In briefPDE-JEPA: Predictive learning of latent dynamics for parametric PDEs
SummaryThis paper studies predictive representation pretraining as an alternative to reconstruction-based learning for parametric partial differential equations (PDEs). It introduces PDE-JEPA, which trains an encoder with masked-latent prediction, aligns latent trajectory geometry with physical field evolution via a geometry projector, and uses a physics-structured predictor that separates parameter-independent evolution from parameter-dependent response. Across nine PDE benchmarks, the authors report average gains of 33.4% in-distribution and 51.4% when extrapolating to unseen governing parameters.
In briefRSI-Router: Subtask-level LLM routing and self-evolving skills for lower-cost agents
SummaryThis paper presents RSI-Router, a framework that assigns subtasks to different language models and develops model-specific skills through recursive self-improvement. It routes between DeepSeek-V4.1-Flash and Qwen3.5-9B. Across five agentic benchmarks, it reportedly outperforms a DeepSeek-only baseline at about 48.3% of the inference cost, and it establishes a stronger performance-cost Pareto frontier than nine other routing methods.
In briefLearning how messages spread in graph networks through feedback
SummaryGraph neural networks benefit from learning local geometry that controls how neighborhood information is compared and combined. This paper proposes GeoF, a recurrent framework that jointly updates node features and propagation geometry through message-passing feedback. Each node keeps a symmetric positive-definite geometry that sets neighborhood weights and maps messages into its local coordinates before aggregation. Experiments on benchmark datasets show that GeoF consistently outperforms state-of-the-art GNN baselines.
In briefFromPitch2Board: Benchmarking LLM Agents in Long-Horizon Football Management
SummaryThis paper introduces FromPitch2Board, a deterministic football-management benchmark that tests five configurable factors through controlled comparisons on one simulator. The authors evaluate four foundation models and four agent scaffolds. They find that GPT-5.6's responsibility ladder rises with recruitment but falls under full management, with its skipped-decision rate increasing from 1.1% to 57.9% across the final responsibility boundary.
In briefResonAct: Streaming Metrics for Self-Healing Multi-Agent Systems
SummaryThis paper presents ResonAct, a runtime self-healing framework that continuously monitors multi-agent systems using streaming operational metrics. It diagnoses failures and localizes root causes, then selects remediation policies, operating as an external control plane without changing application agents. Evaluated on enterprise workflows and AppWorld benchmarks, policy-driven remediation improved task completion rates by up to 10.00 percentage points.
Tarun Chintada, Neelamadhav Gantayat, Ishaan Romil et al.
In briefTriangular Resampling for Stable Long Motion Generation
SummaryThis paper proposes Triangular Resampling (TR), a post-training method that reduces error accumulation in motion diffusion models over long sequences. It replays model-generated partially denoised states during training and clamps them toward ground truth to limit drift. On 120-second motion generation from HumanML3D test prompts, supervised TR reduces FID AUC by 40.9% and FID degradation slope by 55.3% relative to matched post-training without replay.
In briefWhether zeroth-order reward shaping suffices for policy gradient in stabilization control
SummaryThis paper studies which information a reward function must contain for deep reinforcement learning in robotic stabilization tasks, comparing configuration (zeroth-order) and velocity (first-order) terms. The authors find policy gradient methods can solve stabilization without first-order reward terms, and that adding them can cause severe sensitivity as their scale grows. They conclude rewards must be zeroth-order complete over goal-relevant coordinates, while first-order state remains necessary in the policy observation.
In briefPredicting Dutch Train Delays: Explainable AI Tests of Tree-Based Models
SummaryThis paper uses interpretable tree-based ensemble models on the Dutch railway network to predict train delays from topological, weather, and operational features. Prediction is accurate within the same month but systematically worsens when tested on future months. SHAP and dispersion analyses link this decline to volatile environmental features and an unstable statistical target definition. The authors conclude that richer features alone do not solve long-term forecasting, and that season-aware, dynamic models are needed.
Jia Long Bao, Ali Mohammed Mansoor Alsahag, Seyed Sahand Mohammadi Ziabari
In briefUsing LLMs to Spot AI-Written Text: A Cross-Generation Study
SummaryThis paper tests 15 large language models from three generations as both text generators and detectors of machine-written text. Using 1,000 human-written texts and 15,000 model-generated texts, the authors collected over 233,000 binary judgments with explanations. Detection performance depends mainly on the detector's capability rather than the generator's origin, though text from newer models is harder to detect.
In briefUni-TMPO: Matching Trajectory Distributions to Keep Diverse Image and Robot Policies
SummaryThis paper proposes Uni-TMPO, a reinforcement learning post-training framework for diffusion and flow policies. Instead of maximizing expected reward, it converts rewards into a target distribution over trajectory groups and matches the policy to it using forward KL optimization, aiming to avoid mode collapse. Experiments report higher text-to-image rewards with better diversity and efficiency, and improved generalization of vision-language-action models to held-out tasks and scenes.
WhyReplaces reward maximization with forward-KL matching to a trajectory target distribution, reducing mode collapse across T2I and VLA, with experiments including real-robot tests.
In briefVCN-Bench: A Video-Based Navigation Benchmark for Spatial Reasoning from Prior Visual Experience
SummaryThis paper introduces VCN-Bench, a benchmark for testing whether multimodal large language models can use prior video experience to reason about space and navigate closed-loop. Agents watch a prior video covering the start and destination, infer the instructed target, and navigate toward it. Built on Matterport3D, it includes five instruction types, 100k training episodes, and 1,250 evaluation episodes. Experiments show limited navigation performance and frequent failures even when the destination is correctly identified.
In briefJailbreak Context Lingers: How Tool Agents Route Differently After Safety Feedback
SummaryThis paper studies how lingering jailbreak context affects what tool-using AI agents do after receiving safety feedback. Across 192 parent tasks in 42 domains, the authors analyzed 12,148 continuation pairs over eight agents. Identical safety feedback led to different outcomes by model, including rescue, persistent unsafe execution, and over-refusal of benign tasks. Activation patching located causal effects in the final layers, and features from these layers showed moderate predictive signal for full-trajectory outcomes on unseen tasks.
In briefHigh Prediction Accuracy May Hide Weak Responsiveness to Controlled Target States in VLA Readouts
SummaryThe paper argues that accurately decoding object states from vision-language-action (VLA) model representations does not show that predictions faithfully respond to changes in the physical target state. The authors propose an evaluation framework that separates prediction accuracy, target-state responsiveness, and context stability, using physically validated observations that cross target coordinates with robot contexts. They show that high accuracy on natural trajectories can coexist with weak controlled target-state responsiveness, and that comparisons and interventions reveal the three properties provide distinct diagnostic information.
In briefAgentPerfBench: A Benchmark Suite for Inference Performance of Agentic LLMs
SummaryThe paper introduces AgentPerfBench, a benchmark suite for measuring the inference performance of agentic large language models, which perform multi-turn tasks such as coding and tool use. It draws on real traces from SWE-Bench and TerminalBench alongside standard chat baselines, and generates synthetic workloads from empirical distributions of input length, output length, and turn count. The authors argue that existing benchmarks miss realistic context-length growth and do not operate at hardware saturation, and they use Nsight Compute traces to build a multi-dimensional roofline model. The suite also includes scripts to identify bottlenecks on emerging hardware.
Cheuk Hang Lau, Zeyu Cao, Kevin Wong Cheuk Yin et al.
In briefV-Gym: Co-evolving skills and data to improve agentic visual reasoning
SummaryThis paper introduces V-Gym, an autonomous framework that iteratively co-evolves procedural skills and multimodal practice data drawn from agent execution trajectories. Skill updates are kept only when they improve validation performance, while new practice data targets bottlenecks found in trajectories and is added after quality checks. The authors report substantial gains over baselines across several multimodal reasoning benchmarks and backbone models, with evolved skills generalizing across domains and models.
In briefSOLAR: A state-driven online learning rate scheduler for LLM pretraining
SummaryThis paper proposes SOLAR, a framework that adjusts the learning rate online during large language model pretraining. It uses a base schedule as a reference and learns bounded, state-dependent residual corrections for each parameter group, with a Circuit-Breaker that restores training after rare unsafe actions. The authors report that SOLAR achieves lower final perplexity than tuned static schedules and automatic learning-rate tuners across dense models from 60M to 1B parameters and two MoE settings up to 3B.
In briefQuantForge: A System Discovering Residual Decompositions for MXFP4 Post-Training Quantization
SummaryThis paper introduces QuantForge, a system that uses LLM-driven program evolution to discover post-training quantization algorithms for MXFP4 W4A4 quantization of large language models. It records competing explanations, selects controls that distinguish them, and checks that revised code implements the conclusions. The discovered fixed quantizer, HiRes, achieves the lowest seven-model Robust Fit (0.09300) across seven tasks.
In briefFrom Preference to Reciprocity: Decentralized Matching Using LLM Agents and Learning
SummaryThis paper proposes a dynamic bipartite matching framework that combines LLM agents with contextual bandits in a simulated Chinese marriage market. Agents judge locally encountered candidates, while Logistic-UCB models learn whether proposals are reciprocated, without needing full market preference rankings. In a 50×50 experiment, Bandit-UCB achieved higher mean mutual welfare than Gale-Shapley and the fewest blocking pairs among LLM-based policies.
In briefFair Fact-Checking: Using RoSh to Narrow Cross-Lingual Gaps in LLM Truth Judgments
SummaryThe paper tests eight language models on 1,500 encyclopedic claims available in eight languages, finding that English is judged more accurately than other languages on every model, with the gap largest for the smallest models. The authors propose RoSh, a training-free method that applies a per-language shift and rotation to the residual stream at three layers. It improves every model and closes about 75% of the gap on average.
WhyShows models encode correct answers they fail to express, and proposes training-free closed-form residual rotation. Extensive multi-model, multilingual experiments and controls give strong evidence an
Muhammad Ahmad, Fatemeh Seyedin, Adrian Weller et al.
In briefLearning which memories to recall for world-model prediction
SummaryWorld models predict future observations from current experience and actions, and such predictions can depend on distant past observations. This paper proposes Future-Aware Recall (FAR), which trains a retriever using predictive utility measured by the conditional log-likelihood of the realized future, approximated by negative diffusion prediction loss, while remaining future-blind at inference. The retriever learns to weigh retrieval cues such as time, pose, vision, and audio for each query. Across three complementary settings, FAR outperforms hand-designed recall methods.
In briefFestDPO: Aligning Fast Generative Models Using Pairwise Preferences
SummaryFew-step generative models produce high-quality samples in a few evaluations, but their outputs may lack desired properties. This paper introduces FestDPO, an extension of direct preference optimization to few-step generators that uses nonparametric likelihood estimation from samples, making the DPO loss computable. The method is reported to outperform preference optimization baselines on text-to-image generation and protein backbone generation.
In briefEvidence-aligned multimodal self-distillation for detailed image understanding
SummaryThis paper addresses fine-grained visual understanding, where models must recognize small details in complex images. It introduces Evidence-Aligned multimodal on-policy self-Distillation (EAD), which uses a crop-conditioned teacher but weights its corrections by how well they align with a separate evidence reference. The reference is built from the change between the current student's predictions on a masked image and on the original image, which isolates the visual evidence's effect. According to the paper, EAD keeps only 6% of the supervision mass of dense OPSD and still consistently outperforms previous state-of-the-art methods.
In briefAgentic Federated Learning: rule-based agents on client and server sides for adaptive training
SummaryFederated learning trains models across distributed clients without sharing raw data. This paper proposes Agentic Federated Learning (AFL), which adds lightweight rule-based agents at both levels: a client-side agent adapts local training and assesses update reliability, while a server-side orchestrator performs quality-aware client selection and adaptive aggregation. Experiments on CIFAR-10 under IID, non-IID, and noisy-client settings report consistent gains over FedAvg and FedProx.</summary>
In briefHierarchical clustering and noise removal for directed graphs
SummaryThe paper represents a directed graph as a Hermitian matrix derived from its adjacency matrix, capturing both connectivity and edge direction. Using the spectral decomposition of this matrix, it introduces a k-means-based digraph clustering algorithm that is applied recursively to produce nested hierarchical partitions. The authors also build multilevel spline quasi-interpolants from in- and out-degrees of clusters to denoise graph signals. Experiments on synthetic and real-world digraphs report improved clustering and signal recovery measured by RMSE and SNR.
Yi Wang, Sippanon Kitimoon, Hrushikesh N. Mhaskar et al.
In briefTwo-Timescale Fine-Tuning Provably Learns New Features in Two-Layer ReLU Networks
SummaryThis paper theoretically studies fine-tuning of pre-trained models using a Gaussian multi-index setting, where one feature must be learned during fine-tuning and the rest were learned in pre-training. For two-layer ReLU networks, the authors show that two-timescale training, which updates outer weights much faster than hidden ones, learns the new feature while keeping the pre-trained features. The method requires only O(d) fine-tuning samples, regardless of the number of pre-trained features. With random initialization, the same number of samples is insufficient to recover the target parameters.
In briefNumerical Reliability of Differentiable Physics Optimization for Robotic Material Manipulation
SummaryThis paper examines how reliable gradients are when differentiable physics, using Material Point Method simulations of elastoplastic and granular manipulation, is used for robotic optimization. The authors report that GPU summation order can change long-horizon gradients and even reverse the sign of one parameter gradient. Finite-difference checks also become less reliable for longer rollouts, and the choice of observations and loss definitions alters optimization results. They recommend reproducible accumulation, validation that compares perturbation-induced loss changes with repeated-run variation, and explicit reporting of objective construction.
In briefCodoku: A Renewable Benchmark for Program Reasoning by Coding Agents
SummaryThe paper introduces Codoku (code sudoku), a benchmark in which solvers fill typed cells in partial programs so that global static and dynamic constraints, such as a specified control-flow graph and execution path, are satisfied. Because partial programs cannot be run and valid fillings are sparse, tool use and enumeration cannot replace program reasoning. Puzzles are synthesized from scratch with guaranteed solvability and controllable complexity. On 300 puzzles, five frontier models were evaluated, and even proprietary models solved only about half of the large ones.
WhyProposes a renewable program-reasoning benchmark with witness-guaranteed puzzles via semantic reification, addressing execution shortcuts; evaluated on 300 puzzles across five frontier models, moderat
In briefRewarding Novel Deductions: Solver-Guided Training for Logical Reasoning in Language Models
SummaryThis paper proposes SPRING, which uses an SMT solver during training to check each intermediate reasoning step of a language model. Steps that are logically valid, consistent, and make new inferences are rewarded, while contradictory or uninformative steps are penalized. Across three logic benchmarks (ZebraLogic, AR-LSAT, Knights and Knaves) and four models, SPRING outperforms base models, outcome-only reward baselines, and Logic-LM.
In briefCapField-OPD: Merging multiple specialized image-model teachers into one student with continuously adjustable capabilities
SummaryThis paper presents CapField-OPD, a framework that consolidates several specialized teacher models for flow-based generative models into a single student. It builds a continuous capability field from explicit capability coordinates, so capability invocation no longer depends on prompt semantics and can be adjusted at inference time. Experiments cover compositional generation, text rendering, and visual aesthetics.
In briefMinimax last-iterate convergence in matrix games with observed opponent actions
SummaryThe paper studies two-player zero-sum matrix games where payoffs are only revealed as bandit feedback but opponent actions are observed. The authors propose an algorithm achieving a duality gap of Õ(√(d/t)) with high probability at every round, improving the best prior dimension dependence by a factor of d^{3/2}. This rate matches a standard bandit lower bound, establishing minimax optimality up to logarithmic factors, and the method uses O(d) time and memory per round.
In briefA General Add-on Method That Improves Protein Fitness Prediction Models Without Retraining
SummaryThe paper introduces VenusREM-Harness (VRH), a model-agnostic, training-free retrieval-enhanced mutation harness that improves protein foundation models' fitness predictions. It fuses frozen model scores with multiple sequence alignment evidence weighted by model uncertainty, then applies background correction and shrinkage based on structural confidence and solvent exposure. Across 1,211 assays and 3.1 million variants from three benchmarks, all 71 configurations improved Spearman correlation by 0.073 on average. VenusREM2, built with VRH, reaches a ProteinGym Average Spearman of 0.556.
WhyProposes a training-free, model-agnostic retrieval-enhanced mutation framework fusing MSA and structural signals; gains of 0.073 Spearman across 1,211 assays suggest broad impact.
In briefOmniTide: Co-designed algorithms and systems for efficient on-device omni-modal LLM streaming
SummaryThis paper addresses the problem that streaming multimodal inference on devices steadily grows the KV cache, exhausting memory and compute. It proposes OmniTide, which pairs OmniPick, an algorithm that keeps critical cross-modal context based on unit boundaries and modality importance, with OmniPage, a system component that partitions the cache by retention likelihood and compacts surviving tokens. The authors report kernel speedups of up to 12.72x, stream-loop latency reduction of up to 2.40x, and accuracy gains of up to 18.0 points over sliding-window baselines on StreamingBench.
In briefUsing Waveform Images to Guide Retrieval in Medical Time Series Classification
SummaryMedical time series are used in many clinical classification tasks, but existing methods mostly treat them as numerical sequences and underuse the shapes visible in waveform plots. The paper proposes Vision-Informed Retrieval (ViRe), which uses a frozen vision-language model representation of waveform plots as a morphology-aware query to select relevant evidence from numerical features. ViRe reports a 6.42% overall relative improvement over the previous state of the art across six public benchmarks and outperforms ten baselines. Code is publicly available on GitHub..
In briefDirectUV: Image-guided UV texture generation using surface-aware positional encoding
SummaryDirectUV is a diffusion-based framework that generates UV textures for 3D meshes in the latent UV space of a pretrained image VAE, conditioned on a single input image and a coarse UV map. Its Surface-Aware Positional Encoding replaces the usual 2D grid positions with encodings derived from 3D surface coordinates, so attention follows 3D proximity rather than UV-grid distance. This is reported to improve coherence across seams and disconnected UV islands, with a multi-level extension for different surface granularities. The authors state that experiments show sharper, more globally consistent textures than baselines, with the largest gains in occluded and unseen-view regions.
In briefWhen Can Attention Heads Be Statically Defined?
SummaryThe paper studies attention heads that learn similar patterns across different inputs, and whether reusing those patterns can cut training cost. It proposes Selective Attention Freezing (SAF), which picks low-variance heads and replaces their attention weights with fitted post-softmax means halfway through training. Reported results include 1.056x faster optimiser updates after replacing 25% of heads in a 124M model with 4K context, at a 0.77% perplexity increase.
Weixian Waylon Li, Yintao Tai, Marcio Fonseca et al.
In briefSelf-evolving context policies that manage what long-running AI agents see
SummaryThis paper introduces ContextEvo, a framework that learns how an AI agent should manage its context during long tasks. It rebuilds the context the model sees at key decision points, finds failures linked to context, and makes targeted policy updates. Built on the open-source Pi-agent harness, it improves results on three long-horizon benchmarks and performs comparably to or better than Codex, OpenCode, and OpenClaw.
In briefSEmoEdit: Probing and Using Editable Pretrained Speech Flows for Emotion Changes
SummaryThis paper examines whether large pretrained text-to-speech models can edit speech emotion directly, without additional training. The authors analyze editing effects along the generative trajectory and propose SEmoEdit, a training-free method that treats emotion editing as velocity transport between source and target emotions. It supports emotion replacement, erasure, and continuous interpolation. The authors also release SEmoEditBench with 600 editing cases, where SEmoEdit outperforms training-based and activation-steering baselines.
In briefUNBIND: Removing Specific Memorized Code from Code LLMs at Inference Time
SummaryThis paper proposes UNBIND, a method that lets code language models selectively forget specific memorized implementations during inference, without changing model weights. It builds separate directions to identify the hidden states tied to the target code and to suppress its reproduction. Across fourteen baselines, two code models and two corpora, UNBIND reduces target code reproduction by 97.3% to 99.1% as measured by F-BLEU, with at most small losses on HumanEval+ and MBPP+.
In briefNereus: Adaptive Parallelism System for LLM Post-Training
SummaryNereus is a cost-aware runtime that adapts reinforcement learning post-training jobs for large language models as cluster conditions change. It selects memory-feasible execution plans and decides whether to transition using a cost model, while reusing distributed state via Elastic Model Units and a global transition graph. The paper reports a 27.7% reduction in average step latency on a real-data trace and 2.14–7.27× and 1.10–1.47× end-to-end PPO throughput gains over OpenRLHF and Verl.
In briefUniversal Dynamic Portfolios: Online Asset Allocation That Tracks a Changing Benchmark
SummaryThis paper extends Cover's Universal Portfolio to compete with an arbitrary changing sequence of comparator portfolios. It shows that under the log loss, the fast rates from exp-concavity generally disappear in the dynamic setting, and a √(T·P_T) dependence on path length is unavoidable. The authors introduce two structure-aware measures, the Jensen-Shannon distance and the JS^q-path length, and propose a parameter-free Universal Dynamic Portfolio method. Under an additional bounded-gradient assumption, OPS achieves a faster T^{1/3}P_T^{2/3} dynamic regret rate.
WhyExtends universal portfolios to dynamic regret with new structure-aware measures and sharper rates; strong theory, limited empirical validation.
In briefTilted Schrödinger Bridge Matching: post-training a learned bridge with a reward tilt to fit preferences or physical constraints
SummarySchrödinger bridges offer an entropy-regularized framework for unpaired domain translation. This paper introduces Tilted Schrödinger Bridge Matching (TSBM), a post-training method that fine-tunes a pretrained bridge toward a reward-tilted target distribution while keeping the source distribution fixed. It is built as an alternating optimization derived from Adjoint Matching. The method is evaluated on unpaired image-to-image translation for MNIST digit properties and CelebA facial attributes.
In briefBackdoor as Probe: A Test-Time Defense Against Attacks on CLIP
SummaryThis paper proposes Backdoor as Probe (BaP), a test-time adversarial defense for the CLIP vision-language model that requires no retraining. It implants a defender-controlled backdoor in a selected MLP layer through a closed-form edit, which responds weakly to clean inputs and strongly to adversarial shifts, so it can detect adversarial inputs. Detected inputs are then corrected with a small optimized perturbation. Across 16 benchmarks, average robust accuracy rises from 1.0% to 52.3% while clean accuracy is retained, with inference up to 5.7 times faster.
WhyRepurposes backdoor triggers as a closed-form edited probe detecting adversarial shifts; validated on 16 benchmarks, raising robust accuracy from 1.0% to 52.3%.
In briefUniform Race: a parameter-free method for picking samples close to a target distribution
SummaryThe paper studies approximate sampling, where one selects a sample from independent draws of a proposal distribution so that it approximates a target distribution known only up to a constant. The authors propose uniform race, a parameter-free algorithm that divides each importance weight by an independent uniform random variable and returns the candidate with the largest score. They report that it meets the best approximate rejection sampling error bound for every threshold simultaneously, and that test-time scaling experiments on LLM math-reasoning tasks show it remains competitive.
WhyProposes a parameter-free sampler matching the best approximate rejection sampling bound for all thresholds at once, with exponential gains shown on instances and LLM validation.
In briefMemory Prediction Excess: a probabilistic measure of how much past history helps prediction
SummaryThis paper introduces the Memory Prediction Excess (MPE), which measures the average gain in prediction accuracy from using a process's full observed history instead of only its static marginal distribution. It establishes that the MPE is non-negative and bounded, and defines a normalized version and a finite-history variant (FH-MPE). For finite-order Markov chains, the paper shows the minimal memory length needed to match full-history prediction is bounded by the Markov order.
In briefCross-variate interaction in correction space for test-time adaptation in time series forecasting
SummaryThe paper proposes CoRe, a test-time adaptation method for time-series forecasting that models cross-variate interactions on adapter corrections rather than backbone predictions, to avoid propagating backbone errors across variates. It combines Shared-anchor Correction Refinement with a parameter-efficient bottleneck and input-conditioned spectral gating. Across seven backbones, six datasets, and four prediction horizons, the authors report average MSE reductions of 25.82% over backbones and 10.57% over the state-of-the-art TSF-TTA method.
Yuanyuan Deng, Mykola Pechenizkiy, Songgaojun Deng
In briefMechReasoner: A Simulator and Benchmark for Mechanistic Reasoning in Qualitative Physics
SummaryThis paper presents MechReasoner, a qualitative simulator based on confluence-based qualitative physics, along with a benchmark for testing mechanistic inference. The benchmark's 1,120 items are generated deterministically across 18 mechanisms and six task families. GPT-5.5's accuracy falls from 76.1% on the lowest-complexity items to 38.0% on the highest, and this negative association persists after controlling for prompt and answer length.
In briefA persistent state that records how routing influences build up in Mixture-of-Experts models
SummaryMixture-of-Experts models send each token to a few experts, but ordinary routers keep no record of how influences from earlier layers accumulate. The paper proposes SA-MoE, which gives each router a small persistent state that experts cannot see, updated by learned writes whose changes form an exact routing ledger. Tests on sparsely upcycled SmolLM2 and Gemma models, with three seeds each, found the added forward compute under 1% and heavy use by trained routers.
WhyProposes an exact routing ledger over persistent router state, validated across two model families and three seeds with intervention experiments.
In briefDo motion tokenizers for co-speech gesture generation capture gesture meaning? A probing study of whether these codebooks encode communicative gesture properties.
SummaryThe paper examines whether discrete motion tokenizers used for co-speech gesture generation retain gesture-related semantic properties. The authors probe a reconstruction-trained codebook with 19 gesture descriptors, ranging from raw motion to abstract communicative function. Geometry and handedness are readily decodable from token embeddings, while motion category is only weakly decoded despite systematic differences in code usage. The authors conclude that reconstruction objectives alone do not guarantee capture of gesture semantics, and that such evaluations can guide the design of more semantic tokenizers.
In briefGenMem: Generative Symbolic Memory for Self-Evolving LLM Agent Harnesses
SummaryThe paper presents GenMem, a memory system for LLM agents that treats memory management as generating symbolic addresses. Each memory is indexed by a Symbolic Identifier, a multi-level token tuple, so the agent generates stable addresses while the stored content is rewritten over time. GenMem pairs a retriever and an evolver within a multi-agent harness, trained with GRPO using process and outcome rewards. Offline experiments on ALFWorld, WebShop, multi-hop QA, medical reasoning, and deep research compare it against memory-augmented baselines.
In briefLong Time No See: Benchmarking VLMs on Out-of-Sight Reasoning in Egocentric Video
SummaryThis paper introduces Beyond3D, a benchmark testing whether vision-language models can track objects that were moved and then left the camera's view. It contains 9,000 questions over 135 egocentric videos, covering visibility, timing, location, and 3D spatial relations. The best of nine evaluated models reached 42.2% accuracy, above the 29.7% chance level, with the largest errors in identifying when an object was last visible.
Fangzhou Ma, Ivo Alexander Ban, Eren Homburg et al.
In briefDisKO: Deep Koopman Learning in Distribution Space from Unpaired Snapshots
SummaryMany complex systems are observed only as distribution snapshots at unpaired times, which makes learning from trajectories hard. This paper treats the distribution itself as the dynamical state and learns in distribution space. The proposed DisKO extends deep Koopman learning by jointly learning distributional observables, a finite-dimensional Koopman representation, and a generative map back to the full distribution. The authors report state-of-the-art extrapolation across seven benchmarks and recovery of leading Koopman eigenvalues and eigenfunctions on systems with analytic spectra.
In briefCLAD: A constrained abstract domain for verifying neural networks with extra input constraints
SummaryThis paper addresses neural network verification, which formally checks that a network satisfies a property for all inputs in a defined region. Existing abstract domains describe input regions only as Lp-norm balls, which produces loose over-approximations when practical regions also include additional constraints. The authors propose CLAD, which handles combinations of convex constraints by using Lagrange multipliers and a projected primal-dual method. On 1,944 instances, CLAD matches GCPCROWN on standard Linf properties and verifies 22% more instances overall on constrained properties, including 60% more on L2-ball properties.
In briefTesting whether learned physical models reproduce reference dynamics via predictive structure
SummaryThe paper proposes a framework for judging whether machine-learning models of physical systems reproduce the underlying dynamics, not just accurate states. Domain experts define physical representations, and reference trajectories determine which relationships are predictive and kept as tests. The authors apply it to atmospheric forecasting with ERA5 planetary-wave activity and Northern Annular Mode evolution, evaluating Pangu-Weather, GraphCast, and FengWu. The models show distinct departures from reference predictive structure that conventional forecast errors do not capture.
Oskar Bohn Lassen, Joao Paulo de Souza Boger, Simon Driscoll et al.
In briefBMND: Direct Poisson denoising via N-dimensional block matching and collaborative filtering
SummaryThis paper presents BMND, a dimension-independent extension of the BM3D and BM4D block-matching and collaborative-filtering methods for Gaussian and Poisson data. It processes Poisson observations directly, without a variance-stabilizing transform, using noise-aware patch matching and signal-dependent variance propagation. An optional mass-conservation step preserves total observed intensity after aggregation. Experiments on 1D, 2D and 3D data show improved reconstruction and reduced intensity loss in low-count cases, and the non-learning method is released as an open-source library.
Christof Duhme, Lars Schiefelbein, Florian B\"uther et al.
In briefMust Native 3D Texture Generation Rely on 3D Assets for Training?
SummaryThis paper proposes Tex-Zero, a framework that trains native 3D texture generation models without 3D assets. The authors argue that fine-grained color information is essential for training, while geometry can be constructed manually. They convert 2D images into 3D training samples by placing them as planes with patch-wise random rotations and aggregation, then train a VAE and a DiT using only these images. The authors report that the resulting models generate high-fidelity 3D textures with fine details.
In briefLatent Multimodal Memory Reasoning (LMMR): a neuro-symbolic framework that organizes long-term multimodal memory in a continuous latent space
SummaryThe paper proposes LMMR, a framework that moves long-term multimodal memory organization into a continuous latent space. Text hidden states and visual patch features are mapped into a shared Sparse Autoencoder concept space, where they form a turn-level multimodal memory graph. A graph encoder aggregates cross-modal and temporal dependencies and feeds the readout back into the generation backbone as a continuous residual, while native visual tokens are retained. The authors report significant gains over state-of-the-art baselines on nine Mem-Gallery tasks and twelve MemEye settings.
In briefUsing LLMs to generate executable multi-agent system specifications from natural language
SummarySpecifying a multi-agent system requires formal language expertise and labelled data that is rarely available, while the specification must also be executable to allow run-time monitoring. This paper proposes genRTEC, which uses pre-trained large language models to generate executable specifications in the Run-Time Event Calculus from short natural language descriptions of the concepts involved. The authors report an extensive empirical evaluation across various MAS specifications, including qualitative and quantitative assessments. They conclude that genRTEC produces executable specifications with high predictive accuracy without compromising reasoning efficiency.
Andreas Kouvaras, Periklis Mantenoglou, Alexander Artikis
In briefLearning snow water and snow height changes from Sentinel-1 radar data
SummaryThis study compares three machine learning models (XGBoost, U-Net and SegFormer) that estimate changes in snow water equivalent and snow height over the Italian Alps using Sentinel-1 InSAR data. The IT-SNOW reanalysis serves as the reference. SegFormer performs best on both targets, with a mean absolute error of 10.391 cm for snow height and 27.113 mm w.e. for snow water equivalent. Most of the error comes from a systematic offset in the estimated mean change rather than the spatial pattern.
Luca Barco, Lorenzo Innocenti, Bianca Bartoli et al.
In briefProbabilistic Geodesic Flow Matching: a generative method using geodesics over location-scale distributions
SummaryThis paper generalizes flow matching, a generative framework that transports samples via an ODE, to the location-scale family to handle inhomogeneous data with heavy tails or sharp contrasts. It defines probability paths as geodesics on the manifold of probability distributions, which the authors call probabilistic geodesic flow matching. The authors argue that Euclidean optimal transport paths are not necessarily optimal in probability space. Experiments on synthetic benchmarks and scientific datasets at different scales show improved or comparable performance relative to state-of-the-art geometry-motivated generative models.
In briefBeyond Site Agreement: Re-estimating Brain Networks to Generalize Across Sites
SummaryThis paper studies how brain scan–based models generalize across imaging sites in resting-state fMRI. It argues that agreement between sites on full-scan connectivity graphs does not guarantee that predictions hold when connectivity is re-estimated within the same scan. The authors propose BRIO, which uses within-scan re-estimation to guide cross-site alignment, and report up to 3.8% relative accuracy gains over baselines in leave-one-site-out tests on four datasets.
In briefFaster World Action Model inference by adapting reused intermediate states
SummaryWorld Action Models jointly model actions and environment dynamics for future-aware control, but iterative diffusion or flow inference causes significant latency. The paper introduces WAMACHINE, a training-free framework that accelerates inference by preserving and adapting retained inference state as the control loop evolves. It combines Trajectory Remapping across replans, Observation Rebinding across denoising steps, and Residual Rescaling across Transformer layers. On LIBERO and RoboTwin 2.0 with three representative architectures, it reports 1.47–3.05× lower observation-to-action latency and 2.23–3.27× lower GPU inference time per replan, retaining 96.69–99.54% of native task success.
In briefWorldAttention: An Efficient Attention Design for Interactive Video World Models
SummaryThis paper proposes WorldAttention, an attention architecture for text-conditioned interactive video world models that aims to keep long-range history while limiting compute and GPU memory. It combines Hybrid Sparse Attention, which pairs linear global attention with head-adaptive sparse attention, and a Hierarchical KV Cache that stores history in semantically indexed pages across memory tiers. Custom kernels support both designs. The authors report subject consistency scores of 0.9472 on VBench-Long and 0.9668 on InterVBench, which they state exceed prior state-of-the-art methods.
In briefPMOPD: Keeping Task Skills From Interfering in Multi-Teacher On-Policy Distillation
SummaryThis paper studies multi-teacher on-policy distillation (MOPD), a method for combining several specialized skills into one language model, where improving one skill can suppress another. The authors observe that parameter updates for different tasks concentrate in separate low-dimensional subspaces and propose PMOPD, which projects gradients and optimizer updates to remove components that interfere with protected task directions. They also add a conflict probe for task ordering and a cycling strategy for revisiting tasks. On Code, Reason, and Math tasks, PMOPD improves over MOPD, raising the three-task average by 2.54 points on Qwen2.5-7B and 2.09 points on Llama-3.1-8B.
In briefShaping Persistent Representations from Independent Interactions: a training method that uses relations between interactions to help world models learn reusable persistent context
SummaryThe paper introduces SPRII, a training principle that treats shared properties across related trajectories as weak supervision, without numerical property labels. It uses two components, Align, which makes contexts from related interactions agree, and Cross, which predicts one interaction's future from another's context. The authors separate three stages, formation, use, and value, and report average gains above 10% in downstream task performance and above 15% in persistent-property readout across thirteen evaluation settings.
In briefAfter the Fix: Whether Repaired Agent Experience Helps Later Tasks
SummaryThe paper asks whether repairing a failed agent episode makes its experience a better memory for the next task. Using roughly 3,300 runs on ThinkingBox and APEX task pairs under eleven conditions, it compares transferred corrected experience against independent execution. The correction gains vary across settings, and many apparent gains stem from worse uncorrected performance rather than better corrected memory. The authors conclude that repairing experience and reusing it are distinct values.
In briefWhen local gains fail to transfer: frozen Earth-observation embeddings in wildfire mapping
SummaryThe paper tests frozen Earth-observation embeddings (TESSERA, AlphaEarth) against ESA WorldCover and annual Sentinel-2 index summaries for identifying burned land across six wildfires in Greece and Spain. Inside a fire the embeddings lead by 0.05 to 0.13 ROC AUC, but on a fire in another region that advantage disappears. Forecasting a fire twelve kilometres away from one that burned the previous year still works, with TESSERA reaching 0.772 AUC, so the authors argue evaluations should report held-out regions.
In briefLow-Rank Structure of RL Updates in Vision-Language-Action Models
SummaryThis paper studies how reinforcement learning changes vision-language-action robot policies during post-training. Across several flow-based VLA models and four benchmarks, the authors find RL updates are low-rank and concentrated in small Timestep Modules within the action expert. They report that shift update directions in these modules predict task success and that steering along them further improves policies without extra RL training.
In briefSummarize Before Grounding: Query-Guided Chunk Condensation for Long-Video Temporal Grounding
SummaryThis paper addresses locating the video interval that matches a language query in long videos. It proposes SumGround, which splits videos into chunks, compresses redundant visual tokens into query-guided latent summaries, and retrieves the chunk summaries most likely to contain the target event. The authors report that the method outperforms previous state-of-the-art approaches across multiple downstream datasets, with notable gains on long videos.
In briefOne Distribution, Two Channels: Coverage-Calibrated KV Cache Retention and Compensation
SummaryThis paper addresses memory and bandwidth limits of key-value (KV) caches during long-context decoding. It proposes CORE, which distills an offline allocation combining query utility and log-determinant coverage into a lightweight indexer, so one calibrated distribution selects which KV states to keep and redistributes evicted attention mass into latent-memory writes. Reported results show gains of up to 3.78 points over the strongest RULER baseline at 90% compression, with LongBench and repeated-eviction evaluations also reported.
In briefTracking Burn Scars Over Time Using Sentinel-3 Satellite Images
SummaryThis paper studies how to map wildfire burn scars quickly and accurately from Sentinel-3 OLCI satellite images, which is needed for post-fire damage assessment. The authors built a dataset of 246 wildfire activations from 2016 to 2025, using Copernicus Emergency Management Service data with temporally paired Sentinel-3 acquisitions. They compared spatial and temporal (ConvLSTM-augmented) versions of three backbone networks. Temporal modelling improved segmentation only when pre-fire frames were included, and a 5-band subset matched the full 21-band configuration.
Luca Barco, Edoardo Arnaudo, Andrea Bragagnolo et al.
In briefDeep Weighted Bellman Residual Minimization for Q* Estimation
SummaryThis paper tackles offline reinforcement learning problems such as distribution shift, Q-value overestimation, and low sample efficiency. It proposes a weighted Bellman residual minimization framework that uses density ratio weighting to combine expert demonstrations with behavioral data, without relying on the common completeness assumption. The authors derive convergence rates for density ratio estimation and for the excess risk of the resulting deep Q* estimator. Experiments report improved numerical performance and policy generalization over existing methods.
In briefTULIP: Per-input layer-targeted unlearning for large language models
SummaryThis paper examines representation-level unlearning in LLMs, where existing methods apply a single fixed layer to the entire forget set, and asks whether that is sufficient. The authors propose TULIP, which for each input uses the logit lens to locate the layer where the forget answer is formed and read out, then removes the hidden state's alignment with that answer's unembedding vector at that layer. The paper reports that TULIP outperforms output- and representation-level baselines on TOFU, PISTOL, and WMDP across Llama, Qwen, and Zephyr models, and remains robust to paraphrase and quantization attacks.
WhyProposes per-input layer selection for unlearning, backed by a hijacking experiment and comparisons against baselines across several benchmarks and models.
In briefThe Model Knows When to Stop: A Training-Free Early Stopping Method for Long-Context Reading
SummaryLong inputs are often read in sequential chunks, and continuing after enough evidence is found wastes computation. The paper introduces Answer-Convergence Stopping (ACS), a training-free rule that probes the frozen model's current answer after each chunk and stops when that answer is confident and stable, using only output-side generation and token log probabilities. On LongBench-v2, ACS is the only stopping policy that matches or exceeds full-reading accuracy. Across 250 S-NIAH questions, ACS's premature stopping rate ranges from 0% to 12%, compared with 8.4% to 45.6% for the verbalized gate.
Muath Alyobi, Mohamed Eltahir, Almoayyad Abuljdail et al.
In briefReinforcement Learning from Intermediate Renders for Image-to-Code Generation
SummaryThis paper addresses image-to-code generation, such as producing SVG code from a reference image, where reinforcement learning rewards based only on the final rendered output provide sparse feedback. The authors observe that many intermediate code prefixes already render meaningful partial results, and they propose IR4RL, which uses a token-level render-progress reward to turn changes between intermediate renders into localized feedback. The method is evaluated on Image-to-SVG and Image-to-TikZ tasks, where it outperforms supervised fine-tuning and standard GRPO.
In briefAgentWare: A framework that automates the full lifecycle of AI agent apps across edge and cloud
SummaryThis paper presents AgentWare, an AgenticOps framework that automates provisioning, deployment, observability, and evaluation of LLM-based agent applications spread across edge and cloud infrastructure. It supports automated semantic evaluation via LLM-as-a-Judge and generates reproducible reports on correctness, performance, resource use, and energy consumption. The authors demonstrate the framework with a distributed book assistant agent deployed on real edge-to-cloud infrastructure under multiple configurations. The results indicate it reduces manual effort for deployment, instrumentation, and analysis.
Michalis Kasioulis, Moysis Symeonides, George Pallis et al.
In briefScaling compute with small recursive models for combinatorial optimization
SummaryThe paper proposes Tiny Recursive Models, a neural method for combinatorial optimization that scales recursion depth and parallel sampling width. It reports a better balance of solution quality and inference speed than heatmap-based neural solvers on the Traveling Salesman Problem and Maximum Independent Set. The authors also explore self-relabeling, replacing training labels with the model's own better solutions, which they say can reach comparable quality.
In briefIn-game Toxic Detection with Attention-Residual Bidirectional Models
SummaryThis paper studies detecting toxic language in in-game player chat, a hard problem because messages are very short and full of game slang and jargon. It introduces a shared task built on real in-game chat data and proposes BRAR (Bi-directional Representations with Attention Residuals), which performed best on toxic language slot filling. The authors report that BRAR captures global context effectively and outperforms existing baselines on this task.
In briefSpeechCritic: A Speech Judge That Explains Differences from Few Human Labels
SummaryThe paper introduces SpeechCritic, a diagnostic speech judge that decides which of two speech clips is better, along which perceptual dimensions they differ, and which audible cues support the decision. It learns from about 300 human-labeled comparisons by calibrating a frontier audio-language model with acoustic measurements that agree with human judgments. A 7B judge is then trained with SFT, OPD, and RL, and the pipeline is shown to work for English-Japanese and English-Spanish.
In briefCounterfactual Attention Policy Distillation for Video Time Localization
SummaryThis paper addresses the difficulty multimodal large language models have in pinpointing when events occur in long videos. It proposes Counterfactual Attention Policy Distillation (CAPD), which masks each temporal segment and measures how the teacher model's output changes to weight distillation toward evidence that affects boundary prediction. Trained on Qwen3-VL-8B-Instruct with 2,500 samples for one epoch, CAPD improves average recall on TimeLens by 12.0% relative to GRPO. It keeps general video understanding performance, with accuracy comparable to the base model.
In briefCalibrated uncertainty for water pollution monitoring path planning, replacing Gaussian Processes with Deep Ensembles
SummaryThis paper examines whether a well-calibrated Deep Ensemble can replace a Gaussian Process to provide better uncertainty for planning sensing paths in aquatic oil spill monitoring. Five planning strategies are compared, all using a Deep Ensemble trained on physics-based oil spill simulations. The ensemble reduces normalised reconstruction error by 83% relative to the Gaussian Process baseline. With the ensemble, multi-step lookahead planners outperform greedy selection, and Monte Carlo Tree Search is recommended for matching Orienteering's reconstruction quality at much lower computational cost.</summary>
Samuel Yanes Luis, Alejandro Casado P\'erez, Alejandro Mendoza Barrionuevo et al.
In briefDiffusion Subgoal Planning (DSP) for Long-Horizon Offline Goal-Conditioned RL: a method using diffusion models to generate intermediate subgoals that guide agents toward distant goals.
SummaryOffline goal-conditioned reinforcement learning struggles in long-horizon tasks, where sparse rewards and discounting make value-based guidance unstable. The paper proposes Diffusion Subgoal Planning (DSP), which uses a diffusion model to generate high-level subgoals by learning conditional and unconditional flows and applying classifier-free guidance at inference. The authors report that DSP outperforms prior methods on navigation and manipulation benchmarks, with especially strong results in maze environments requiring multi-step subgoal planning.
Hengrui Zhang, Yuhu Cheng, C. L. Philip Chen et al.
In briefNudgeability: Reasoning Models Follow Confidence Signals Without Tracking Their Competence
SummaryThis paper examines whether reasoning language models that can call tools decide to answer on their own or delegate, and whether stated confidence or doubt changes that choice. Researchers inserted a first-person sentence of confidence or doubt into identical reasoning traces and compared the resulting continuations. Models were consistently sensitive to these signals, with doubt increasing delegation and confidence decreasing it. However, responses were poorly targeted: a median 42% of induced flips were well-targeted, only 2 percentage points above a random baseline.
In briefPersonaManifold: Finding Curved Geometry in LLM Persona Representations
SummaryThis paper shows that common methods for controlling LLM personas by adding and interpolating vectors in activation space fail in systematic ways, suggesting the representations are not linear. The authors propose PersonaManifold, which models personas as points on a curved low-dimensional manifold and steers between them along geodesic paths instead of straight lines. They also introduce the Behavioral Similarity Triplet benchmark, which measures persona similarity through situational behavioral responses rather than self-report questionnaires. Experiments on three open-source LLMs report that geodesic distance predicts behavioral similarity better than Euclidean distance, and geodesic steering helps most in high-deviation regions.
In briefA review of agentic agri-robotic phenotyping, moving from fragmented sensing to a unified PhenoAgent approach
SummaryThis review traces plant phenotyping from manual trait measurement to high-throughput, robotic, and AI-driven crop monitoring, noting that current systems remain fragmented across sensing methods, traits, and growth stages. The authors frame phenotyping as an integrated seed-soil-plant-environment-management problem and propose a conceptual PhenoAgent framework for evidence-based crop-state interpretation and uncertainty-aware decision support. The paper also discusses challenges including dataset scarcity, annotation, benchmarking, generalization, and explainability.
Muhammad Owais, Ehtesham Iqbal, Samee Ullah Khan et al.
In briefRetracing Hodgkin-Huxley: Recovering States Does Not Prove the Mechanism
SummaryThis paper asks whether machine learning can recover the hidden-state reasoning behind the Hodgkin-Huxley neuron model. The authors train structured latent models on simulated data, withholding gate identities and trajectories, and test prediction, state recovery, protocol transfer, and agreement with HH dynamics. They conclude that accurate prediction does not establish recovery of the underlying mechanism, so state and dynamics should be evaluated separately.
In briefFlowState: Execution State as Memory for Long-Horizon LLM Agents
SummaryThis paper proposes FlowState, which treats an agent's execution state as memory that can be retained and revisited across requests. It stores typed state nodes, their relations, and references to raw tool observations, and uses incremental state updates and progressive state access to retrieve history as needed. Compared with a full-context baseline using the same DeepSeek-V4-Flash model, FlowState improves average success on MemoryArena by 4.55 points and on τ³-Bench by 13.95 points, while cutting total token use by 43.2% and 40.6%.
In briefGLF-Q: A Global-Local Feature Post-Training Quantization Method for Vision Transformers
SummaryThis paper proposes GLF-Q, a post-training quantization method for Vision Transformers that requires no retraining. It guides block reconstruction by aligning global and local features and uses offline Hadamard transformations with zero runtime overhead to spread activation outliers. According to the paper, GLF-Q with standard uniform quantizers substantially outperforms state-of-the-art methods at 3-bit image classification and achieves speedups under 8-bit GPU deployment.
In briefRethinking Latent Visual Reasoning: Grounding Reasoning in Visual Evidence
SummaryThis paper studies latent visual reasoning, where multimodal language models perform intermediate computation in continuous latent tokens instead of words. The authors find that latent tokens respond weakly to image changes that alter the correct answer, which they attribute to missing supervision during GRPO training. They propose ReaLVR, which adds visual-evidence supervision by contrasting correct and wrong answers and relevant and mismatched visual evidence. The method outperforms baselines across three model families and scales to models up to 235B parameters.
In briefSingle-layer MeMo as a randomized Hamming-kernel classifier
SummaryThis paper studies the single-layer form of MeMo, a language model that stores context-to-next-token associations in a correlation matrix memory. The authors show its ideal retrieval rule is a multiclass classifier based on the positional Hamming kernel, with inputs and outputs encoded by Gaussian random codes. They bound the errors from context sketching and output decoding and give a margin-based guarantee for recovering the ideal prediction. On a restricted WikiText-2 next-token task, it is compared with classical baselines and is reported to offer a useful trade-off among accuracy, memory, and throughput.
In briefBrain-signal-conditioned action policies for neural motor decoding
SummaryThis paper proposes BrainVLA, a framework that uses a pretrained vision-language-action (VLA) model, aligned through language, to decode neural motor signals while reducing reliance on scarce paired neural-action data. The authors adapt the OpenVLA-OFT policy with LoRA fine-tuning and train a neural encoder via neural-language alignment to convey motor intention to the policy. Evaluated on two datasets with different action dimensionalities, it outperforms the evaluated baselines in cross-session decoding R² and task success rate, with high training data efficiency.
In briefACPruner: Pruning Visual Tokens in Vision-Language Models by Maximizing Attention Coverage
SummaryLarge vision-language models are slowed down by the many visual tokens they process. This paper treats visual token pruning as a coverage problem, selecting a compact token subset whose attention in the vision encoder covers the image while favoring informative regions. The authors propose ACPruner, a training-free framework that scores tokens using intra-modal saliency and inter-modal relevance, then greedily selects tokens to maximize coverage. Experiments on several LVLM backbones report strong performance retention along with substantial end-to-end inference speedups.
In brief"RoPE is Dead, Long Live RoPE": Scalable Data-aware Positional Encodings
SummaryThis paper argues that Rotary Position Embedding (RoPE), the default in many language models, is biased toward nearby tokens, and its slow frequency bands expose models to unseen angles when extrapolating beyond the training context. The authors propose Data aware RoPE (DaRoPE), which keeps standard RoPE on fast bands but uses bounded coordinates learned from contextual representations on slow bands. Compared under matched conditions across synthetic tasks, symbolic music, genomics, neural signals, and language models from 124M to 50B parameters, DaRoPE leads on non-text benchmarks and mitigates recency bias.
WhyReplaces RoPE's slow bands with learned contextual coordinates, improving extrapolation; broad multi-domain comparisons up to 50B parameters support the claims.
In briefSkillRubric: Co-Evolving Guidance and Evaluation Rubrics for Multimodal Agents
SummaryThis paper introduces SkillRubric, which pairs each reusable skill with actor-facing guidance and an evaluator-facing rubric. A multimodal verifier uses screenshots and tool outputs to give completion and progress rewards to the relevant turns. The authors also propose an alternating co-evolution scheme that validates guidance and rubric revisions separately, and report consistent gains across multimodal agent benchmarks.
In briefTraining-free spatial guidance that limits image edits to the instructed region
SummaryThe paper introduces Attention-Scoped Guidance (ASG), a sampler wrapper for instruction-guided image editing that requires no training, no external mask, and no extra network evaluation. It derives a soft support map from the editor's instruction attention, weakening text guidance where support is low and strengthening image anchoring where it is high. On the MagicBrush and PIE-Bench++ splits, ASG improves preservation-oriented metrics, and a dose-matched control shows that the spatial placement itself drives the gain.
In briefPulseInfer: I/O-Centric Sparse KV Cache Offloading for Faster Long-Context LLM Decoding
SummaryThis paper addresses the decode bottleneck in long-context LLM serving, where large KV caches limit batch size and GPU utilization. It presents PulseInfer, a system that offloads sparse KV cache blocks to CPU memory and manages the resulting CPU-GPU recall I/O through interruptible layer-wise scheduling, adaptive offloading admission, and coalesced transfers. Implemented on SGLang, the authors report up to 4.7x higher decode throughput and up to 76% lower TPOT while preserving near-lossless accuracy.
In briefReinforcement learning to improve branching in neural network verification
SummaryThis paper presents RSB, a reinforcement learning framework that refines branching heuristics used in formal verification of neural networks. It trains an actor-critic architecture to maximize long-term cumulative rewards rather than immediate scores, using attention weights derived from neuron features and graph embeddings to rescale baseline heuristic scores. On 600 challenging instances, the authors report that RSB outperforms state-of-the-art branching heuristics, solving 11% more instances while reducing branch exploration by 50%.
In briefParameter symmetries and conservation laws in gradient flow, examined through a unified geometric framework.
SummaryThis paper studies how parameter-space symmetries relate to conservation laws in neural network training under gradient flow, drawing on Noether's theorem from physics. The authors develop a unified geometric framework that specifies when symmetries correspond to conservation laws. They introduce compositional identifiability and use it to establish an inheritance principle for complete characterizations in multilayer networks. The framework is applied to multi-head and grouped-query attention, polynomial neural networks, and square deep linear networks.
In briefSGG-ReflAct: Sub-goal guided reflection with structured planning for reliable long-horizon reasoning
SummaryThe paper addresses LLM agents drifting from their goals in long multi-step tasks due to inconsistent intermediate beliefs. It proposes SGG-ReflAct, which feeds sub-goals from a single-path LLM planner into the reflection process, and BeamSGG-ReflAct, which uses beam search for plan exploration. Experiments on ALFWorld, ScienceWorld, and Jericho show it outperforms REFLACT in nearly all settings and reduces hallucinated actions.
In briefActionLens: A Benchmark for Diagnosing Who-Does-What-When Errors in Video AI Models
SummaryThe paper introduces ActionLens, a diagnostic benchmark of 6,701 multiple-choice video questions testing whether vision-language models can link the right action to the right person at the right moment. Answers are derived from 1.58 million per-second, per-person annotations across five diagnostics, including gaze detection. Across 20 models, the best scores 65.9% on the human-reviewed subset versus 91.0% for humans, with gaze detection near chance. The authors also find models often select the wrong actor's action, and release data and code publicly.
WhyProposes a fine-grained diagnostic benchmark for spatial-temporal binding, with deterministic labels from per-second annotations and controlled tests across 20 VLMs.
Gueter Josmy Faure, Min-Hung Chen, Hao Ping Wang et al.
In briefMemDream: Agents That Test and Repair Their Own Memory Before Failures Happen
SummaryThe paper proposes MemDream, a framework that lets LLM agents proactively probe and repair their memory graphs instead of fixing them only after retrieval failures appear. It runs periodic offline "dream" cycles in which three specialized agents (Dreamer, Analyst, Consolidator) collaborate, and a policy trained with Group Relative Policy Optimization learns which repairs yield lasting gains. Experiments on LoCoMo and MemoryAgentBench report a 4.5-point F1 improvement on LoCoMo and a 9.1-point higher overall score on MAB over the strongest reactive baselines.
