In briefEasy to predict, hard to compute: boundary dependence finds outputs entropy patching misses
SummaryThe paper shows that the Byte Latent Transformer (BLT), which places patch boundaries based on next-byte entropy, overlooks positions whose type is predictable but whose value must be computed, such as numbers after "=" in math solutions. On GSM8K, the entropy rule exactly answered 19.0% of such computed results, while a boundary after each "=" reached 51.8% and entropy combined with a label-free boundary-dependence signal reached 67.1%. Similar gaps were reported for byte models of several sizes trained from scratch.
WhyShows entropy-based byte patching systematically misses computed values; boundary dependence signal finds them, with multi-scale experiments and paired significance tests.
In briefFrom Sparse Representations to Behavioral Insights for Multimodal Depression Assessment
SummaryThis paper introduces BehavDep, a framework for multimodal depression assessment that decomposes behavioral representations into sparse latent factors linked to behavioral concepts via a semantic bridge. It learns video-level depression tendency scores under weak supervision and aggregates observations for user-level assessment. Experiments report the best overall assessment performance along with interpretable modality contributions and behavioral patterns.
In briefCompile the Table: Query-Calibrated Operator Compression for Efficient Tabular In-Context Learning
SummaryThis paper proposes QCOC, a method for compressing in-context examples used in tabular in-context learning. It compiles the full KV cache of the examples once into compact memory that is shared across later queries, using prototype clustering and closed-form value calibration against attention query vectors. On 64 held-out OpenML-CC18 datasets it reaches the highest mean accuracy among compared compression and retrieval methods. Compressing 8,192 examples to 512 memory slots gives a 10.5x cache compression ratio and faster online inference than full-context baselines.
In briefDetecting alcohol impairment in e-scooter riders during trips while limiting false alarms
SummaryThis paper presents a method that uses onboard inertial and throttle sensors to assess a rider's impairment during a trip and raises an alarm once enough evidence accumulates. The detector comes with a provable bound on its false-alarm rate. In experiments on 141 rides, it detected 91% of rides at the higher blood alcohol level and 50% at the lower level, with median detection times of 25 and 27 seconds. An embedded implementation was shown to meet real-time requirements.
WhyProposes in-ride impairment detection from inertial and throttle data with a provable false-alarm bound, validated on 141 rides; strong novelty and evidence, impact mainly limited to e-scooter setting
In briefSkill-V: A Verifiable Self-Evolving Skill Library for Interactive Agents
SummarySkill-V is a self-evolving skill library for interactive agents that represents each skill as a versioned, falsifiable contract linking intent to observable behavior. Task failures trigger new skills, while disagreements between contract evaluations and task outcomes guide revisions to existing skill boundaries. On ALFWorld and WebShop, Skill-V reports success rates of 95.3% and 85.9%, with a more compact library than growth-oriented baselines.
In briefDPPM: Dual-Path Parametric Memory for Personalized Language Models
SummaryThis paper proposes Dual-Path Parametric Memory (DPPM) to help language models personalize responses across sessions. An evidence path pools representations of interaction history to keep earlier evidence, while a delta path sequentially updates an associative state to capture changes. The fused outputs generate history-conditioned LoRA adapters. The authors report that DPPM outperforms baselines across several backbones, reaching 54.22% on PersonaMem-v2 and 86.79% on PrefEval.
In briefRouterInterp: Explaining How Mixture-of-Experts Models Route Tokens to Specialized Experts
SummaryThe paper proposes the Superposed Specialisation Hypothesis, which holds that experts in sparse Mixture of Experts models specialise in a disjoint union of fine-grained features rather than one broad domain. Building on this, the authors introduce RouterInterp, a method that identifies Sparse Autoencoder features most predictive of routing decisions and produces unified natural language explanations. On gpt-oss-20b, RouterInterp achieves about 65% higher detection accuracy than prior token-statistics-based methods.
In briefGuard models should also catch agents' skipped required safety steps
SummaryThis paper argues that guard models for LLM agents must detect required safety actions an agent fails to perform, called obligations, not just forbidden actions. The authors introduce ObligationBench, a benchmark of 240 expert-validated trajectories, and find most evaluated models identify obligations poorly. Their ObligationGuard model, trained on 40,000 synthetic examples, outperforms all evaluated models on recall and exact-match rate.
In briefOptimal random quantisers for spherically symmetric distributions: choosing a few random points that best approximate a round-symmetric distribution
SummaryThis paper studies how to choose random quantisers, which approximate a distribution with a finite set of points, so that the quantisation error is minimised for spherically symmetric distributions. The authors show that optimising over this class is a convex problem and give an equivalence theorem for global optimality along with a constructive algorithm. They report that random points uniformly spread on a sphere of suitably chosen radius perform very well for moderate sample sizes.
In briefFrom Video Clips to Creation Trajectories: The Sora100K Dataset for AI-Native Video Creation
SummaryThis paper introduces Sora100K, a dataset that models AI-native video creation as structured creation trajectories rather than isolated clips. It splits the records into text-to-video generation roots, single-turn editing edges, and multi-turn editing trajectories, with VLM-based annotations. Lightweight adaptation of LTX-2 models shows gains in visual quality, multi-shot generation, and cross-shot consistency, but following multi-turn editing instructions remains challenging.
In briefNarrow and Deep: An Ontology Tower as Knowledge for an Industrial Equipment LLM Agent
SummaryThe paper proposes an ontology tower, a narrow but deep ontology of a single equipment system that grows through quantities derived from measured points via physical relations and through operating-journal lessons incorporated as knowledge nodes. In a preregistered evaluation of nine tasks on a real low-humidity air-handling test plant, supplying this knowledge raised the rate at which agents avoided each task's most plausible misjudgment by about 20 percentage points. In live runs through an invariant safety layer, agents brought the controlled variable into its target band in 12 of 14 runs.
In briefSame Outcome, Different Evidence: Intent Recovery in LLM Safety Evaluation
SummaryThis paper argues that attack success rate (ASR), the common metric for LLM safety, cannot tell whether a model recognized and refused a harmful task or never understood it. The authors pair ASR with an operative understanding rate (UR), which checks whether a response identifies the evaluated task and treats it as one to answer. Results show similar ASR values can hide very different recovery rates, and that recovery rises with more explicit prompts while ASR does not follow the same pattern.
In briefThinking Inertia: LLMs Keep Reasoning Even When Told Not To
SummaryThis paper studies the "no-thinking" behavior of large language models, which is when they are meant to answer without explicit reasoning. The authors propose three metrics to measure whether pre-answer text is empty, relevant, or truly inferential. Testing six models with six prompting interventions across Boolean, multiple-choice, and open-ended questions, they find "Thinking Inertia": explicit inference persists under strict controls and grows as the answer space opens.
In briefRAGenome: Long-Context Retrieval-Based Genomic Language Model
SummaryThis paper introduces RAGenome, the first retrieval-based genomic language model, which extends pretraining to contexts 100 times longer than existing MSA-based models. Trained on whole-genome alignments from 100 vertebrates, it captures both cross-species evolutionary relationships and within-species long-range interactions. The authors report that gene finding performance rises from 0.45 to 0.60, while performance on evolution-based tasks such as prioritizing pathogenic variants remains competitive.
Frederikke Isa Marin, Panagiotis Antoniadis, Dionysia Danai Brilli et al.
In briefMemory Forcing: a streaming video generation method that keeps mid-length history accessible
SummaryAutoregressive video diffusion can generate video causally without a full bidirectional pass, but existing few-step models keep only the opening and most recent frames in a fixed KV cache. Events that leave this window become unreachable, which the authors call mid-horizon forgetting. Memory Forcing preserves this missing history under the same cache size, using Archive and Working banks and Bank-aware RoPE. The authors report stronger results on longer clips and subject or scene preservation on leave-and-return, and they extend the design to Wan2.2 5B.
In briefSR-TTA: Label-free adaptation to keep breathing detection working despite interference
SummaryThis paper studies detecting breathing with a 64-antenna cell-free massive MIMO base station. A learned beamformer uses spatial nulling to suppress motion interference, and a label-free method called SR-TTA adapts it at deployment by maximizing consistency across antenna subsets with an out-of-band spectral veto. Under simulated in-band interference, respiration-rate error drops from 5.8 to 0.8 bpm, and on real testbed recordings failures fall from 47% to 7%.
In briefWhat Output-Only Review Cannot Verify: Study Contracts for Research Agents
SummaryThe paper proposes study contracts that link declared experimental choices, run obligations, and claim scope to recorded execution evidence, so AI-generated studies can be checked against what was approved beforehand. This is explicitly distinguished from verifying scientific truth. In a diagnostic with eight self-authored clean and mutated pairs, a deterministic checker detected all eight registered mutations, while LLM judges given isolated packages flagged 104 of 144 mutated cases. The authors note these results do not measure verifier quality or agent reward hacking.
In briefDINO-forcing flow models: predict representations instead of denoising them
SummaryThis paper proposes a simpler way to train flow matching image generation models. Instead of co-denoising pretrained representations such as DINO through a second denoising trajectory, the model directly predicts the pretrained representation and is then conditioned on its own prediction. The authors report faster convergence and better FID scores, outperforming latent-space state of the art on ImageNet with 2x fewer epochs and improving pixel-space FID by more than 20% over comparable prior methods.
Arijit Ghosh, Lucas Degeorge, Paul Couairon et al.
In briefWhy draft trees miss target probability, and how exit-guided verification and training fix it
SummaryThis paper studies tree-based speculative decoding, which verifies multiple draft continuations in one target-model pass. The authors identify a canonical exit law and prove that target coverage bounds the expected output-block length of any exact path verifier. They propose Tree Exit Verification (TEV), an exact level-parallel verifier, and Exit-Guided Draft-Tree Training (ExitTrain), which uses missing target probability as feedback. Experiments report a 13% increase in output-block length from ExitTrain, and a 15% verifier latency reduction and 14% end-to-end speedup over DDTree from TEV.
In briefDEX: Digit-Level Early Exit for Energy-Efficient Neural Network Inference
SummaryThis paper presents a hardware accelerator for quantized U-Net brain-tumor segmentation that uses most-significant-digit-first arithmetic to make decisions before a full result is computed. Four runtime mechanisms act on the output digit stream, including exact early negative detection and two approximate methods chosen offline under an accuracy constraint. On BraTS, the approach reduces digit cycles by 38.38% while achieving a mean Dice score of 80.58% on 73 held-out cases, compared with 81.20% for the floating-point model.
In briefStreaming diffusion model for real-time video upscaling using cross-step attention
SummaryThe paper proposes a streaming-aware framework that adapts pretrained single-image latent diffusion models for efficient video super-resolution. Its Cross-Step Attention mechanism reuses intermediate denoising features across adjacent frames and diffusion steps to exchange temporal information, and Trajectory-Coupled Diffusion Scheduling improves temporal coherence. The authors report better perceptual quality and temporal realism on REDS4 and YouHQ40-Test, and over 40 FPS at 512×512 after cold start..
In briefThe Ball and the Box: Two Geometries of Computation in Superposition
SummaryNeural networks can encode more features than they have dimensions, a phenomenon called superposition. This paper studies how many dimensions are needed to compute Boolean gates from such representations using a threshold layer. It derives sharp dimension thresholds under two error criteria. The expected error count can vanish at fewer dimensions than needed for all outputs to be correct with high probability, because shared reads let rare cases cause many errors at once.
In briefTraceRelay: attention-aligned recurrent model over rolling traces
SummaryThe paper presents TraceRelay, a recurrent architecture that spreads persistent representations across a rolling sequence of low-dimensional traces, combining right-looking and left-looking attention. The authors ran 36 small-model experiments on Equal Repeats, bounded Dyck closing-type prediction, and causal Most-Freq generation. At trained length 256, models with recurrent phase inheritance reached 98.81–99.69% accuracy on Equal Repeats versus 50.73–51.63% without inheritance, and accuracy dropped sharply beyond the training length.
In briefCorrelational training: a new weight-update method for morphological neural networks
SummaryThe paper argues that standard back-propagation may be suboptimal for morphological layers, whose weight Jacobians are sparse and can yield poor gradients. The authors propose a weight update inspired by Multiplicative Weights Update, treating each morphological perceptron as a learning-from-experts problem in logarithmic space with a correlation-based reward. Evaluated on nine benchmarks, the method improves eight of them by up to 32.84 percentage points and substantially reduces run-to-run variability.
In briefScalable uncertainty quantification for neural networks using generalized Laplace active subspaces
SummaryThis paper proposes a low-rank generalized Laplace approximation for estimating uncertainty in neural networks, building a local Gaussian posterior over a small set of data-informed curvature directions. It compares the standard Bayesian posterior scaling with a mean-loss scaling. The authors report that mean-loss scaling yields a more stable, lower-dimensional active subspace and produces calibrated, coherent predictive intervals.
In briefFinding and fixing collider bias in Bayesian PINNs
SummaryBayesian physics-informed neural networks (B-PINNs) infer parameters from sparse or noisy data, but their common collider formulation can systematically bias the posterior over physical parameters, even when the prior is centered on the true values. The authors propose a hierarchical chain model, where physics generates trajectories that generate observations, and sample its posterior exactly with particle MCMC after discretizing the dynamics. Experiments confirm the predicted bias in standard B-PINNs and show the chain formulation avoids it.
In briefPEBench: A Unified Benchmark for Evaluating Video Prompt Enhancers
SummaryPrompt enhancers turn short instructions or multimodal references into structured cinematic plans for video generators. This paper introduces PEBench, the first unified benchmark for directly evaluating prompt enhancers across text-to-video, image-to-video, and reference-to-video tasks, with 1,100 expert-verified cases. It also presents an evidence-grounded evaluation framework using 24 criteria, and human validation shows its scores align closely with expert judgments.
In briefOnboard Marine Anomaly Detection on Φsat-2: From Simulation to In-Orbit Demonstration
SummaryThis paper describes a lightweight marine anomaly detection pipeline developed, integrated onboard, and deployed on the European Space Agency's Φsat-2 satellite to process imagery directly in orbit. The pipeline combines sea segmentation, self-supervised encoding of marine regions, and anomaly detection based on deviations from normal sea conditions. Early in-orbit tests showed a significant mismatch with simulated data, so the system was retrained on real Level-1 imagery using an improved annotation strategy, which substantially enhanced performance.
Clotilde Szywala, Thomas Goudemant, Marjorie Bellizzi et al.
In briefTimer-M1: A Multivariate Time Series Foundation Model Learned from Primitives
SummaryThis paper presents Timer-M1, a pretrained multivariate time series foundation model for zero-shot forecasting. It uses a primitive-based pipeline that synthesizes series with shared temporal patterns and assembles them into multivariate samples using relational patterns, with gated two-dimensional Transformer blocks. Across three large-scale forecasting benchmarks, it ranks first on FEV and TIME and second on GIFT-Eval among recent foundation models.
In briefMemTrial: Learning When an LLM Portfolio Agent Should Trust Its Memory
SummaryThis paper presents MemTrial, a large language model agent for portfolio management that tests how much each past experience actually changes its decisions, rather than crediting it with market-wide outcomes. It compares drafts made with and without each experience, then pools these estimates with a hierarchical Bayesian model and uses them only after they predict unseen dates, otherwise falling back to an equal-weight 1/N portfolio. On four benchmarks, it performs best on a semi-synthetic benchmark with known experience quality and limits its losses relative to 1/N to at most 2.2% on PortBench and InvestorBench.
WhyNovel counterfactual credit assignment via Banzhaf values and hierarchical Bayesian shrinkage; validated on four benchmarks with a semi-synthetic ground truth, giving fairly solid evidence.
In briefSpectral Weight Decay: A regularization method that encourages low-rank neural network weights
SummaryThe paper proposes spectral weight decay, a post-step nuclear-norm update that shrinks the spectrum of each weight matrix, unlike standard weight decay, which treats weights as vectors. On LLaMA models from 124M to 500M parameters, it lowers effective rank and improves SVD-LLM compression at matched validation loss. At 500M parameters and a 4% distortion budget, it reaches 1.89x compression and 1.18x GPU inference speedup, versus 1.14x and 1.01x with standard weight decay.
Dmitrii Andriianov, Andrey Veprikov, Aleksandr Beznosikov
In briefMAP4CS: A Multi-dimensional Data Pruning Framework for Efficient Code Retriever Fine-tuning
SummaryThis paper proposes MAP4CS, an adaptive data pruning framework for fine-tuning code retrievers used in retrieval-augmented generation. It selects a small, high-quality subset by combining syntactic structure, semantic diversity, and distributional representation, followed by rule-based filtering. On two large-scale datasets, using only 5% of the training data, it outperforms random sampling and performs comparably to or better than full-dataset fine-tuning.
In briefReducing gendered overcommitment in meme reasoning under ambiguity
SummaryThis paper studies gendered economic dependence in image-text memes and argues that models often infer roles without enough evidence, producing biased outputs. The authors propose CGER-Net, which checks whether an input provides sufficient evidence and abstains when it does not. On the EconMeme-GE dataset, it reduces the Gender Overcommitment Rate by up to 44% on ambiguous cases while keeping accuracy comparable on unambiguous ones.
Kushal Kanwar, Dushyant Singh Chauhan, Kapil Rana et al.
In briefMultiWorldBench: Do Independently Controlled Views Show One Shared World?
SummaryThis paper introduces MultiWorldBench, a diagnostic Minecraft benchmark with 495 case configurations across seven task suites and ten capabilities, testing whether independently controlled views stay consistent with one shared, persistent world. Using Engine GT as a reference, it evaluates Solaris (MP), Gamma-World, and the single-player Solaris (SP). Gamma-World scores highest among generated systems, yet all generated systems remain far below Engine GT, especially on state persistence and structural consistency.
In briefA 4-tensor attention model that predicts the next semantic scene state for video and robot planning.
SummaryThe paper proposes a 4-tensor attention model that predicts the next semantic state of a scene, intended for video generation and robot planning. It is tested on ROCStories, where each window poses a next-sentence prediction task at the semantic layer. On the validation split, it reports lower last-sentence cross-entropy than a one-dimensional free-running transformer baseline across three matched settings, with training finishing in 2.4 hours versus 45.2 hours.
In briefInternalizer: A Portable Hypernetwork That Turns Documents into Adapters for Very Large Language Models
SummaryThis paper introduces Internalizer, a hypernetwork that maps a document directly to a LoRA adapter so a language model can store that context in its weights. It generates document-specific adapters for the frozen 284B-parameter DeepSeek v4 Flash, a model about two orders of magnitude larger than targets in prior work. On unseen documents of up to 4096 tokens, teacher-forced top-1 and top-5 accuracy rose from 63.4% and 83.5% for the base model to 84.9% and 97.8% with the generated adapters.
In briefDeltaReplay: Step-Level Memory Reuse for Mobile GUI Agents
SummaryThis paper proposes DeltaReplay, a framework that lets mobile GUI agents decide, step by step, how to reuse stored execution trajectories without modifying them. Trajectories are saved as paths in a transition graph, and each action is split into a task-independent operation and task-specific parameters. At reuse time, each recorded step is either followed, executed with replaced parameters, or left to the base agent. On AndroidWorld and SPA-Bench, it improves task success rate over a base agent with the same backbone by up to 10.3 and 25.0 percentage points, respectively.
In briefDoes lighting-aware training help detect face swaps? A controlled study of Temporal Self-Blended Images
SummarySelf-blended images are commonly used to train face-swap detectors, but they mainly capture blending artifacts. This paper tests whether adding lighting inconsistencies helps, using Temporal Self-Blended Images (T-SBI), which transfer lighting statistics between frames of the same video. Across four datasets, no illumination-specific gains appeared, with AUC differences within seed variability. T-SBI does shift prediction scores, changing optimal thresholds substantially, so fixed-threshold comparisons can mislead. It did improve robustness to heavy JPEG compression on DFDC.
In briefChecking whether latent directions in synthetic vascular models match real interventions
SummaryThe paper argues that complete orthogonal predictive coordinates do not by themselves link latent directions to named interventions, and tests this with a mathematical and synthetic audit. Predictions can be exactly equivalent while interpretations remain valid only within the span of known perturbation signatures, and coordinate transport removed the large untransported edit error. The authors separate prediction, semantic support and scientific acceptance in executable controls. These are synthetic numerical audits, not clinical validation.
In briefWhat is the goal of unsupervised machine learning?
SummaryThis paper argues that unsupervised learning, unlike supervised and reinforcement learning, is a heterogeneous field that cannot be captured by a single goal. The author identifies four goals: estimating the data distribution, generating new data points, extracting features for downstream tasks, and understanding the data.
In briefA framework to compare AI puzzle-solving methods on cost, autonomy and approach
SummaryThis paper proposes a three-dimensional framework for analyzing AI methods for sequential decision making, covering formalization as a Markov decision process, degree of autonomy, and skill with computational cost. The authors use the Tower of Hanoi puzzle as a controlled benchmark to compare graph-based, reinforcement learning, and large language model approaches. They report that LLM-based methods shift complexity to inference-time verification because of weakly constrained action spaces, resulting in substantially higher memory and runtime costs than Neurosolver and FBRL.
In briefStructured sentiment analysis framed as sequence labeling of dependency graphs
SummaryThis paper studies structured sentiment analysis, which aims to produce a fine-grained sentiment graph whose nodes are spans of sentiment holders, targets, and expressions, and whose arcs capture relations among them. The authors treat the task as dependency graph parsing but solve it with sequence labeling, using linearized graph encodings that assign a label to each input word. Experiments on seven datasets in five languages (English, Spanish, Norwegian, Basque, and Catalan) show performance competitive with leading, more complex single-model approaches.
Muhammad Imran, Ana Ezquerro, Carlos G\'omez-Rodr\'iguez et al.
In briefMapping the Enzymatic Reaction Space: A Unified Benchmark and Pretrained Model for Enzyme Reactions
SummaryThis paper frames enzymatic reactions as a space linking reactants, products, and Enzyme Commission (EC) annotations, since reactions depend on both molecular structure and catalytic function. The authors introduce VenusRX-Bench, a benchmark covering forward reaction prediction, single-step retrosynthesis, and EC-number prediction, and VenusRX, a T5-style sequence-to-sequence model. According to the paper, VenusRX achieves the best or competitive results on most evaluated settings, and EC information improves reaction prediction while learned representations support EC prediction.
In briefCan Jev Serve as a Q-Function or Policy in Reinforcement Learning?
SummaryThis paper examines whether Jev, a decision model that returns calibrated, typed answers in one forward pass, can be used inside reinforcement learning. The authors analyze what each RL component requires of its answers and build algorithms that use Jev as a reference policy, an exploration judge, and a replay rater. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, while Jev itself stays untrained.
In briefEpistemic disturbance in conflict-resolution graph models: actions that change beliefs without changing the state
SummaryThe paper extends the graph model for conflict resolution (GMCR) with state-preserving actions such as announcements, exercises, leaks, and selective disclosures, which leave the physical state unchanged but alter other decision makers' beliefs. Using a four-valued extension that separates evidence for and against a move, it shows how such evidence can enable or disable perceived moves and how this shifts stability judgements. It distinguishes capability assessments, which affect all sanction-based stability concepts, from intention assessments, which affect only sequential stability. The 1995 DVD format negotiation illustrates that sequential stability can distinguish phases that general metarationality cannot.
In briefHI3D 3.0 (Twinkle3D): High-Resolution Generation of Object-Specific 3D Assets
SummaryThis paper presents HI3D 3.0, an image-to-3D system whose Twinkle3D geometry model generates watertight triangle meshes at 2048³ resolution, aiming to reproduce object-specific details such as inscriptions and brand marks. It improves surface quality, scales diffusion generation to long geometric token sequences, strengthens the initial generation stage, and adds a fine-grained image-3D cross-modal interaction mechanism. The authors report that the system outperforms four commercial systems on all reported metrics, recovering 82.1% of inscribed characters at 98.2% precision.
In briefCamera-noise residuals for face-swap detection: redundant with RGB, not complementary
SummaryThis paper tests on FaceForensics++ whether a Noiseprint++ camera-noise residual adds information complementary to an RGB Xception backbone for detecting face swaps. Ablations show that late fusion does not beat RGB alone, and the residual-only branch is near chance. The authors find the noise maps carry a statistical signal in per-sample moments, which an InstanceNorm layer at the noise-branch input removes, lowering cross-validated AUC from 0.747 to 0.554. They conclude the residual is redundant with RGB on this manipulation distribution.
In briefTRACE: Diagnosing Verifier Brittleness in Agentic Evaluation
SummaryThis paper introduces TRACE, a protocol for testing whether changes in an AI agent's evaluation score reflect real capability or changes in the scoring itself. Across synthetic and public tau²-bench tasks, renaming tools or reformatting outputs could shift scores, and identical reruns flipped 15-36% of outcomes. Two frontier LLM judges also disagreed on 57% of the same records.
In briefVESSI: A Vision-Language Model Framework for Surveillance and Investigation Support
SummaryThis paper proposes VESSI, a framework that uses vision-language models to analyze video by answering prompts about salient visual features converted into text, aiming to aid automated surveillance and post-incident investigations. The authors test four state-of-the-art models and introduce the Composite Model Utility Score (CMUS) to evaluate performance on largely unlabeled data. According to the abstract, the most reliable model flagged potentially relevant activity in over 66% of videos while cutting review time by more than 85%.
Saverio Cavasin, Pietro Tedeschi, Mattia Tamiazzo et al.
In briefEarly Signs of Diffusion Model Memorization via Basin Geometry and Cyclic Denoising
SummaryThis paper shows that diffusion models encode memorization in the geometry of their learned energy landscape before generated samples reproduce training images, a state the authors call latent memorization. Using score divergence and basin volume, they find localized basins form around training samples and separate them from held-out samples before the first memorized sample appears. They probe these basins with cyclic denoising, which repeatedly applies partial noising and denoising, and report results on CelebA, CIFAR-10, and Stable Diffusion v1.4.
WhyIntroduces latent memorization detected via energy-landscape basin geometry and cyclic denoising, backed by a proof and experiments on CelebA, CIFAR-10, and Stable Diffusion.
Nikhil Verma, Siddharthan Dileep, Anoop Singh et al.
In briefPerception-based, meaning-aware benchmark for judging speech-driven gesture generation
SummaryThis paper presents a benchmark for evaluating co-speech gesture generation models that combines standardized comparison, human-validated metrics, and semantic evaluation. The authors curate 13 objective metrics and propose Semantic Gesture Preservation (SGP), which measures how well semantic gestures from the ground truth are retained, using annotations on the BEAT2 dataset produced with a multi-modal LLM. A perceptual study with 101 participants found that SGP aligns with speech-related human judgments, while Semantic Score does not, and that composite metrics better match human ratings. The results indicate that objective metrics need validation against subjective evaluations.
In briefσTransfer: Moving Uncertainty Settings from Small to Large Networks under μP
SummaryThis paper addresses the difficulty of choosing the prior precision for Laplace approximations in large neural networks, where selecting it normally requires an expensive search. Under the Maximal Update Parametrization (μP), the authors derive a rescaling of the prior covariance that keeps the selected precision stable as model width grows. They propose σTransfer, which picks the precision on a smaller model and transfers it zero-shot to a larger one, and verify it on regression, image classification, and Transformer readouts.
Richard Bergna (1, 2), Fernando Ruiz Mazo (1) et al.
In briefAutoRegressive Retriever: Using Item Feedback to Improve Query Understanding in Universal Multimodal Retrieval
SummaryThis paper introduces the AutoRegressive Retriever (ARR), a multimodal retrieval model that alternates between retrieving an item and updating the query embedding before ranking the collection. It is trained with stepwise contrastive fine-tuning and reinforcement learning that uses the final reciprocal rank of a relevant item as the signal for selecting feedback items. The authors report that ARR outperforms the compared baselines on average on both in-domain and zero-shot benchmarks.
In briefEvi-VN: Hard-region-guided virtual node evidence injection for GNN-based fraud detection
SummaryThis paper observes that several graph neural networks make overlapping errors on the same hard samples, suggesting latent fraud evidence that graph structure and standard features miss. The authors propose Evi-VN, which uses feature-isolated evidence chains across structured, text, image, and audio data to correct these shared blind spots. The evidence is applied selectively to likely hard samples through virtual class nodes, leaving reliable predictions and existing GNN inputs unchanged. Experiments cover bot, fake-review, refund-evidence, and telecom-fraud tasks.
In briefStrategy-Selecting Game AI for Real-Time Strategy Games
SummaryThis paper proposes a two-level agent for the MicroRTS real-time strategy game: a Thompson-sampling bandit selects strategic commands based on an estimate of the opponent's strategy from in-game observations, and a constrained command-conditioned PPO policy executes unit-level actions. In a controlled comparison with a flat PPO baseline using the same architecture, budget, curriculum and self-play league, the system won significantly more often against three of the four strongest training-map opponents, including the two strongest held-out ones. No significant difference was found against the remaining opponent.
Nick Leenders, Roy Lindelauf, Joost van Oijen et al.
In briefLamarck's Driving School: Evolutionary Competition for Self-Driving Training Strategies
SummaryThe paper proposes a Lamarckian evolutionary framework that trains self-driving systems by letting candidate training scenario distributions compete, instead of relying on surrogate criteria such as realism or difficulty. The outer level uses Darwinian crossover, mutation, and selection, while the inner level learns policies and passes acquired capabilities to later stages. The authors distill recurring stage-wise patterns into a lightweight reusable strategy. Experiments report performance loss reductions of up to 25.07% for the full framework and 19.13% for the lightweight strategy.
In briefDIAL-OPD: Learning more from fewer tokens in on-policy distillation
SummaryThis paper studies token selection in on-policy distillation (OPD), where a student model learns from teacher signals on its own generated text. The authors find that training on fewer tokens can outperform using all tokens. They propose DIAL-OPD, which scores tokens by combining the log-ratio of teacher and student probabilities with the logarithmic mean of their probabilities, and keeps only the highest-scoring tokens. Across 4 teacher-student pairs and 7 math reasoning benchmarks, keeping 40% of tokens, DIAL-OPD outperforms Vanilla OPD and its full-token variants.
In briefDisFace3DNet: Explaining Facial Attractiveness Scores via 3D Component Breakdown
SummaryThe paper proposes DisFace3DNet, a model that predicts facial attractiveness by separating faces into seven components such as shape, skin, hair, and expression, learning their scores from overall ratings without human component labels. A constrained fit combines these component scores into the overall rating, showing each component's numerical contribution. On SCUT-FBP5500, it reaches a Pearson correlation of 0.8904 ± 0.0063 with average human ratings across five folds. Skin, hair, and facial shape show the largest variation, and human evaluation supports these score directions, though agreement for expression is weaker.
In briefHarness Evolution Hits a Ceiling: When to Train Weights for LLM Agents
SummaryThe paper studies two ways to improve long-horizon LLM agents: evolving the harness around a frozen model, or training its weights. It proposes reading failure composition to choose between them, since process failures are repaired by the harness while content failures need weight training. On DeepPlanning and WebArena-Lite, LoRA adapters trained on evolved-harness trajectories internalise the harness gains, with a shuffled-answer placebo adapter performing worse than the base model.
WhyProposes a failure-composition diagnostic for choosing harness versus weight training, validated with self-evolving harnesses, LoRA adapters, placebo controls, and two benchmarks.
In briefTabula Rasa: Monte Carlo estimation of unit-variance noise with controlled spatio-temporal correlation
SummaryThis paper proposes a method for generating time-varying Gaussian noise whose variance and temporal correlation can be controlled. Such noise supports downstream tasks that require temporal control and temporal coherence. The core idea frames the problem as joint Monte Carlo estimation of pixel reconstruction and variance, using the database-literature concept of sketching. The authors report that the method enables temporal control in downstream tasks with simpler and faster code than previous approaches.
In briefTwo-stage multimodal learning for navigating toward moving sounds using sight and hearing
SummaryThis paper addresses dynamic audio-visual navigation, where an agent must find sound sources that move, using both vision and hearing. The proposed two-stage framework first aligns sound intensity with visual geometry through an Intensity-Calibrated Geometry Encoder, then fuses multimodal features with a Transformer-based Cross-Modal Scene Aggregator. Experiments on Replica and Matterport3D reportedly achieve state-of-the-art success rates and better generalization to unheard sound categories.
In briefLocal-LLM interviewer that adapts questions to participant expertise, with traceable evidence
SummaryThis paper presents an interviewer architecture built on a locally hosted large language model that adjusts question depth throughout a conversation based on participant responses. The system profiles each participant's expertise in real time to generate suitable questions and transition messages. Evaluated with 246 participants, the expertise profiling module achieved 78.9% exact agreement with self-reported expertise (weighted Cohen's K 0.80), and participants rated relevance, engagement, and satisfaction highly.
Aisvarya Adeseye, Jouni Isoaho, Adeyemi Adeseye et al.
In briefSkillContrast: Difference-based text selection for reranking agent skills
SummarySimilar agent skills often share instructions but differ in when they should be used. SkillContrast is a training-free selector that compares retrieved skills and keeps their differing text with local context for a pretrained reranker. On 1,235 requests from SameCapRisk-Bench, it gave 54 to 72 more clean hits than TF-IDF query selection at identical input lengths. It also used 51.1% to 58.8% fewer model-input tokens than full skill bodies.
In briefOne Skill Too Many: How Co-Installed Similar Skills Conflict in Coding Agents
SummaryThis paper studies conflicts between similar skills installed together in coding agents, where the model chooses between them using only their names and descriptions. When a similar skill runs instead, the installed skill can lose core functions, such as a ban on touching git, while the task still passes, so benchmarks that only check task completion miss these cases. Using snapshots of 20,947 repositories and experiments on three models with 312 confirmed skill pairs, the authors find such conflicts are common.
WhyFirst systematic study of co-installed skill conflicts, validated across 6,368 runs and 20,947 repositories, with a pre-tool hook fix demonstrated.
In briefSyllable-Aware Tokenization for German Speech Recognition Across Domains
SummaryThis paper compares three tokenizer types for German end-to-end speech recognition: a multilingual character set, byte-pair encoding over spelling, and phonologically informed units based on syllabification and grapheme-to-phoneme conversion. Across 40 fine-tuning runs, all tokenizers perform similarly on in-domain read speech. Under domain shift, syllable-aware tokenization helps more at small vocabulary sizes, especially on dialectal and spontaneous speech. The authors also find that the acoustic encoder dominates the error patterns.
Christopher Witzl, Tobias Bocklet, Korbinian Riedhammer
In briefCacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration
SummaryThis paper studies how to speed up diffusion transformer inference by reusing cached features instead of full computation, without any training. The authors find that the placement of exact-computation steps matters far more for fidelity than the complexity of the reconstructor. They propose CacheScope, which normalizes probe-output change by input trajectory displacement to detect high-risk reuse steps online and falls back to full computation there. The method reportedly achieves up to 3.57x acceleration and reduces median LPIPS by up to 70.1% over the strongest probe-based baseline on FLUX.1-dev, Wan2.1-T2V-1.3B, and Open-Sora 1.2.
In briefRevisiting Handcrafted Minutiae Detection: A Simple, Effective Open-Source Fingerprint Baseline
SummaryThis paper introduces SBMEX (Skeleton-Based Minutiae EXtraction), a fast, deterministic fingerprint minutiae detection method integrated into the open-source pyfing package. It uses a dual look-up table design to speed up Crossing Number computation and skeleton tracking, and adds a continuous quality score based on path length, ridge-valley skeleton fusion, and spatial density decay. On NIST SD302, it reports accuracy comparable to or better than traditional open-source baselines without fine-tuning, with much lower detection latency than classical Crossing Number Python implementations.
In briefUXBench Pro: A Benchmark for Personalized User Experience in Multi-Turn Dialogue
SummaryThe paper presents UXBench Pro, a benchmark of 1,000 test instances drawn from real user interactions across 12 task scenarios and 82 domains, with each instance paired with a seven-facet user profile. It combines a personalized User Reward Model for third-person judgment with Sim4Eval, a user simulator that supports multi-turn interaction and first-person evaluation. Two meta-benchmarks, URMBench and USimBench, are introduced to test how faithfully these evaluators reflect real human preferences and behaviors.
In briefNeural networks that estimate how fast a person's arm gesture is performed, for controlling robots
SummaryThe paper benchmarks eighteen neural network architectures on ten abstract sequence tasks and identifies four compact models (BiGRU, TCN, Conv1D, GRUReLU) that perform consistently well. Three of these are then applied to estimating the speed of dynamic arm gestures from skeletal keypoint sequences captured with OpenPose. The best configuration reaches a mean absolute error of 0.198 on the peak-count speed measure, roughly 5% relative error.
In briefMinimax Gaussian Mechanisms for Continual Machine Unlearning: a privacy-based method for updating models under sequential deletions
SummaryThe paper studies machine unlearning, which updates a trained model after records are deleted so it approximates exact retraining without repeating full training. The authors design Gaussian mechanisms for Newton updates under sequential deletion requests and use Gaussian differential privacy with adaptive composition to show that the released model sequence is hard to distinguish from exact retraining. They derive error and noise-variance bounds for independent and random-walk Gaussian noise and evaluate them with simulations and a credit default data analysis.“
In briefBeyond Action Entropy: Quotient-Space Exploration for Genome-Scale Metabolic Model Repair
SummaryThis paper addresses repairing genome-scale metabolic models (GEMs) using functional observations, where many different reaction edits can represent the same biological mechanism. The authors propose QuotientPO, which merges equivalent repairs into canonical mechanisms and explores directly in this quotient space, along with a kernelized Rényi estimator that measures graded crowding among repair cores. On 2,212 held-out GEMs, Success@32 rises from 17.93% to 20.10% (12.1% relative), with more distinct successful repair cores found under the same sampling budget.
In briefBuilding structured decision sources for consensus-based pseudo-label learning
SummaryThis paper studies how to make multiple predictors supply genuinely different evidence so that consensus-based pseudo-label learning becomes more reliable. The authors build decision sources by controlling within-class structure, varying center granularity and neighborhood mixing, checking each source's stability, and selecting a complementary subset via node pair coassignment. On Cora, CiteSeer, and PubMed, the constructed sources raise fixed-budget pseudo-label precision by 1.19 to 4.39 points over three conventionally initialized GCN sources, suggesting source construction matters more than model count.
In briefWhere to Adapt Matters: Layer-Selective Fine-Tuning to Retain Capabilities
SummaryThis paper studies how fine-tuning large language models can erode general abilities, and asks which Transformer layers to adapt. It uses layer-wise empirical Fisher information to measure sensitivity and proposes input-output cosine similarity as a cheaper, forward-only proxy. Based on this, the authors introduce LS-LoRA, which places LoRA adapters only in layers with low input-output similarity. On mathematical reasoning and code generation, LS-LoRA improves average target-task performance and retains more commonsense reasoning than standard all-layer LoRA.]
In briefModel-free policy gradients for mean-field control using randomized transport maps
SummaryThis paper proposes Transport REINFORCE, a model-free policy gradient method for discrete-time mean-field control. Standard REINFORCE estimators capture only the policy's effect through controlled dynamics, missing its effect through the population distribution. The method perturbs the population distribution, via simplex mixing in finite spaces and Gaussian-mixture-preserving transport maps in continuous spaces, to estimate this missing contribution. The authors prove consistency of the perturbed objective and gradient and derive bias and mean-square error bounds for the sample-based estimator. Experiments on several benchmarks show improvement over standard REINFORCE.
In briefTAM: Task-aware memory distillation for efficient spatiotemporal prediction
SummaryThis paper proposes TAM, a framework that stores a frozen teacher model's knowledge in a bounded, retrievable memory for spatiotemporal prediction distillation. The student either matches the teacher's similarity distribution over shared references or regresses observation-conditioned residual prototypes. TAM is used only during training and leaves student inference unchanged. Across video prediction, weather forecasting, and traffic flow prediction, adding TAM improved SSIM on all six video datasets and reduced MSE on five relative to the corresponding KD baselines.
In briefAgentEvolver: System-Wide Self-Evolution Through Task Execution
SummaryThis paper presents AgentEvolver, a system that lets an AI agent develop reusable capabilities while it executes tasks, with the foundation model kept fixed. Eight entity families, such as operations, methods, and agents, are revised through a shared versioned lifecycle. The authors report an 82.08% resolution rate on SWE-bench Pro Public with evolution, above their baseline without it, and six application cases show retained capabilities reused later, along with some incomplete objectives and a failed strategy.
In briefPointVGGT: Zero-Shot Multiview RGB-D Point Cloud Registration Using Visual Geometry Foundation Priors
SummaryThis paper introduces PointVGGT, a training-free method for aligning multiple unordered RGB-D scans into a single metrically consistent coordinate frame. It uses a visual geometry foundation model such as VGGT to estimate global poses, grounding its scale-ambiguous predictions against metric depth, then refines them via voxelized spatial hashing for dense correspondences and robust bundle adjustment. The authors report strong zero-shot registration accuracy and efficiency on indoor, object-centric, and outdoor datasets.
In briefBeyond Report Imitation: Clinically Aware Multi-Image Ultrasound Report Generation
SummaryThis paper addresses generating ultrasound reports from multiple images, noting that archived key frames capture only part of a dynamic exam, so imitating original reports misaligns with visual supervision. The authors propose CAMEO, a three-stage framework covering ultrasound vision-language learning, cross-view evidence grounding, and clinically aware preference alignment. On the primary USReport-Distilled benchmark, CAMEO improves BLEU-1, ROUGE-1, METEOR and ClinicalScore over EchoVLM.
In briefS³Geo: A method that matches drone and satellite images to find locations more reliably.
SummaryThis paper presents S$^3$Geo, a framework for cross-view geo-localization, which estimates locations by matching images taken from drone and satellite viewpoints. It combines a Decoupled Query Pooling module for region-aware structural features, an optimal transport-based query-level contrastive learning scheme for soft cross-view correspondences, and semantic distillation from a frozen CLIP teacher. The authors report that it consistently outperforms state-of-the-art methods on University-1652 and SUES-200 without increasing inference complexity.
In briefNanoProof: Open and Efficient Automated Theorem Proving in Lean 4
SummaryThe paper introduces NanoProof, described by the authors as the first factorized execution-guided theorem prover for Lean 4 whose training data, tooling, pipeline, and weights are all released. The authors also release a dataset of structured proof trees and a tool for programmatic interaction and data extraction within the Lean 4 verifier. NanoProof reaches 50.8% pass@16 on MiniF2F-Test, which the authors report with far less compute than comparable systems and AlphaProof.
In briefModeling how errors spread to pinpoint the root cause of failures in LLM multi-agent systems
SummaryThis paper addresses the problem of finding which agent and step caused a failure in LLM-based multi-agent systems, defined as the decisive error whose correction would recover the failed run. The proposed method, EMFA, builds a structured representation of the failed trajectory, models cascading error propagation and persistent interaction loops, and uses propagation-aware screening followed by counterfactual verification. On the Who&When benchmark, EMFA achieves state-of-the-art step-level attribution accuracy and remains competitive at the agent level.
In briefLLM version turnover weakens screening of AI-assisted scientific writing
SummaryThe authors paired 4,000 pre-ChatGPT PNAS abstracts with rewrites from 23 LLM versions released between June 2023 and August 2026, then tested how detectors hold up as models change. Detectors trained only on past versions can fail at generation boundaries: calibrated to flag 1% of human abstracts, they caught over 99% of rewrites just before the sharpest boundary but only 3.8% just after. The paper argues that detector benchmark accuracy should be treated as provisional and re-verified with every LLM release, including earlier versions.
WhySystematically quantifies how LLM version turnover breaks detectors using 23 versions and 4,000 abstracts, revealing collapse at generational boundaries with direct research-integrity policy implicati
In briefRunnable commit untangling to help coding agents organize and repair large patches
SummaryCoding agents often produce large patches mixing several purposes, which are hard to review. This paper proposes RucTangle, a method that splits such patches into ordered commits while keeping the code runnable after each one, and TangleEval, a framework measuring how untangled histories help agents fix bugs. On 131 agent-generated patches, RucTangle's histories were all runnable, whereas baselines produced 20.6% to 37.4% unrunnable histories. Adding RucTangle histories to agent context improved pass@1 by 5.2% absolute on regression repair.
WhyProposes the first agentic untangling keeping each commit runnable, with an evaluation showing 5.2% pass@1 gain on 453 regression patches.
In briefTesting whether language models can reason deeply over long contexts using Prolog logic puzzles
SummaryThis paper examines whether large language models can perform deep deductive reasoning over long contexts, rather than just retrieving information. The authors built ProloNg, a synthetic benchmark in Prolog whose problems vary in reasoning depth, with the hardest case at depth 22 and 62k context length. Testing 8 reasoning models from 5 frontier LLM families, they find performance drops substantially as reasoning depth grows, with most models near chance beyond depth 10.
In briefAssessing whether AI chatbots give culturally fitting maternal health advice in India
SummaryThe paper introduces MH-INDIC, a framework for judging whether maternal-health conversations with large language models fit the cultural context of urban and semi-urban North India, using ten dimensions of maternal-health reasoning. A 26-item survey of 102 pregnant and postpartum women was used to evaluate ten LLMs. Several models roughly matched the overall human population distribution, but all varied far less across demographic and household profiles than the human group did. Using human-grounded prompts to condition generated dialogues produced stronger profile alignment and higher quality ratings.
In briefAdapting English Quality Classifiers for Multilingual LLM Pretraining Data Selection
SummaryThis paper proposes a method to convert an existing English data-quality classifier into a multilingual version covering over 100 languages. It trains a small multi-layer perceptron on Transformer encoder embeddings, using multilingual text as input and scores from the English classifier on machine-translated text as labels. Experiments at 1B, 3B and 8B scales show it matches existing multilingual filtering baselines on downstream LLM benchmarks without harming regional and cultural knowledge benchmarks.
Vinko Sabol\v{c}ec, Bettina Messmer, Yassine Turki et al.
In briefEmbedding bias in conditional independence tests: why replacing text or images with embeddings can lead to false rejections
SummaryThis paper studies conditional independence testing when an embedding of a text or image replaces the original variable. The authors show that if information lost by the embedding remains correlated with the variables, the rejection probability under the null can approach one. They propose a robust test that keeps its level approximately under a declared tolerance, and report results on synthetic data and text embeddings.
In briefSDPAD: A Fully Spike-Driven Pipeline for End-to-End Self-Driving Planning
SummarySDPAD is a fully spiking neural network pipeline that plans vehicle trajectories end to end for autonomous driving, converting a pre-trained perception model into integer-spike form. On the nuScenes open-loop benchmark it reports an average L2 error of 0.40 m and a 0.12% collision rate while using 69.9 mJ of energy, under 2% of recent ANN baselines. On the NAVSIM closed-loop test it scores 86.3 PDMS, ahead of the earlier spiking planner SAD by 4.3 points.
WhyProposes the first fully spike-driven end-to-end planner, matching ANN accuracy on nuScenes and NAVSIM at much lower energy; impact is mainly edge autonomous driving.
In briefUncertainty-aware trajectory prediction for physics-aware highway vehicle forecasting
SummaryThis paper proposes uncertainty-aware extensions of X-TRACK, a physics-aware framework for predicting vehicle trajectories in safety-critical settings such as autonomous driving. The method models aleatoric and epistemic uncertainty in predicted motion variables and propagates it into trajectory space, then applies conformal prediction to the trajectory covariance to reach a target marginal coverage. On the highD dataset, X-TRACK-DE improves accuracy over the deterministic baseline, and both uncertainty-aware variants can be conformally calibrated to the desired coverage level.
In briefSV-TAD: Native Sparse Convolutions for Efficient Temporal Action Detection
SummaryThis paper addresses the inference cost of adapting large Vision Transformers to long videos by introducing native sparse 2D convolutions that operate directly on pruned token sets. Integrated into the SV-TAD framework for temporal action detection, it reduces VideoMAEv2-L computation by up to 64% and achieves 2.2x faster inference. It maintains state-of-the-art accuracy on THUMOS-14 and ActivityNet-1.3.
Ricardo Pizarro, Roberto Valle, Jos\'e M. Buenaposada et al.
In briefChronos: Helping Code Agents Reason Over Software Evolution History
SummaryThe paper introduces Chronos, a test-time framework that turns merged pull requests into structured experience cards linked by a typed graph of code, developer-intent, and organizational relations. LLM-based code agents use semantic search and multi-hop graph expansion to retrieve relevant past changes, and the same memory guides patch generation and selection. On SWE-Bench Verified, the full workflow raises the mean resolution rate of SWE-Agent across six LLM backbones from 69.2% to 72.9%, reaching 79.8% with MiniMax M2.5. In a human evaluation, graph-grounded retrieval increased the mean number of useful cards per task from 1.24 to 2.87 compared with flat semantic retrieval.
In briefCoCam4D: Geometry-Aware Cooperative 4D Perception for LiDAR-Free Driving
SummaryAutonomous vehicles often have limited perception due to occlusions and sensor range. This paper proposes CoCam4D, a Bayesian collaborative perception framework that models geometric uncertainty and shares compact 3D Gaussian primitives so vehicles can reduce each other's depth uncertainty without LiDAR. It also introduces 35-byte Dynamic Object Primitives for efficient C-V2X communication. The authors report gains of 11.48% on OPV2V+ and 10.62% on DAIR-V2X-C over recent vision-only methods.
In briefSmoothing Sparse Expert Selection in Mixture-of-Experts Models
SummarySparse Mixture-of-Experts models use a fixed top-k rule to pick experts, which creates a hard cutoff that separates experts with similar scores into fully trained and untrained groups. The paper proposes Elastic Expert Routing, which randomly samples the number of active experts from a discrete distribution centered on k, gradually softening this threshold during training. The method keeps the expected compute cost and inference budget unchanged. It improves downstream macro-averages by 0.84 and 2.02 points in fine-tuning on OLMoE-1B-7B and Qwen3-30B-A3B, and in from-scratch pretraining it lowers validation LM loss by up to 0.04 and beats the static top-k baseline by 1.6 points on average.
In briefMemory Types Differ: Using Diverse Strategies to Improve Long-Term Memory for LLM Agents
SummaryThis paper argues that existing retrieval-based memory methods for LLMs treat all memories the same way, which limits performance. The authors introduce TriMEM, a multi-class memory dataset with type annotations, and MemoType, a framework that uses a learned router to classify memories and queries and apply matching retrieval strategies. Experiments on three datasets show MemoType outperforms existing methods, with up to a 16.18% gain in Recall@1.
In briefPAM-ToD: A plug-in that adapts 3D Gaussian Splatting road scenes to different times of day
SummaryThis paper proposes PAM-ToD, a lightweight plug-in that adapts a pre-trained 3D Gaussian Splatting road scene to a new time of day using a few anchor images, while keeping the original parameters fixed. It learns color corrections that scale existing colors and add brightness, such as street lamps at night, without separately estimating albedo and illumination. The authors also introduce CARLA-ToD, a benchmark with matching geometry, camera poses, and moving-object trajectories across three times of day. The paper reports higher PSNR and lower LPIPS than baselines in static and dynamic settings.
Kota Shimomura, Sungho Moon, Tsubasa Hirakawa et al.
In briefLooped Transformer model that keeps reasoning stable across tens of thousands of test-time iterations
SummaryThe paper argues that looped Transformers degrade as iterations increase, because noisy state updates overwrite correct intermediate results. The authors propose InfiLoop, a loop-native residual connection that uses content-based weighting and learned temporal decay to decide which past computations to keep. Its memory stays constant as the loop count grows. A 7M-parameter model reportedly reaches 97.9% exact accuracy on Sudoku-Extreme and 13.6% pass@2 on ARC-AGI-2.
In briefSera: Combining Knowledge and LLM-Based Degradation Semantics to Forecast Battery Health More Reliably and Interpretably
SummaryThis paper proposes Sera, a framework for forecasting battery state of health (SoH) that adds degradation semantics to temporal models. It builds two representations from time series using rule-based domain knowledge and LLM-based interpretation, then integrates them with temporal features through gated aggregation. Experiments across multiple prediction horizons and temporal models show consistent accuracy gains, with up to a 37.3% reduction in prediction error. Counterfactual analysis indicates forecast responses align with the meanings of key degradation descriptors.
In briefIncremental Open-Ended Deep Research with Structured Harness: a method for continuously updating research reports
SummaryExisting open-ended deep research systems usually write reports from scratch, which is inefficient when reports must be kept current as new information appears. The paper proposes Incremental-OEDR, which treats a report as an evolving research state that keeps valid knowledge, revises outdated content, and adds new information. It also introduces Structured Harness, which represents reports as structured outlines, sections, and evidence to support selective updating and evidence reuse. Experiments report competitive report quality with better continuity and lower research cost.
In briefWorkerville: An organizational behavior view of agent safety
SummaryThis paper proposes organizational behavior (OB) as a framework for studying the safety of advanced AI agents, and formalizes counterproductive work behavior from human workplace research as Agentic Counterproductive Behavior (ACB). The authors introduce Workerville, a controlled benchmark that applies 16 organizational configurations to 210 tasks, yielding 3,360 challenges, and evaluates six frontier LLMs. They report that combined negative organizational antecedents amplify behavior non-monotonically, with the unauthorized-disclosure rate rising from 16.5% to 60.1% under two antecedents and falling to 50.3% under three.
In briefSelf-supervised parallel MRI reconstruction using Hankel subspace learning
SummaryThis paper addresses parallel MRI reconstruction from undersampled data, an ill-posed inverse problem. It proposes HSSRecon, a scan-specific framework that splits data by physical acquisition units before Hankel lifting and applies multiplicity normalization to avoid data leakage in self-supervision. The network learns a compact complex-valued Hankel subspace operator, and reconstruction is solved over the original k-space variables with a conjugate-gradient solver enforcing hard data consistency. On fastMRI brain data with three contrasts and three sampling masks, the method reports competitive PSNR, SSIM, and NMSE across six aggregated conditions.
