-the-art neural density estimators. Results on real-world datasets also show better anomaly detection.
In briefCommunityKV: Faster Long-Context Decoding via Graph-Based Token Grouping
SummaryThis paper addresses the high cost of self-attention and key-value cache memory transfer when Transformers process long contexts. It proposes CommunityKV, a training-free sparse attention framework that treats token selection as community detection on a graph built from QK^T scores already computed during prefill. New generated tokens are assigned to communities with a constant-time local update during decoding. On Qwen3 and Llama-3.1 models across three long-context benchmarks, it reports up to 1.25x end-to-end generation throughput over dense attention, and up to 1.71x with query-group graph aggregation at comparable accuracy.
Joe McKenna, Anastasios Alexandridis, Nathan Susanj et al.
In briefDyRAD: Radar View Synthesis for Moving Driving Scenes
SummaryDyRAD is a method that reconstructs dynamic driving scenes from radar data using static background reflectors and motion-tracked dynamic reflectors, rendering complete range-azimuth-Doppler tensors. Reflectors are rendered through a fixed analytic point-spread function, so sensor-induced spread is kept out of the scene representation, which also enables transfer across radar configurations without refitting. On the RADIal dataset, DyRAD recovers radar detections for 90.7% of reference-detected objects, compared with 26.9% for the strongest baseline.
WhyRenders full RAD tensors via static and tracked dynamic reflectors with an analytic PSF; achieves 90.7% detection recovery versus 26.9% baseline on RADIal.
Merav Keidar, Tomer Borreda, Rajalakshmi Nandakumar et al.
In briefIdentifying Data Source Is Not the Same as Testing Data Quality
SummaryThe paper studies whether repeatedly training on model-generated data degrades later models, and whether provenance can guide which generated examples to reuse. Using financial-risk text, generator attribution is 98.7% accurate on original passages but drops to 53.1% after paraphrasing and 29.0% after style rewriting, while generated-versus-human detection stays near perfect. Over three rounds of generation and retraining, the two selection rules, one based on source information and one on a reference model score, produced no detected stable difference in model degradation. The authors conclude that identifying data origin and judging data usefulness for training are separate problems.
In briefDynamic LoRA Experts and Prototype Ensembles for Class-Incremental Learning
SummaryThis paper tackles class-incremental learning, where models must learn new classes without forgetting old ones. It proposes DLEPEM, a two-stage rehearsal-free method that assigns a task-specific LoRA expert to each new task to reduce cross-task interference. It then combines frozen pre-trained-model prototypes with task-adaptive LoRA-expert prototypes for task-level discrimination. The authors report strong performance on standard and few-shot CIL benchmarks under the evaluated protocols.
Hongwei Zhao, Rui Liu, Yansong Liu (School of Computer Science et al.
In briefSteganography is easy to learn, but steganographic reasoning is hard to learn
SummaryThis paper examines whether language models can learn to hide their reasoning inside innocuous-looking text, a concern for chain-of-thought monitoring. It compares steganographic reasoning with two related abilities, steganographic messaging and encoded reasoning, across reinforcement learning, in-context learning, and supervised fine-tuning. The authors find that steganographic reasoning is much harder to learn, usually appearing only under fine-tuning and requiring at least twice as much training as messaging. It can be learned under all three methods on a cover task where hiding information is convenient.
In briefFast regularized policy optimization using one-step TD critic updates
SummaryThe paper studies policy mirror descent (PMD) in regularized Markov decision processes, pairing it with a persistent critic updated by one temporal-difference step. For finite discounted MDPs, it proves global linear convergence in value under exact coordinate-wise Bellman updates. For the stochastic version under a single off-policy Markov trajectory, it gives an expected value gap of ε after roughly Õ(1/((1-γ)^5 σ̃_b ε)) transitions, without trajectory resets, generative-model access, or nested policy evaluation.
WhyCombines PMD with a one-step TD critic, proving global linear convergence and reset-free sample complexity, a substantive theoretical advance.
In briefSpherical interpolation to keep new and old multimodal embeddings compatible
SummaryContrastive vision-language models embed images and text in a shared normalized space, but upgrading to a new model usually makes old and new embeddings incompatible, forcing costly re-computation of the whole gallery. The paper studies whether interpolating along the spherical geodesic between the old-model query and the aligned new-model query can improve retrieval without re-indexing. Experiments across multiple benchmarks and model families show it outperforms orthogonal alignment alone and restores backward compatibility in most settings.
WhyApplies spherical geodesic interpolation to backward-compatible retrieval; multi-benchmark experiments and a margin-based certification support the claims, with impact mainly on model upgrades.
In briefRainAtlas: A Multi-Continental Dataset for Precipitation Downscaling
SummaryThis paper releases RainAtlas, a large-scale, machine-learning-ready dataset for precipitation downscaling covering three continents. It harmonizes heterogeneous hourly kilometer-scale observations onto a common 2-km grid, with about 210,000 aligned low- and high-resolution precipitation pairs per region, from ERA5 reanalysis and direct observations. The authors benchmark state-of-the-art ML downscaling models and find that out-of-domain generalization varies substantially depending on the training regions.
Pierre-Louis Lemaire, Luca Schmidt, Wietze Suijker et al.
In briefP-SRM: Recovering Wrongly Rejected Predictions in Visual Tracking
SummaryMany visual trackers use rejection mechanisms to filter out unreliable predictions, but these can also discard correctly localized candidates. The paper proposes P-SRM, which reassesses rejected candidates using spatial responses, past accepted states, and native decision margins, and selectively restores reliable ones. Tested on six trackers and four datasets across nine configurations, P-SRM improves ranking of rejected candidates and overall tracking performance.
In briefEstimating label-noise transition matrices with guarantees using selective classification
SummaryThe paper proposes a method for estimating the label-noise transition matrix for noisily labeled data, based on one-sided selective classification. It avoids estimating class posteriors and provides finite-sample performance guarantees, while allowing flexible binary classification learners. The authors also introduce algorithms implementing the approach and give refined finite-sample performance bounds.
Xabier de Juan, Santiago Mazuelas, Yilun Zhu et al.
In briefSynthetic pre-pretraining still helps at scale, but not because of a grammar prior
SummaryThis paper studies pre-pretraining (PPT) on synthetic non-natural language data before standard language model pretraining. It covers five PPT tasks, four pretraining data mixtures, parameter scales from 500M to 7B, and budgets up to 100B tokens. The authors find that PPT's downstream and token-efficiency gains persist at scale, but the gains do not consistently come from a grammatical prior; instead they arise from tasks that improve long-range retrieval.
Atsuki Yamaguchi, Tatsuro Inaba, Joel Niklaus et al.
In briefA Chess Benchmark for Optimizing Prompts of Large Language Models
SummaryThis paper introduces a chess benchmark of 1,118 Lichess puzzles for studying automatic prompt optimization of frozen large language models without updating their weights. The benchmark offers deterministic scoring, renewable puzzles with adjustable difficulty, and links puzzle gains to short game-play rollouts. The authors evaluate six optimization algorithms on eight target models, finding that even the strongest model, Gemini 3.5 Flash, solves only about 55% of puzzles.
Timoth\'ee Lesort, Alejandra L\'opez de Aberasturi G\'omez, Tristan Karch et al.
In briefSelf-evolving failure bank that turns runtime feedback into persistent VLA policy learning
SummaryThis paper introduces FailBank, a framework that turns runtime safety-shield corrections into lasting improvements for vision-language-action robot policies. A fixed CBF-based safety module acts as an observe-only teacher, and outcome-aware admission selects corrections and successful actions for guarded LoRA updates. On the VLA-Arena benchmark with two backbones, FailBank raises task success by 8.5 and 6.9 points and cuts policy-induced cumulative cost by 35.6% and 23.8%.
In briefShould the AI show its support info? A method for choosing when to disclose.
SummaryThe paper studies when to reveal support information to a human decision-maker under a budget constraint. It shows the optimal policy is a threshold rule on the Value of Information and proposes a practical plug-in policy with bounds on its degradation and regret. Experiments and two user studies indicate selective disclosure can outperform no or full disclosure, though the advantage varies across tasks.
Carlotta Giacchetta, Alessando Bogani, Cesare Barbera et al.
In briefBeyond Accuracy: Certified Prefix-Invariant Low-Precision Fast Matrix Multiplication
SummaryThis paper studies fast matrix multiplication in low precision, where mixing token rows during rounding can let later tokens affect earlier language model outputs, breaking prefix invariance needed for likelihood scoring. On Qwen2.5-14B-Instruct, two FP8 fast realizations with normal-looking accuracy changed likelihood-based answers on 5.83% and 10.00% of 240 OpenBookQA items. The authors build certified two-level Strassen realizations on bounded integer codes using 49 block multiplications instead of 64, guaranteeing bitwise equality to a row-local classical int8 operator.
In briefBackward-state policy is part of the learning algorithm, not just a memory detail.
SummaryLow-precision training rounds tensors that the backward pass reads again, and each use may read the rounded forward value, the original, or a new rounding. The paper argues this policy is part of the learning algorithm, since copy accuracy and final loss cannot show whether it is correct. The authors derive per-use requirements for which value each use must read and check them on single operators without training. In three tests with PyTorch and Transformer Engine, these requirements correctly predicted whether reuse changes the average backward computation.
In briefDo Lower Prediction Errors Mean Better Flow Physics? A Study of Sim2Real Neural Operators
SummaryThe paper studies whether lower error in machine-learning surrogates for airfoil flow also means lower error in physical flow statistics, using paired CFD simulations and PIV measurements of a NACA4418 airfoil. Removing velocity fluctuations from the most energetic 10% of observed cells changed forecasts more than equal-area random removal, though the masks were not matched for removed energy, so this reflects sensitivity rather than physical importance. One neural operator (CNO) had lower velocity-field error but much higher fluctuation-energy error. The authors recommend reporting physical diagnostics alongside aggregate prediction scores.
In briefA Benchmark for Source-Free Universal Domain Adaptation on Time Series
SummaryThis paper presents the first benchmark for source-free universal domain adaptation on time series, where adaptation occurs without source data while handling mismatched label sets across domains. It also provides the first study of pretrained foundation models as feature extractors for time series domain adaptation. The authors identify that the unknown-sample rejection threshold in existing methods is highly sensitive and propose a plug-in auto-thresholding module for any such method. Experiments on three time series datasets support the module, and show that foundation models do not consistently outperform classical backbones.
Romain Mussard, Fannia Pacheco, Maxime Berar et al.
In briefStress-testing LLM lie detectors when models role-play false personas
SummaryThis paper examines whether probes that detect lies from a language model's internal states reliably flag falsehoods when the model adopts an anti-factual persona, such as a conspiracy theorist. The authors built a dataset of 8,916 human-reviewed responses from three LLMs and tested eight existing probes, many of which failed. Confounder experiments showed these probes often track concepts like instruction compliance or response likelihood that are spuriously correlated with truth. The authors introduce a simple linear probe that performs best on both stress tests, but conclude current lie detection probes remain unreliable.
Maximilian von Klinski, Sebastian Lapuschkin, Wojciech Samek et al.
In briefRATIO: Optimizing Reasoning in Quantized Models by Penalizing Overthinking Tokens
SummaryPost-training quantization shrinks large language models to cut memory and inference cost, but on reasoning models it can lower accuracy and cause longer reasoning chains. RATIO identifies model-specific overthinking tokens by comparing full-precision and quantized models, then assigns each a tailored penalty without extra training. Compared with quantized baselines, it reportedly improves accuracy by up to 9.8 points and reduces chain-of-thought length by up to 51.3%.
In briefFrom Image Latent Space to Fuzzy Rules: Interpretable Analysis of a Gastrointestinal Foundation Model
SummaryThis paper proposes a prototype-based fuzzy-rule framework that explains patch-level features from inner layers of pretrained medical foundation models without fine-tuning. It learns class prototypes by clustering and classifies features with human-readable IF-THEN rules. On capsule endoscopy, gastrointestinal endoscopy classification, and colonic polyp segmentation, it reaches accuracy comparable to black-box classifiers. The extracted rules are also used to examine where synthetic medical images depart from real tissue.
Michael D. Vasilakakis (1), Dimitris K. Iakovidis (1) ((1) Department of Computer Science, Biomedical Informatics et al.
In briefFinite-horizon Fisher memory in recurrent systems with bounded power
SummaryThe paper analyses how information is allocated, written and retained in finite-horizon linear-Gaussian noisy recurrent memories. It shows that the directional Fisher memory matrix has trace equal to the number of units, so non-normality redistributes information without raising its spherical average, and it gives finite-horizon error bounds. Experiments training input masks and linear readouts on fixed recurrent carriers are reported as matching the theoretical predictions.
In briefProbabilistic Adversarial Training: a paper on improving robustness by separating two distributions
SummaryThe paper views adversarial examples as arising from the overlap between a distance-based distribution and a victim-classifier-induced distribution. It proves that a KL-divergence-based quantity minus a log normalizing constant lower-bounds probabilistic robustness (PR), and maximizing this bound serves as a tractable surrogate. The authors show the objective recovers a scaled form of adversarial training, and experiments report consistent PR improvements.
In briefInline Memory and Reusable Skills: A Memory-Centric Framework for Vision-Language-Action Models
SummaryVision-Language-Action (VLA) models show promise for general robotic manipulation, but adapting them to new tasks is costly and can cause forgetting of earlier tasks. This paper proposes Optimus-R, a memory-centric framework that treats robotic adaptation as explicit query-skill memory tuning, using an inline memory interface, a query-skill memory bank, and a lightweight Bridge-and-Adapt mechanism. Experiments on in-domain adaptation, cross-domain adaptation, and lifelong learning show data-efficient skill learning with reduced catastrophic forgetting.
In briefTopTimeNet: A Time-Series Classifier Using Topological Features
SummaryThis paper introduces TopTimeNet, a model that distinguishes periodic from chaotic dynamics in time series. A fixed, non-learned stage extracts a 42-dimensional geometric and topological descriptor from Takens delay embeddings and persistent homology, followed by a lightweight learnable classifier. On a benchmark of 49 nonlinear dynamical systems, a 1,638-parameter configuration matches the mean accuracy of a version with 33 times more trainable parameters. However, robustness drops sharply when noise is added to the raw signal and the full feature pipeline is recomputed.
In briefPILOT: Deciding When to Retrain Online Forecasters Using Forecast Errors
SummaryThis paper addresses when to retrain time series forecasting models under non-stationary data streams, since retraining is costly. The authors propose PILOT, which uses completed forecast errors to decide when to retrain, creating pseudo-labels from future error increases and training a lightweight scorer. PILOT is a plug-in module that works with arbitrary forecasting backbones. The paper reports that across eight benchmarks and three backbones, it achieves the best average-rank performance among retraining policies with a favorable efficiency trade-off.
In briefSafety of Latent Communication in Multi-Agent Systems: A study of risks when agents exchange information directly in internal representation space
SummaryThis paper studies latent communication, where agents in a multi-agent system exchange information through trainable links that map sender representations into the receiver's input space, reducing text-based overhead. The authors find that even benignly trained links can increase harmful compliance compared with text communication, even though the agents themselves stay unchanged. They also introduce a reinforcement-learning attack and show that adjusting the rewards can repair compromised links, reducing harmful compliance without updating the agents.
WhyShows benign latent-link training raises harmful compliance; an RL attack lifts scores from 27.9 to 76.9 across three topologies and four benchmarks, with repair experiments.
Muhammad Huzaifa, Sina Mavali, Thorsten Eisenhofer
In briefContext Injection and Long-Horizon Reliability: Separating Recognition, Processing, and Counting
SummaryThis paper evaluates language model reliability on a 400-step ledger task by scoring three components separately: recognising records seen 5 to 300 steps earlier, processing the current record, and maintaining an incrementing counter. Experiments on Qwen3-8B show duplicate-recognition errors rising late in the episode, while the counter made no errors in 8,000 steps. Inserting irrelevant content before each step made the counter fail in most seeds, and external exact reminders were tested as a way to re-align it.
In briefExplore-over-Graph: hybrid embedding and LLM reasoning for question answering over incomplete knowledge graphs
SummaryThis paper introduces XoG, a framework for multi-hop question answering over knowledge graphs that are missing some facts. When missing facts break reasoning paths, XoG recovers them from learned graph structure, using relation statistics and KG embeddings, rather than from the LLM's own knowledge. The LLM acts as a semantic selector and reasoner within an iterative planning-exploration-reasoning process. Experiments on WebQSP, CWQ, and BRINK show that XoG outperforms comparable methods without task-specific training under incompleteness and cuts LLM token consumption by up to 33%.
In briefLearning to Segment Objects from Motion Without Labels
SummaryThis paper introduces MoSA, an unsupervised framework that learns an object prior from unlabeled videos without manual annotations. It generates multi-granularity motion pseudo-labels, trains a Perceptual Grouping Model via contrastive learning, and transfers the prior into a prompt-guided segmentation architecture. Across seven zero-shot benchmarks, MoSA significantly outperforms prior unsupervised methods and achieves performance comparable to the fully supervised SAM.
WhyLearns transferable objectness from unlabeled motion, matching fully supervised SAM across seven zero-shot benchmarks, a novel and impactful direction.
In briefCORD: Reusable Degradation Representations Learned Across Different Physical Systems
SummaryThis paper proposes CORD, a method for modeling equipment degradation across different types of physical systems. It pairs type-specific observation interfaces with a shared degradation backbone and uses two self-supervised objectives to learn structure within observations and degradation evolution across observation histories. It is evaluated on bearings, batteries, and cutting tools, and tested on turbofan engines, a type absent from pretraining. Results indicate that joint multi-domain pretraining yields representations reusable across devices, datasets, and system types.
In briefEchoPress: Training-free KV cache pruning that reuses prefill computations to match KVzip accuracy faster
SummaryEchoPress is a training-free method that reduces the memory used by long-context language model inference by pruning key-value cache entries. It approximates KVzip's context-reconstruction importance scores using queries and keys already computed during standard prefill, reconstructing only the first chunk to calibrate scores for the rest. On LongBench and RULER with Qwen3-8B and Llama-3.1-8B-Instruct, it matches KVzip's task accuracy at eviction ratios of 50% to 90%. It cuts compression overhead by 1.7 to 19.6 times and total prefill time by up to 2.9 times.
WhyApproximates KVzip's reconstruction-based importance scores with prefill queries and keys, training-free, a substantive improvement. Matches KVzip accuracy on LongBench and RULER with 1.7-19.6x lower
In briefVANDAM: Reading DNA sequences using molecular priors from DNA science
SummaryCurrent genomic foundation models treat DNA as a string and pretrain with masked token prediction, without explicitly modeling biochemical and structural properties. VANDAM extends their training with DNA molecular priors that can be estimated from sequence: in self-supervised training it predicts regional molecular properties from pooled representations, and when functional labels exist it also injects local features at the input. The authors report consistent gains over token-based objectives across four architecture families and nine held-out genomic tasks, and probing experiments suggest the priors generalize to other unseen molecular properties.
In briefGraphMAS: A Benchmark for Multi-Agent Coordination in Graph Learning
SummaryThis paper introduces GraphMAS, a systematic benchmark for evaluating how multiple AI agents can coordinate on graph learning tasks. It builds a shared pool of graph reasoning specialists and organizes coordination by inter-agent interaction and runtime adaptivity, yielding four paradigms and seven methods. Evaluated across seven text-attributed graphs, three domains, and two tasks, coordinating specialists outperforms single specialists and single-agent reasoning, while richer interaction does not reliably help. Instance-adaptive specialist selection gives the best accuracy-efficiency trade-off.
In briefRiemannian flow model fine-tuned with reinforcement learning for molecular crystal structure prediction
SummaryThis paper introduces CG-OMatG, an equivariant Riemannian flow-based generative model for predicting molecular crystal structures. It treats molecules as rigid bodies and learns their centroid positions, orientations, and lattice parameters, trained on subsets of the OMC25 and CSD datasets. The model is further fine-tuned with policy gradient reinforcement learning to favor low-energy candidates, and its outputs are validated on the CSP blind test benchmark using COMPACK packing-similarity analysis.
Thomas Egg, Harry Winston Sullivan, Maya M. Martirossyan et al.
In briefHow local loss-landscape geometry changes during language model pre-training
SummaryThis paper studies language model pre-training dynamics through the geometry of the local loss landscape and identifies two phases. In the first, high initial sharpness causes instability at large learning rates, and the landscape becomes flatter early on, which explains the need for learning-rate warmup and longer warmup for larger peak rates. In the second, the landscape is governed by gradient noise scale, and the authors propose a depth-flatness trade-off that motivates a dynamic batch-size schedule that starts small and increases late in training.
In briefMemCodex: Self-Programming Hierarchical Memory for Language Agents
SummaryThis paper introduces MemCodex, a self-evolving hierarchical memory system that organizes agent experience into executable memory programs covering summaries, relational knowledge, reusable skills, and latent memory. The system rewrites how each layer is built, indexed, retrieved, and routed, and at query time reads move from coarse to fine layers, stopping once enough evidence is found. The authors also present MemArena, a unified runtime for heterogeneous data and memory systems. According to the paper, MemCodex improves average task success by 10.1% over the strongest adaptive-memory baseline while using 3.4x fewer context tokens and running 2.1x faster.
In briefDiffWAM: A fast, efficient world-action model for drone navigation
SummaryThis paper presents DiffWAM, a geometry-conditioned navigation world-action model that turns the predictive features of a frozen video model directly into continuous UAV camera trajectories, avoiding future-video synthesis and reconstruction at deployment. It also introduces FastDreamer, which runs computation in parallel with flight and hands off trajectories asynchronously. On the 1,000-sample DiffWAM-1000 benchmark it reports a trajectory RMSE of 0.3492 m and an endpoint success rate of 74.40%, and an onboard version reaches 1.08 s model-pipeline latency on NVIDIA Jetson AGX Thor.
In briefRevisiting on-policy adversarial black-box distillation by calibrating groupwise reward geometry
SummaryThis paper studies how to transfer capabilities from API-accessible large language models, which expose only text outputs, into smaller student models. It identifies a mismatch in the existing GAD method, where the critic's training objective differs from the within-group relative rewards used by GRPO, which can cause collapsed reward scales or fragile margins. The authors propose Groupwise Reward Geometry Conditioning (GRGC), which uses Gaussian groupwise optimal transport calibration during critic training and group power modulation during policy optimization to shape reward groups. Experiments across diverse teachers, student models, and datasets report effectiveness on in- and out-of-distribution evaluations with negligible overhead over GAD.
WhyBuilds on GAD with calibrated groupwise reward geometry for advantage construction; broad experiments across teachers, students, and datasets show gains, impacting black-box distillation.
In briefMeasuring Farmland Ground Height Changes with Drone Photogrammetry and a Deep Learning Model
SummaryThis paper presents a U-Net-based algorithm that estimates vertical ground surface displacement from drone (UAV) photogrammetry point clouds, as an alternative to traditional ground filtering. The model predicts elevation corrections and also estimates their uncertainty through heteroscedastic regression. Tested on four datasets against several ground-filter methods, it performed slightly better in most cases. On one dataset, its subsidence RMSE was 6.1 cm, compared with 7.7 cm for the next best model, SMRF.
Wojciech Gruszczy\'nski, Edyta Puniach, Pawe{\l} \'Cwi\k{a}ka{\l}a et al.
In briefHierarchical RL that keeps routes valid for joint routing and switch placement in chip design
SummaryThis paper tackles joint routing and switch placement, a combinatorial optimization problem in chip design, using a hierarchical reinforcement learning framework. The method builds solutions step by step at the level of logical communication routes through switch expansion, switch placement, and route refinement, so every communicating pair keeps one loop-free route. Exploration uses Gumbel Monte Carlo Tree Search, and the authors report that neural-guided search substantially improves solution quality over non-learning optimization methods. Pretraining across floorplans is also reported to provide a strong initialization for fine-tuning on unseen instances.
Dorian Gailhard, Ugo Lecerf, Enzo Tartaglione et al.
In briefFAST: A scalable Transformer that matches pixels across images of a scene
SummaryThis paper introduces FAST, a model that finds dense correspondences between pairs of images of the same scene by estimating 2D displacements. The authors show that query-key projections from single-view vision foundation models can be reused for cross-view matching, converting selected self-attention layers into cross-attention without adding parameters. They also assemble a 6-million-pair training corpus. According to the paper, FAST reaches state-of-the-art results on a range of benchmarks and improves with larger backbones and more training data.
In briefNixtlaverse: An Open-Source Forecasting Ecosystem
SummaryThis paper presents Nixtlaverse, a set of open-source Python libraries for time series forecasting. The libraries share common long-format panel data and keyed forecast output formats, while each model family keeps its own specialized implementation rather than a single unified interface. The authors demonstrate the design on the public M5 data through a unified evaluation, runtime and memory profiling, and reconciliation of forecasts across the M5 hierarchy.
Olivier Sprangers, Max Mergenthaler Canseco, Marco Peixeiro et al.
In briefLatentHarness: learning latent actions for memory and reasoning through counterfactual policy distillation
SummaryThis paper addresses two bottlenecks in long-context reasoning: retaining evidence across long inputs and sustaining computation over many reasoning steps. LatentHarness unifies memory access and latent reasoning as sequential selection of latent actions, choosing at each internal step among THINK, RECALL from a fast-weight memory, or EXIT to emit the next token. The policy is trained with counterfactual policy distillation, which branches each action and scores its effect on the emitted token. The authors report that at 1.4B parameters, it improves on the strongest baselines by 2.8% and 10.0% relative across six benchmarks and runs 5.9x faster than the strongest long-context baseline.
In briefStable Transformers for Graph Generation: a study of how the denoiser's dynamics affect deep graph generative models
SummaryThe paper finds that graph Transformers used in graph generation lose effectiveness with depth, as repeated self-attention contracts node representations and causes vanishing gradients. The authors analyse this from a dynamical systems perspective and build a permutation-equivariant Transformer with non-dissipative transport, along with a damping mechanism that interpolates between non-dissipative and contractive regimes. Experiments on synthetic and molecular graph benchmarks show that stronger contraction progressively harms generation, so the denoiser's dynamical regime is identified as a key design factor for deep graph generative models.
In briefCNCGEN: A Dataset and Framework for Generating Machining Plans and Toolpaths from B-rep CAD Models
SummaryThis paper introduces CNCGEN, a dataset and learning framework for generating three-axis machining process plans and toolpaths from B-rep CAD models. The dataset includes about 50k geometrically verified synthetic machining flows and 800 held-out real CNC records. CNCGEN generates operations and toolpaths step by step while updating a compact machining state, and uses a learned surrogate verifier to give material-removal feedback during training. Experiments show it yields better workpiece geometry with less residual material and overcut than adapted CNC generation baselines.
In briefA library for differentiable signal processing and machine learning on spherical data
SummaryThis paper presents torch-harmonics, a library offering efficient, differentiable implementations of signal processing and machine learning methods for data on the sphere. It includes the spherical harmonic transform, vector spherical harmonics, discrete-continuous and spectral convolutions, and global and neighborhood spherical attention. The library also provides building blocks for spherical transformer architectures for scientific and engineering applications.
In briefOverForge: Strategy-and-Tactics Reasoning for Cooperative Lifelong Adaptation
SummaryThe paper presents OverForge, a training-free hierarchical architecture that separates persistent coordination strategies over roles and labour division from tactical action reasoning within each agent. A metacognitive module links the two levels by forming strategy-action branches and imagining their outcomes with a forward model. In the OvercookedV2 environment, OverForge delivers 7 soups in a connected kitchen, compared with 3 for each flat LLM baseline, while keeping agreed roles and adopting roles proposed by unfamiliar partners.
In briefLetting a Carrier Element Absorb Adversarial Changes to Keep the Subject Intact
SummaryThe paper introduces a "carrier," a secondary visual element that gives adversarial attacks an auxiliary region so that the main subject of an image is distorted less. The authors report that the carrier absorbs a larger share of globally normalized attack updates and improves cross-model transferability. In successful targeted attacks, humans still perceive the personalized subject as the primary content while the classifier is misled.
Linfeng Jiang, Steven McDonagh, Yuhang Chen et al.
In briefTIAgent: Generating Verifiable Computational Pathology Workflows from Natural Language
SummaryThis paper presents TIAgent, a framework that turns a natural-language description of an analysis task and its input dataset into a verifiable computational graph built from existing computational pathology operations. The graph can be inspected and executed, supporting uses such as biomarker discovery and tissue analysis. The authors benchmarked 14 large language models on 90 unit tasks, and Kimi-K2.7-Code achieved the highest Unit Workflow Success of 0.967.
Aakash Madhav Rao, Mark Eastwood, Danielle Belgrave et al.
In briefTrust Is Not a Score: Runtime Assurance Contracts for High-Risk AI Agents
SummaryThe paper proposes a Runtime Assurance Contract (RAC), a formal policy schema that ties an AI agent's autonomy to observed evidence during consequential tasks. A failed or unknown mandatory gate forces retry, escalation, deferral, or stop, so aggregate scores cannot authorize action. Experiments use synthetic failure-injection cases in agentic coding and compare RAC against a score-only rule and a restricted baseline. The authors state the results establish neither deployed safety nor cross-domain effectiveness.
In briefArchitectureIQ: A Benchmark for Measuring Model Training Intuition
SummaryThis paper introduces ArchitectureIQ, a benchmark comparing how well language models and human researchers predict which training recipe will perform best on synthetic datasets. The authors find that LLM intuition is good but imperfect, with frontier models trailing the best human researcher on architecture questions. They report that extra chain-of-thought compute does not help much, that a 20-item knowledge base substantially boosts weaker models, and that model intuition is insensitive to dataset properties.
In briefMeddeX: LLaMA-3 Medical Report Analysis with Retrieval and Gemini Vision
SummaryThis paper presents MeddeX, a system that analyzes medical reports, including text PDFs, scanned PDFs, and radiological images, to produce concise, evidence-backed summaries. It combines a QLoRA fine-tuned LLaMA-3 (8B) model, Retrieval-Augmented Generation, and a Gemini Vision layer, along with features such as garbage-text detection and longitudinal report comparison. Evaluated against the base LLaMA3-Med42-8B model, it shows gains across metrics, including +71.4% BLEU and +11.8% ROUGE-L.
Rajdeep Pandhare, Rohit Kundaikar, Yakub Kojar et al.
In briefDrift Inspector: Measuring How Research Fields Shift Using Atomic Contribution Claims
SummaryThis paper presents Drift Inspector, an open-source system that uses an LLM to extract standalone contribution claims from each abstract and clusters them over time into an interactive map. Applied to six years of EMNLP papers, it shows the field moving from classic NLP tasks toward LLM-era capabilities such as reasoning and multimodality. The same pipeline has processed the full ACL Anthology, and extraction was human-validated.
Vsevolod Karimov, Stepan Ostarkov, Anastasia Poroshina et al.
In briefBTC3D: A training-free method that blends image tile conditioning to preserve fine details in image-to-3D generation
SummaryThis paper addresses the loss of fine detail in diffusion-based image-to-3D generation, where global image features compress spatial information. It proposes BTC3D, an inference-only framework that extracts local conditioning signals from split image patches and gradually increases their influence in later diffusion stages. According to the authors, the method improves texture quality and visual fidelity while keeping global structure consistent, without any retraining.
In briefUniBuc's SemEval-2024 Task 2 method for clinical trial language inference
SummaryThis paper describes the UniBuc team's entry to SemEval 2024 Task 2, a safe biomedical natural language inference task on clinical trials. The team used SOLAR Instruct without fine-tuning, relying on input manipulation and tailored prompts for individual clinical trial report sections in zero-shot and few-shot settings. The approach achieved a consistency score of 0.72, ranking 14th on the leaderboard. The error analysis found the model tends to take shortcuts and rely on simple heuristics, especially under semantic-preserving changes.
Marius Micluta-Campeanu, Claudiu Creanga, Ana-Maria Bucur et al.
In briefPredicting When Masking Helps or Hurts Compressed CLIP Robustness Before Deployment
SummaryThis paper proposes a label-free, pre-deployment diagnostic, the Spurious Inversion Metric (SIM), to predict whether masking-based token pruning will improve or harm worst-group robustness in CLIP. Across 8 spurious-correlation benchmarks, masking's effect proved highly unstable, improving accuracy on some datasets and degrading it on others. Gating deployment by SIM's sign recovers masking's benefits while avoiding its worst failures, matching or exceeding a strong pruning baseline on 7 of 8 datasets.
In briefA Helps B While B Hurts A: Directed Transfer in Instruction-Tuning Mixtures
SummaryThis paper studies how to choose instruction-tuning tasks for a language model under a fixed budget, showing that transfer between tasks is directional: task A can help task B while B harms A. The authors introduce a transfer map, a signed estimate of how each source task helps or hurts each held-out target. Fitted from hundreds of fine-tuning runs on Qwen3 and Mistral models from 0.6B to 32B parameters, the map selects helpful tasks and raises accuracy on reasoning targets by up to 14 percentage points.
In briefValues as Style: Separating Values from Meaning to Steer LLMs with Less Damage
SummaryThe paper proposes an editable semantic-value interface on frozen residual states, using a one-way pathway from semantics to values so that an LLM's value priorities can be changed while the scenario, facts, and task constraints are preserved. On two instruction-tuned backbones, the method improves semantic preservation and reduces benign refusals at comparable value alignment. On LLaMA-3.1-8B, compared with validation-selected prompting, it achieves similar alignment (0.750 vs. 0.748), higher BERTScore (0.938 vs. 0.923), and fewer contradictions (5.1% vs. 7.6%).
In briefLLM-as-a-Judge for Romanian: Adapting Ragas and Comparative Ranking
SummaryThis paper tests whether large language models can act as judges for evaluating Romanian retrieval-augmented generation (RAG) systems, adapting the Ragas framework with Gemini 2.5 and Gemini 3. The authors built AdminRo-Eval, a dataset of Romanian administrative documents annotated by native speakers, as ground truth. Three evaluation methods were compared, and the best method varied by metric: granular decomposition gave the highest human alignment for Faithfulness (96% with Gemini 2.5 Pro), while comparative ranking performed better for Answer Relevance (90%).
In briefGFD-OPD: Distilling Large Diffusion Models into Smaller Ones Across Scales
SummaryThis paper studies on-policy distillation of diffusion models from a large teacher into a much smaller student, a setting where the standard recipe was found to fail. The authors introduce Fixed-State KL to measure the student-teacher distribution gap during training. They attribute the difficulty to the smaller student's limited capacity and to classifier-free guidance amplifying discrepancies between conditional and unconditional branches. They propose GFD-OPD, which reduces the gap while avoiding CFG error amplification, and report state-of-the-art results on all benchmarks.</summary>
In briefShieldCLIP: Selective Safety Alignment to Reduce Harmful Content in Multimodal Models
SummaryThe paper proposes ShieldCLIP, a framework that performs safety alignment based on the observed safety state of each modality rather than the sample's origin, so safe content is preserved while only unsafe content is redirected. The authors also release ViSUv2, a 195k-quadruplet dataset with independent per-modality safety labels across 578 concepts and 28 categories. Evaluations on cross-modal retrieval, text-to-image generation, and image-to-text generation report fewer harmful outputs than prior safety-aligned encoders and mitigation baselines, while preserving the original embedding space's utility.
Tobia Poppi, Silvia Cappelletti, Samuele Poppi et al.
In briefBetter Supervision Is Nearby: Neighborhood On-Policy Self-Distillation
SummaryThis paper proposes Neighborhood On-Policy Self-Distillation (N-OPSD), an improvement on on-policy self-distillation for training mathematical reasoning models. Locally perturbed parameter settings provide complementary reference-aligned corrections that cover more reference positions than the unperturbed privileged teacher. A frozen expert pool is built offline by greedy selection, and online routing chooses experts for student-visited states. On AIME 2024, AIME 2025, and HMMT February 2025, N-OPSD improves Average@12 over OPSD across Qwen3-1.7B, 4B, and 8B.
In briefRACE: Residual-aware test-time adaptation for forecasting with related histories
SummaryThis paper addresses time-series foundation models (TSFMs) forecasting when many related historical series are available, a setting common in continuous glucose monitoring and web/cloud workloads. The proposed RACE framework retrieves query-compatible neighbors, aligns their residuals, and aggregates coherent evidence into a training-free correction, with an optional lightweight domain-specific gate deciding when to apply it. The authors report that across four TSFMs, RACE improves all three domain-aggregate metrics on both primary domains, with the largest gains on high-error queries, and that gates transfer across backbones.
In briefRoboCoach: World Models as Active Coaches for Improving Composable Robot Skills
SummaryThis paper presents RoboCoach, a framework that uses a world model to imagine failures and decide which subtask demonstrations to collect and which skill modules to update. Across simulation and real-robot platforms, imagined and deployed success are strongly correlated. With only 150 additional subtask demonstrations, success rates rose substantially, and the improved skills transferred to held-out task compositions.
WhyUses imagined failures in a world model to target demonstration requests and expert updates, a novel closed loop. Validated on two simulation suites and two real robots against matched baselines.
In briefUnapologetically Distributed: A Case for Decentralized Document Analysis
SummaryThis paper introduces "Unapologetically Distributed," the first comprehensive study of distributed learning in Document Analysis across three axes: tasks, architectures, and fine-tuning strategies. It challenges the view of federated learning as a necessary performance trade-off for privacy, noting that its potential to improve robustness to out-of-distribution data has often been overlooked. The authors report that various distributed training approaches improve generalization on tasks such as Table Recognition, handwriting recognition, and Word Spotting, especially during transfer learning. They conclude that decentralization can be an opportunity to strengthen model robustness and adaptability in real-world settings.
In briefMC-PanDA++: A Simpler, Stronger Method for Adapting Panoptic Segmentation Across Domains
SummaryThis paper addresses unsupervised domain adaptation for panoptic segmentation, which lets models learn from labeled synthetic data and unlabeled real data to reduce annotation costs. The authors extend their earlier MC-PanDA method with self-supervised vision encoders for more robust initialization, per-class self-adapting mask-wide loss scaling, and a single-stage training pipeline. The authors report that the resulting method is conceptually simpler, performs better, and is more robust.
Ivan Martinovi\'c, Josip \v{S}ari\'c, Yuki M. Asano et al.
In briefSE-ADD: Self-Evolving Audio Deepfake Detection Using Mistake-Driven Supervision
SummaryThis paper addresses how audio deepfake detectors must keep working when new spoofing attacks appear after deployment. It proposes SE-ADD, a framework that iteratively adapts an audio language model with LoRA, using supervision built from the model's own verdicts and self-generated forensic cues, with extra cue-augmented training on misclassified samples. Experiments on two audio language models show better generalization to unseen attacks, with equal error rate (EER) reduced from 36.72% to 7.52% for Qwen2-Audio and from 19.93% to 3.97% for MOSS-Audio.
In briefConflict Between Logic and Memory: Training Conditions Shaping Optimizer-Dependent Rule Learning
SummaryThis paper studies why different optimizers fitting the same task can learn different generalizing relations in single-hidden-layer ReLU networks. Using composite evidence tasks, parameter-level interventions, and a strict-parity scan, the authors find that trainability of nuisance-connected weights reshapes both shared failures and relative optimizer advantages. Fixing first-layer nuisance-connected weights at initialization substantially raises accuracy on the hardest stage, whereas masking inputs after full training does not.
In briefNodeGround: A Node Classification Benchmark for the Graph Foundation Model Era
SummaryThis paper introduces NodeGround, a benchmark that compares graph foundation models with dataset-specific supervised methods for node classification under one shared evaluation framework. It covers 51 datasets, six graph foundation models, and 15 supervised methods, measuring both prediction quality and workflow costs such as adaptation, training, tuning, and inference. The results favor carefully tuned graph neural networks overall, and the benefits of reusing pretrained models depend on tuning budgets. The authors conclude that pretrained reuse does not yet offer a broadly reliable route to better predictions or cheaper workflows. The evaluation pipeline, records, and leaderboard are publicly released.`
In briefChronoGraph: A Functional 4D Scene Graph and Vision-Language Model for Understanding Interactions and Planning Actions
SummaryThe paper introduces ChronoGraph, a functional 4D scene graph that links actions on affordance parts to semantic and geometric state changes, representing observed and anticipated transitions in the same form. The authors build ChronoGraphBench from human-interaction videos and simulated robot trajectories, and train ChronoGraphVLM in two stages, combining graph-as-chain-of-thought supervised fine-tuning with joint 4D graph reinforcement learning. Experiments report gains over the corresponding pretrained baselines and zero-shot transfer to VLM4D, and real-world demonstrations show graph-based planning supporting mobile manipulation with existing robot skills.
Chenyangguang Zhang, Malgorzata Gwiazda, Guanlong Jiao et al.
In briefText-in-image attacks that mislead AI image detectors built on vision-language models
SummaryThis paper studies whether vision-language models (VLMs) used to judge if an image is AI-generated can be misled by text printed inside the image. The authors systematically test typographic attack strategies on detection-focused, open-weight, and commercial VLMs, in both real-to-fake and fake-to-real directions. They find reasoning modes are more vulnerable than direct modes, and larger models show higher clean accuracy but also higher attack success. The work also examines robustness under image and text transformations and exposes limitations of current VLM-based detection systems.
In briefHow attention in large language models has evolved: mechanisms, trade-offs, and new trends
SummaryThis survey reviews how attention mechanisms in large language models have developed, treating context as the model's internal memory. It compares research using five dimensions: memory representation, update, access, readout, and integration. Drawing on 59 release records from 14 major model lineages and 11 open-weight models, it finds that explicit-memory and recurrent-state methods increasingly overlap in function, and that heterogeneous architectures coordinate memory across network depth. The authors argue that efficient sequence design now centers on organizing and selectively using contextual memory.
In briefBAM! Bayesian Anything Model: A Lightweight Foundation Model for Physics-Aware Generative Imaging
SummaryThe paper introduces BAM (Bayesian Anything Model), a lightweight 36M-parameter foundation model for few-step, physics-aware posterior sampling in computational imaging. It upgrades the RAM backbone into a conditional flow map so that instrument physics can be specified at inference time, and is pre-trained on large image corpora and libraries of forward operators. The authors report that on linear inverse problems across several image benchmarks and a camera-shake benchmark, BAM outperforms specialised and zero-shot methods in sample quality using only 3 steps, at lower computational cost.
Alessio Spagnoletti, Charlesquin Kemajou Mbakam, Jonathan Spence et al.
In briefGraph Residual Conjugate Diffusion: A Diffusion Model for Graph Signals with Equalized Signal-to-Noise Ratio
SummaryThe paper introduces Graph Residual Conjugate Diffusion (GRCD), which gives each graph-frequency mode its own clock so that all modes share the same conditional signal-to-noise ratio, and fits a Gaussian reference at a finite terminal SNR. It is evaluated on five settings, including METR-LA traffic and Molene weather, against seven comparators. At four function evaluations, GRCD reduces averaged maximum mean discrepancy by 22 to 36 times over the best comparator on all five settings.
In briefDiffusable latents via structure-agnostic distillation of pooled teacher features
SummaryThis paper studies distilling pretrained foundation models into an autoencoder bottleneck to make latents easier for diffusion models to learn. The authors find that aligning a single pooled image-level descriptor to the teacher performs as well as or slightly better than dense position-wise distillation. The approach extends to 1D token-sequence latents and cross-modal settings, such as distilling a text encoder into an image autoencoder.
Adrien Ramanana Rahary, Nicolas Dufour, Patrick P\'erez et al.
In briefFANVIDv2: Judging Video Upscaling by Whether Faces and Plates Can Be Recognised
SummaryThis paper introduces FANVIDv2, a benchmark for video super-resolution that scores methods by whether face and licence-plate recognition work on their outputs, rather than by PSNR and SSIM. Low-resolution clips are created with randomised compound degradation, including blur, noise, JPEG compression and downsampling. With a 2.3M-parameter baseline, the authors report higher face and plate recognition scores. The annotations and evaluation code are released.
Kavitha Viswanathan, Vrinda Goel, Shlesh Gholap et al.
In briefFrom Modes to Memories: Scale-space dynamics of diffusion models
SummaryThis paper views diffusion models as deterministic dynamical systems indexed by noise scale, studying how the denoiser's fixed points and attractors behave at each scale. It introduces a critical scale σc, the largest noise level at which an example is still retained, to quantify memorization from duplication, overfitting, and outliers. Experiments in controlled settings and on Stable Diffusion show that σc tracks memorization and identifies fully and partially memorized examples.
Cristina L\'opez Amado, Marco Fumero, Francesco Locatello
In briefAllocating limited upload bandwidth across model layers for federated learning under extreme budgets
SummaryFederated learning clients often upload large model updates, creating communication bottlenecks, and uniform compression wastes bandwidth on less important layers. The paper proposes LBAT, which treats communication under strict uplink budgets as a resource allocation problem, estimating each layer's transmission value from local training signals. An exact byte dynamic programming allocator then selects rank and bit configurations within the budget. The authors report that LBAT outperforms uniform rank, uniform quantisation, and fixed compression baselines across various extreme budget settings on tabular prediction and data generation tasks.
