In briefReSight-SMC: Two-stage power sampling with visual scouts for vision-language model reasoning
SummaryThis paper introduces ReSight-SMC, a verifier-free, training-free two-stage sampling method for large vision-language models. The first stage uses isolated SMC islands to keep independent reasoning trajectories and directs visual scouts to prefix-relevant image regions. The second stage aggregates importance mass by canonical answer and samples an answer with a supporting trajectory. Across four backbones and five benchmarks, it outperforms Power-SMC in aggregate on reasoning and perception tasks and remains competitive with reinforcement-learning-trained models.
In briefReference-Tail Trust: Certified Probability Floors for Learned Updates in Deployed GNNs
SummaryThe paper addresses how graph neural networks can use improved message passing after deployment without undermining predictions already trusted. It proposes Reference-Tail Trust (RTT), which allows learned updates inside a frozen GNN and certifies the prediction actually served through per-node probability floors and a call-level budget. In the ogbn-arxiv audit, RTT reports a mean gain of 6.5×10⁻³ nats per call with upper bounds on regression and negative-flip rates, and further evaluations cover molecular, image, and tabular settings.
WhyProposes certified probability floors for learned updates inside frozen GNNs, with derivations and broad experiments across datasets and shifts.
In briefDual-Stream Simultaneous Translation Using 2D Grid Attention
SummaryThis paper proposes a dual-stream attention framework that represents source and target streams as a two-dimensional grid and models their interaction with four attention types under joint normalization, for simultaneous machine translation. Two approximations reduce per-layer complexity, while a self-guided training loop generates read/write supervision without external alignment. On Chinese-to-English translation, the model outperforms the Wait-k baseline at comparable latency, gaining +5.66 BLEURT and +10.36 COMET.
In briefDiagnostic Harness for Bounded Autonomy in Regulated AI Agents
SummaryThis paper introduces RegLLM, a diagnostic harness that measures six trustworthiness signals for AI agents in regulated workflows, including citation validity, escalation correctness, and unsafe-action rate. In a small offline run, enabling governance raised escalation recall from 0 to 0.67 and lowered unsafe-action rate from 0.33 to 0.08. The authors find that configuration variance across small pilot runs can overwhelm apparent tuning effects, so the results are diagnostic only and do not establish reliable adapter effects or production readiness.
In briefFILIGREE3D: Scaling sparse latent flow matching for ultra-high-resolution single-image 3D generation
SummaryThe paper presents FILIGREE3D, which generates 3D geometry from a single image at voxel resolutions up to 2048³, with extensibility to 4096³. It combines structure-aware sparse scaling with alternating local-global attention, selects training samples by high-resolution geometric gain, injects multi-scale image features into a sparse 3D DiT, and applies visibility-aware voxel regularization. The authors report that peak GPU memory stays within practical limits and generation takes about one minute, with better geometric fidelity and detail preservation than existing baselines.
In briefColNanoVDR: Training a small, text-only query encoder for multi-vector document retrieval without encoding document pages
SummaryThe paper presents ColNanoVDR, a framework that distills the query encoder of multi-vector visual document retrieval models into a small student without encoding or caching training pages. It aligns student and teacher query tokens using entropic optimal transport with learned token weights. The 149M-parameter text-only students keep about 95% of their teachers' NDCG@5 on ViDoRe v1-v3 and encode queries up to 26x faster.
WhyBrings document-free distillation to multi-vector visual retrieval with an optimal transport alignment, proves a MaxSim score bound, and retains about 95% teacher NDCG@5 on ViDoRe.
In briefRole-guided mixture-of-experts encoder for pathology slide image classification
SummaryThis paper addresses whole slide image classification, where frozen pathology foundation models may not adapt well to target tissue patterns. It proposes a role-guided mixture-of-experts feed-forward network that is trained in two stages: first distilling pathology priors from a frozen Virchow2 teacher into a lightweight DINOv2-small student, then refining the experts with asymmetric prototype-guided optimization. The resulting encoder produces patch representations compatible with standard MIL aggregators. Experiments on the public BRACS dataset and a private PAROTID dataset across five backbones show consistent gains over the strongest baseline.
In briefDeShortcut-Align: Reducing Spurious Shortcuts for Robust Safety in Reasoning Models
SummaryThis paper examines why safety-aligned large reasoning models often over-refuse and lose general abilities, attributing it to learned spurious shortcuts such as formatting templates and sensitive keywords. It proposes DeShortcut-Align, a three-stage framework covering refusal sensitivity attribution, attribution-guided contrastive augmentation, and counterfactual consistency regularization. Experiments on 7B and 14B models report improved robustness to template-stripping attacks, substantially reduced over-refusal, and better preserved reasoning capability.
In briefLVMT: A Transformer for Tracking Objects in Long Videos
SummaryThis paper addresses online video segmentation, which struggles to track objects through long videos with long-term occlusions. It proposes a lightweight GRU-based temporal propagation module that learns which object information to keep and pass across time, and a Truncated Query Propagation training strategy that propagates information between frame chunks while backpropagating only within chunks. The resulting Long-term Video Mask Transformer (LVMT) reportedly sets a new state of the art on six benchmarks and is 10 times faster than the prior best method. Code is publicly available.]
WhyAdds a GRU-based temporal propagation module and chunked training, achieving state of the art on six benchmarks with substantial gains.
Narges Norouzi, Niccol\`{o} Cavagnero, Idil Esen Zulfikar et al.
In briefECHO: Event-augmented wrist-only robot manipulation with memory and prediction
SummaryThis paper presents ECHO, a learning-based robot manipulation method that uses only a wrist-mounted camera, augmented with event camera data. Event cameras capture pixel-level brightness changes, helping when RGB images degrade under extreme exposure. ECHO includes a hindsight module that keeps past gripper trajectories from the event stream, and an outlook module that anticipates future scene changes. The authors report that ECHO outperforms RGB and RGB+event baselines on wrist-only RLBench tasks and in real-world wrist-mounted experiments.
In briefConformal prediction for tabular foundation models with better conditional coverage
SummaryThe paper notes that tabular foundation models can give prediction regions that under- or over-cover even when point predictions are accurate. It proposes C-USIM, a lightweight form of highest-predictive-density split conformal prediction that handles multimodal predictions without extra training or inference. Experiments with TabPFN and TabICL show better marginal coverage accuracy and lower average conditional and group coverage errors.
In briefFewer Assumptions by Design: A Reusable Skill for LLM-Assisted Verus Verification
SummaryThis paper studies whether LLM agents can write strong Verus specifications for doubly linked lists, a self-referential Rust data structure that is hard to verify, while relying on as few trusted assumptions as possible. It compares manual verification, property-specific verification, and a reusable skill that encodes domain knowledge and a task-decomposition strategy. The authors report that an LLM agent equipped with a carefully designed skill can produce strong, low-trust specifications for doubly linked lists in Verus.
Andrada-Livia Antoneac (Alexandru Ioan Cuza University of Ia\c{s}i, Bitdefender), Dorel Lucanu (Alexandru Ioan Cuza University of Ia\c{s}i) et al.
In briefSubRot: Signed gradient subspace calibration for rotation-based quantization of vision-language models
SummaryThis paper studies post-training quantization for vision-language models, aiming to keep multimodal performance at low bit widths. It proposes SubRot, which uses eigendecomposition of the empirical Fisher matrix of activation gradients to find a stable, sensitivity-separated channel subspace, and guides quantization error with signed first-order and second-order constraints alongside MSE. Across five VLMs and five benchmarks, it reports higher average scores than FlatQuant under W4A6 and W4A4, with a 1.4-point gain on LLaVA-NeXT-7B.
In briefOne Readout, Many Repairs: diffusion-guided search that fixes failed tool-agent calls more cheaply
SummaryTool agents use language models to act through external tools, but successful calls can still leave user requests unfulfilled. This paper presents ReCommit, a training-free framework that uses a single parallel readout from a masked diffusion language model to reuse operation-type scores across repair trials, guiding hierarchical search over operation supports. On real failures across four enterprise services in the Agent-Diff benchmark, it reports relative recovery gains of 75.9% and 63.2% at budgets B=3 and B=13, with mean full-budget repair time reduced by 61.3% and 51.3% versus the strongest evaluated 8B method.
In briefSPOC-Net: A Network That Identifies GNSS Jamming Components in Combined Signals
SummaryThis paper addresses the difficulty of extending GNSS jamming recognizers that treat each mixture as a separate class. It proposes SPOC-Net, which identifies the set of basic jamming components instead, using time-frequency features, learned component queries, and a decoder that selects a valid set. Training uses only measured single-component records, combined on demand into labeled mixtures. On 14,220 test mixtures it reports 80.69% exact-set accuracy and a 92.84% micro-averaged F1 score, and 80.89% exact-set accuracy on compositions excluded from development.
Zhihan Zeng, Kaihe Wang, Jos\'e A. L\'opez-Salcedo et al.
In briefP4Q: Jointly Designing Visual Token Pruning and Quantization to Speed Up Vision-Language Models
SummaryVision-language models are costly to run in memory and computation, which makes deployment hard. This paper presents P4Q, a framework that jointly optimizes visual token pruning and low-bit quantization, selecting tokens using fake-quantized features and calibrating the quantized model on the retained-token distribution. On LLaVA-NeXT, P4Q reports an average end-to-end inference speedup of 2.8x across eight test sets while keeping higher accuracy than prior compression and quantization methods.
In briefFrom Attention Sensitivity to Layer Role: Rethinking Mixed-Precision Quantization of Transformers
SummaryThe paper proposes JAB, which defines a loss over the joint Q, K, and V weights of an attention block, measured on the real attention output, and uses it both to fit quantized weights and to allocate bit widths. JAB works for attention-only quantization of Mistral-7B but stops working once MLP layers are included. A role-aware offset rule, which needs no sensitivity estimate, outperforms JAB on GPT-2's MLP and the full Mistral-7B model.
Nafiseh HosseinpourFardi, Negar Alihadi, Mahmoudreza Babaei et al.
In briefLimits of Metacognitive Monitoring in LLMs
SummaryThis paper examines the confidence reports of four frontier language models across 15 benchmarks to see whether models can tell when their answers may be wrong. It finds that high task accuracy can coexist with weak error discrimination; one model solves 97% of competition math problems, yet its answer-time confidence separates correct answers from errors barely better than chance. The authors conclude that hard questions and errors shared by models remain difficult to catch, even with prompted self-review or peer evaluation.
In briefRevisiting to Segment: Distilling Self-Generated Memory for Reasoning Segmentation
SummaryThis paper studies multimodal large language models that segment images by reasoning about visual content. It finds that reusing the model's own earlier reasoning and localization outputs as working memory improves segmentation. The proposed SWiM framework selects high-quality attempts to build this memory, then distills guidance from a memory-conditioned teacher into a student that sees only the image and query. The authors report state-of-the-art results on reasoning segmentation benchmarks.
In briefWhen Text Matters: Design Principles for Visual Token Pruning in Vision-Language Models
SummaryThis paper examines how to prune visual tokens in vision-language models to cut computation without losing key visual information. The authors propose a training-free method that first prunes using vision-encoder attention, keeps extra candidates until the decoder midpoint, and then uses text-to-visual attention to select the final tokens. Across eight benchmarks and three models, it outperforms the best baselines by average performance recovery gains of 11.10 and 16.84 percentage points at 80% and 90% pruning.
WhyA training-free staged pruning method defers text-guided reselection to decoder midpoint, yielding large performance recovery gains across eight benchmarks and three models.
In briefRobust multi-agent communication under variable bandwidth using attention-based hierarchical variational bottleneck
SummaryThis paper studies learned communication among multiple agents when bandwidth is limited, so that only the first part of each message is received. The authors propose AH-VIB, which combines a variational information bottleneck, attention-based autoregressive message generation, and a hierarchical robustness loss. It is evaluated on a custom cooperative object-inspection and occupancy-mapping task against baselines including MADDPG and CommNet. The results report competitive mean return and improved reliability under the most constrained bandwidth conditions.
Lukas Koch Vindbjerg, Qi Zhang, Yury Brodskiy et al.
In briefBeyond Verbalized Confidence: Training Reasoning Models to Report Calibrated Confidence
SummaryReasoning models trained with reinforcement learning from verifiable rewards tend to be overconfident. Existing methods have the model state a numerical confidence as text, which adds variance and is not differentiable. CREDO instead reads confidence deterministically from a dedicated token pair and trains it with differentiable regression, and it weights rollouts by the gap between confidence and outcome. On mathematical and code reasoning, CREDO achieves the best accuracy and calibration, with gains extending to abstention and selective prediction.
In briefAUV-Bench: A Benchmark for Evaluating Aesthetic Understanding and Generation in UI Design
SummaryThe paper introduces AUV-Bench, built with professional UI designers, containing about 1,395 executable web interfaces and four tasks: aesthetic scoring, diagnosis, repair, and text-to-UI generation. Evaluating 12 models, it finds moderate agreement with designers on holistic scoring, but exact diagnosis-chain success peaks at only 24.7%. Correct judgments and successful repairs often do not coincide, revealing a Judgment-Action Gap, and the authors conclude current models show only partial aesthetic competence.
In briefEviSplat: Keeping Multi-View Evidence in 3D Gaussian Splatting for Text-Based Object Segmentation
SummaryThis paper introduces EviSplat, a method for locating and segmenting objects in 3D scenes from free-form text queries. Instead of merging multi-view observations before the query is known, it keeps individual observation features as evidence and learns each Gaussian's supported visual appearances. For a given query, it scores instances using their most relevant observations and combines this with local evidence to score each Gaussian. Experiments across diverse datasets and protocols report state-of-the-art performance.
In briefLearning precise motion control where mistakes cannot be undone, shown with billiard shots.
SummaryThis paper studies high-risk, high-precision motion control, where actions have irreversible outcomes and the reward landscape has sharp peaks and ridges. Using computational pool as a representative example, the authors propose SCOOT, a deep reinforcement learning algorithm built on advantage-weighted regression. It adds elite-sample policy optimization, a mixture-of-experts policy, and a distance regularization term with a learning curriculum. The authors report that SCOOT learns physically based billiard shots with high action precision and discovers multiple shot strategies for a given ball configuration.
Nam Hee Kim, Markus Kirjonen, Perttu H\"am\"al\"ainen
In briefFrom Soft Targets to Reward Signals: How Target Assignment and Reward Objectives Interact
SummaryThis paper studies how assigning a fixed set of preference strengths to different response pairs changes the reward signals learned by different reward objectives. It introduces an assignment geometry framework comparing mean-matched smoothing with within-stratum reassignment. Across five reward objectives, soft targets that keep intact correspondence retained the largest clean preference margins within a common accuracy-equivalence budget.
In briefCombining sparse rewards and dense distillation: training dynamics of on-policy distillation
SummaryThis paper studies why training that mixes sparse reinforcement-learning rewards with dense teacher-distillation signals can become unstable. Using a neural tangent kernel analysis, the authors introduce a token-level cross-signal statistic measuring how reward and distillation gradients align at each position. They identify two failure modes: magnitude drowning and localized directional conflict.
In briefCan We Trust the Teacher? Decoupling Credit Direction and Magnitude for Self-Distillation
SummaryThe paper argues that when teacher supervision updates step-level credit direction and magnitude together, both become vulnerable to teacher errors. It proposes Decoupled Credit Self-Distillation (DCSD), which uses belief-margin probing for direction and marginal information gain for magnitude. Across 11 benchmarks, DCSD reports the best overall scores against GRPO, OPSD, RLSD, and RLCSD.
In briefContext-dependent time-series forecasting with HyperReservoirs
SummaryThis paper proposes the HyperReservoir, an extension of reservoir computing for time series containing multiple dynamical regimes. It pairs a main reservoir with a smaller context reservoir that modulates the main reservoir's output weights, while keeping linear-regression training. The model is compared with a conventional echo state network and a full-matrix Conceptor on Lorenz and Rössler prediction tasks. The HyperReservoir achieved the lowest mean test error on all three tasks.
Kohei Tsuchiyama, Takatomo Mihana, Ryoichi Horisaki et al.
In briefORAV: A Benchmark for Audio-Video Generation from Mixed Reference Inputs
SummaryThis paper introduces ORAV Bench, a benchmark for generating audio-video content from several heterogeneous reference inputs. It contains 380 task instances with 2 to 10 references each, spanning 9 semantic roles and 30 role compositions. The authors also propose a reference-aware pairwise evaluation protocol that reaches 86.08% effective agreement with human judgments. Testing five frontier systems, they find that strengths vary across reference compositions and that a recurring failure is reproducing unintended source content instead of the requested result.
In briefQiYao-M: A Multimodal Time Series Model That Treats Internal and External Data Differently
SummaryQiYao-M is a multimodal time series foundation model that models endogenous and exogenous modalities separately. Endogenous modalities are handled by a predictor and supervision scheme that learn how they evolve from history to the future. Exogenous modalities are handled by a retrieval enhancer that adapts downstream without updating model parameters, and it is trained using endogenous proxy data, without exogenous multimodal pretraining data. Experiments on unimodal and multimodal benchmarks show strong forecasting performance with and without exogenous modalities.
In briefAdapting Only Meaning, Not Structure: Few-Example Schema Tuning for Scientific PDF Extraction
SummaryThe paper addresses the costly manual tuning of extraction schema field definitions before large language models can reliably run them, especially when only a few verified extractions exist. It proposes CPSE, which calibrates extraction prompts and field semantics from a few gold annotations while keeping the output structure fixed, separating identity discovery from record completion. On expert-annotated polymer-science documents, CPSE improves extraction by 9.93 points over an execution-matched baseline, with gains confirmed by an independent judge and a blinded expert audit.
In briefPain Signals as Control: Sensing and Memory for Agents in a Single Body
SummaryThe paper studies agents deployed in one body that cannot learn wear rates through repeated trials. The authors propose a load-gated nociceptive channel paired with a memory of felt costs, which shifts allocation toward the best-paid work not yet felt. In simulations of 2,000 floor-layer knees anchored to published loss rates, working life extends from 55.2 to 59.6 years and career output rises from 33.7 to 36.1, with 69.3% of bodies gaining and none losing.
In briefOpenWhistle: A Large Public Dataset and Benchmark of Dolphin Whistles
SummaryThis paper introduces OpenWhistle, a public dataset of bottlenose dolphin vocalizations with about 180,000 whistles (114 hours) recorded over five years from a stable pod of five dolphins in a semi-natural setting. It includes 8,354 expert-annotated whistles and reproducible evaluation protocols for whistle-type detection and classification, along with the full processing pipeline. The authors pretrained a Wav2Vec2.0 model on this corpus, which outperformed general bioacoustic models AVES and BioLingual on both tasks.
Faadil Mustun, Chiara Semenzin, Roberto Dessi et al.
In briefDivOPD: Batch selection for asynchronous on-policy distillation of multi-turn agents
SummaryThis paper introduces DivOPD, a learner-side batch-selection method for asynchronous on-policy distillation of multi-turn student agents. It spreads a fixed turn budget across more rollouts and prioritizes turns with larger teacher-student disagreement, excluding turns without usable teacher feedback. Experiments on simulated ALFWorld, ScienceWorld, and WebShop benchmarks report higher success rates and geometric-mean speedups of about 1.84x in training tokens and 1.87x in learner GPU time over vanilla OPD.
In briefMW-Nowcast: Six-hour ensemble radar forecasting of extreme rain
SummaryThe paper presents MW-Nowcast, a six-hour ensemble radar nowcasting model. It combines a deterministic predictor for shared, organised rainfall structure with a generator that produces diverse local residuals around that prediction. On independent test data from the United States, Europe and China, it detects heavy and extreme rain more skilfully than leading methods across the 6-hour horizon. For the most intense rainfall, the authors report that warning time roughly doubles in all three regions.
WhyCombines deterministic prediction with generative residuals for six-hour ensemble nowcasting, validated on independent US, European, and Chinese data with doubled extreme-rain lead time.
In briefTransform-aligned learned features for lossy point cloud attribute compression
SummaryThis paper proposes Transform-Aligned Learned Features (TALF), which applies the attribute transform to learned spatial representations so they align with the coding targets. TALF is integrated into a transform-based attribute codec with coefficient prediction and conditional residual entropy modeling under a unified rate-distortion objective. Experiments on three benchmark datasets and multiple transform bases show improved rate-distortion performance over conventional and learned baselines.
In briefSemantic mapping for city scenes, with calibrated observations and regularized policies
SummaryThis paper addresses semantic mapping in urban environments, where objects range from pedestrians to buildings and viewing distances vary widely. The authors introduce a large-scale urban mapping dataset with realistic layouts, high-fidelity rendering, and multi-scale instance annotations. They propose a category-aware likelihood calibration policy, a mutual-information regularizer to reduce dependence between calibration and motion policies, and category-wise value estimators optimized jointly as a Pareto problem.
In briefBV Loss: A Training Objective for Block Diffusion Speculative Decoding
SummaryThis paper proposes Block Verification-aware loss (BV loss), a training objective for diffusion drafters in speculative decoding that aims to maximize the expected acceptance length of a drafted sequence under block verification. Across math, code, and chat benchmarks, using DFlash and DSpark with Qwen3-4B and Qwen3-8B, BV loss increases the mean number of accepted tokens per verification call by 13.0% to 21.0% over cross-entropy training, without changing the inference procedure. The authors report that BV loss also outperforms TV and LK losses and that its gains extend to token verification and greedy decoding.
In briefStructured neural SDEs for calibrating to path-dependent targets
SummaryThe paper introduces SLiSDE, a family of neural stochastic differential equation models built from structured linear stochastic layers that can be simulated in parallel over time. Expressivity comes from gated stacking, where previous-layer paths modulate the next layer's latent flow. For tasks where rare paths dominate the loss, an optional Girsanov tilt serves as a learned importance sampler with an exact likelihood-ratio correction. The authors report well-posedness, a discretisation error bound, universality of terminal laws, and better performance than fully neural SDE baselines on functional calibration benchmarks.
In briefOptimizing automatic task construction for scientific extraction agents from short goals
SummaryThis paper studies how to build the task-specific output schema, extraction instructions, and training rubrics for a scientific extraction agent from only a short goal and unannotated reference documents. The framework keeps schema construction and extraction instructions editable during optimization, focusing textual-gradient feedback on lower-scoring documents. On a heterogeneous-catalysis literature corpus, optimizing both schema construction and extraction instructions performs best across all four judge-rubric settings, with ablations and blinded human evaluation supporting the approach.
In briefSimulating whole survey respondents with coherent answers across questions, not just single items
SummaryThe paper proposes FR-LLM, which fine-tunes two LLMs: one for each item's marginal response distribution and one autoregressive model for dependencies across answers. A Marginal-Constrained Joint Projection step combines them so simulated questionnaires keep realistic cross-item relationships while staying accurate per item. On two real social survey datasets, FR-LLM better reproduces multi-question response patterns and generalizes to unseen populations and questions. In a small commercial-survey dataset, its simulated responses led to the highest realized profit in pricing and stocking decisions.
In briefWM-VLM: Internal World Models for Mixed Text-and-Image Reasoning
SummaryWM-VLM adds a lightweight world model branch to a pretrained vision-language model so it can generate intermediate visual states while reasoning. Training happens in two stages: first learning to generate the next visual state, then using it to answer questions. The authors build spatial reasoning tasks with verifiable intermediate states and test them on 2D and 3D mental rotation. WM-VLM outperforms the supervised fine-tuned backbone by up to 39.25 points, and removing or corrupting the generated states sharply reduces performance.
In briefGaussian Neural Networks: A New Regularizer Placing Priors on Activations
SummaryThis paper proposes Gaussian neural networks (GaNNs), which treat activities from earlier layers as signals with Gaussian noise and use an additional unsupervised loss to predict the properties of that noise distribution. During training, this unsupervised loss penalizes unexpected activities, allowing larger weight updates in less surprising directions. The authors report that GaNNs outperform standard neural networks on a range of classification and regression tasks. They also examine the models' ability to quantify uncertainty.
In briefGenerative Residual Factorization: splitting next-patch prediction uncertainty into removable and irreducible parts
SummaryUnder a shared-factor model, this paper factors the conditional distribution of the next image patch into a posterior over a shared scene factor and a residual kernel given that factor. It shows that conditional entropy splits into a residual entropy that no factor observation can remove and a posterior term that a better representation of the past can reduce. The paper also analyzes next-embedding prediction and notes that the split holds as an equality in a scalar Gaussian model, where it can be evaluated in closed form.
In briefCEO Arena: A Benchmark for Long-Horizon Multi-Agent Decisions in Competitive Markets
SummaryThis paper introduces CEO Arena, a benchmark for evaluating LLM-based agents acting as company CEOs in a competitive simulated market. Eight companies make sequential pricing, procurement, marketing, R&D and service decisions over 500 simulated days. Using matched replacement evaluation, the authors measure each agent's operating returns and its effects on rivals and the market. In the main evaluation, most agents have negative mean returns, and private gains can coincide with market losses.
In briefTesting Whether Directory and File Names Alone Improve One-Shot Code Search
SummaryThe paper tests whether a language model can use only directory and file names to find annotated code files that fixed lexical queries miss. On 82 issues from 11 repositories, name-only directory routing recovered 0.465 of gold files within eight candidates, above FTS5 at 0.352 and a full-issue rg query at 0.245. The authors note they cannot attribute the gain to hierarchy or show that it improves issue resolution.
In briefAccelerator choice alone is not enough for AlphaFold2 inference on Cloud TPUs
SummaryThis paper compares AlphaFold2 inference across a Colab CPU, an NVIDIA T4 GPU, and an eight-chip Cloud TPU v5e slice. A single TPU chip was much faster per call than the T4 in the same campaign, but the default execution path uses only one of the eight chips. The authors also find that software choices such as vmap and pmap strongly affect throughput, and that reruns five weeks later did not reproduce the cloud baselines.
In briefESTHER: Estimating and Reconstructing 3D Hands from Wearable Stereo Cameras in the Wild
SummaryThis paper presents ESTHER, a model that estimates and reconstructs metric 3D hand structure from egocentric stereo cameras worn by a person. The authors also build ESTHER3D, a dataset with model-generated labels for in-the-wild training and motion-capture ground truth for testing. Experiments report state-of-the-art accuracy and robustness to missing views, dropped frames, and lighting or motion blur.
WhyProposes an end-to-end egocentric stereo hand model and real-world benchmark, with metric scale retained after monocular collapse; experiments test generalization and robustness.
In briefPolylogarithmic Nash regret in matrix games with bandit feedback
SummaryThis paper studies minimizing Nash regret in unknown finite matrix games, where the learner receives bandit payoff feedback and observes opponent actions. The authors propose Optimistic Payoff Balancing (OPB), which achieves instance-dependent O(log²T) Nash regret against arbitrary adaptive opponents, including games with nonunique equilibria. The result extends the polylogarithmic guarantee of Maiti et al. (2025) from 2×2 games to arbitrary finite dimensions.
WhyResolves an open problem by extending polylogarithmic Nash regret from 2x2 to arbitrary finite matrix games, with a proof handling nonunique equilibria.
In briefUniOPSD: Combining Outcome and Hindsight Feedback for Training Language Model Agents
SummaryThis paper introduces UniOPSD, a method for training language model agents that merges environment rewards with hindsight feedback from successful peer responses. It uses adaptive local credit arbitration to decide how much each feedback source influences each decision. The authors report success rates on ALFWorld and WebShop and accuracy on Search-QA using Qwen2.5-3B-Instruct and Qwen2.5-7B-Instruct, and state a 7.0 point gain over SDAR on 3B WebShop.
WhyAdaptively arbitrates outcome and hindsight credit locally, with experiments on three agent benchmarks and a 7.0-point gain over SDAR.
In briefFrom Perception to Integration: Inside Reasoning in Vision-Language Models
SummaryThis paper studies why vision-language models struggle when a question requires several visual judgments, using controlled tasks on feature binding, numerosity, spatial relations, and amodal completion, plus a combined Composite task. Hidden-state readouts and interventions show that the answer becomes decodable during reasoning, often before the model stops on its own. The authors train a small detector to predict this readiness and stop reasoning early. On MMStar and RealWorldQA, this cuts mean reasoning tokens by 79.1% and 74.5%, while accuracy rises by 3.13 and 3.30 percentage points.
In briefControlTrace: Recovering the Control Field to Recognize Hidden Image Content
SummaryThis paper studies hidden content, such as words or contours, embedded in ordinary-looking images by spatially conditioned diffusion models, which vision-language models often fail to recognize. The authors build FreqBlind, a 6,000-image benchmark, and propose ControlTrace, which uses an 8.4M-parameter U-Net to predict the grayscale control field from the carrier image before a VLM identifies its content. In the evaluated setting, ControlTrace reaches 60.2% open-ended contour recognition accuracy, and recognition stays stable under JPEG compression, Gaussian noise and downsampling.:
In briefTemporal Binding Mechanisms in Large Audio Language Models
SummaryThis paper uses mechanistic interpretability to study how three open-source Large Audio Language Models represent sound events and bind them to temporal positions. It finds that event location concentrates in event name representations at intermediate layers, encoded along a low-dimensional, curved relative time trajectory. Steering these representations along this trajectory systematically shifts before/after judgments but does not reliably change predicted onset timestamps, suggesting coarse temporal reasoning and precise localization use distinct mechanisms.
In briefSIPO: Selective-Inference Policy Optimization for Tree-Structured Search Agents
SummaryThis paper addresses a statistical bias in tree-structured reinforcement learning for search agents, where branches selected for expansion may have values reflecting selection history rather than only continuation quality. The authors propose SIPO, which uses a scale-free branch criterion, exchangeable branching, and order-statistic correction to adjust credit estimation. Across seven QA benchmarks with Qwen3-4B, Qwen3-8B, and Qwen2.5-7B, SIPO reports the highest multi-hop and single-hop averages among compared methods, ranking first on six of seven benchmarks with Qwen3-8B.
In briefSeparating personal from population gains when tuning EEG models for new users
SummaryThis paper tests how well three pretrained brain-signal (EEG) models, CBraMod, REVE and LaBraM, adapt to 235 unseen people across three motor-imagery datasets. Personal adapters beat both the general population model and adapters borrowed from other subjects, but the advantage shrank when the population model was trained with more data. Calibrating with only a few labels was not reliably helpful. The authors recommend judging personalization against both a population baseline and swapped parameters, across different training budgets.
In briefPass or Fail? Testing LLMs on Two Greek Entrance Exam Benchmarks
SummaryThis paper introduces Prot-Ex and Pan-Ex, two Greek-language benchmarks built from questions in entrance exams for Greek Model and Experimental schools and the Panhellenic national university entrance exams. The benchmarks cover subjects such as Modern Greek, Mathematics, and Physics, with closed, structured, and open-ended formats, and evaluate several text-only LLMs. The localized KriKri-8B clearly outperforms its base model and rivals much larger models on linguistically demanding humanities tasks, while few-shot prompting improves closed-ended accuracy but degrades structured-task performance in 8B models.
In briefSTRIDE: Automatically Checking Whether Text-Described Scenarios Match Generated Pedestrian Trajectories
SummaryThe paper presents STRIDE, a framework for evaluating how well pedestrian trajectories match the context in scenario descriptions, without needing real human trajectory data. It uses the VRDST protocol, grounded in sociological theory, to break context into behavioral questions answered by a deterministic measurement tool library. The authors build STRIDE-Bench with 1K scenarios across 30 real-world maps and report 80% agreement with human judgment. Evaluation of several text-to-trajectory models shows limited context alignment and difficulty with fine-grained conditioning.
In briefInfiMed2: A general medical multimodal foundation model using stage-aware data and stability-based supervision
SummaryThis paper presents InfiMed2, a family of 4B and 27B general medical multimodal foundation models built around stage-aware data design. The authors curate a 55.68B-token corpus and adjust data mixtures across continued pretraining and post-training stages. For supervised fine-tuning, they regenerate visual question-answering responses using answer stability and related selection methods, and apply reinforcement learning with verifiable rewards to the 4B model. Across five medical multimodal benchmarks, the reported mean accuracies are 66.73% for InfiMed2-4B and 73.72% for InfiMed2-27B.
In briefCompact Underwater Image Enhancement via Physics-Guided Distillation for Robots
SummaryThe paper proposes physics-guided spectral distillation (PSD), a method to shrink underwater image enhancement models for real-time use on resource-constrained underwater robots. It uses a multilevel Haar wavelet transform to transfer low-frequency color and illumination information and high-frequency structural details, weighted by degradation-aware estimates and reliability masks. Experiments on UIEB, LSUI, and EUVP datasets validate the method, and deployment on a self-developed ROV is demonstrated.
In briefD²-VLA: A dual-memory model that helps robots remember cues during long moving-object tasks.
SummaryThe paper presents D²-VLA, a vision-language-action model that combines dual memory and dual-frequency control at the KV-cache interface of a pretrained policy. It introduces DOMINO-Long, a ten-task benchmark where robots must use earlier visual cues while manipulating moving objects. The authors report higher complete-task success rates than the compared baselines on DOMINO and DOMINO-Long, along with gains on real-robot tasks, LIBERO-Long, and RoboTwin 2.0.
In briefFinite-time convergence rates for projected two-timescale stochastic approximation with Markov noise
SummaryThis paper studies finite-time concentration and convergence rates for projected two-time-scale stochastic approximation driven by a controlled Markov chain, where the slow iterate is projected onto a compact convex polyhedron. Using the Skorokhod map, it gives explicit high-probability bounds for tracking the fast equilibrium and projected slow dynamics, and establishes almost-sure convergence with last-iterate rates under step-size conditions. The results are applied to actor-critic methods, projected TD(0), and projected stochastic gradient descent.
WhyGives explicit finite-time high-probability bounds and convergence rates for projected two-time-scale stochastic approximation with Markov noise, handling discontinuous boundary vector fields, support
In briefInstance-adaptive prompts as context for time-series foundation models
SummaryThe paper introduces PaCTS, which generates a small set of learned, instance-adaptive continuous prompt embeddings from the visible context to act as compact context for frozen time-series foundation models. The prompts are built from global statistics and refined with segment-level temporal information. Experiments report better forecasting across context lengths and architectures, with a shorter input outperforming double-context baselines at lower inference cost, and better out-of-distribution generalization than weight-space adaptation.
In briefBEHAVE: Self-improving hardware design and verification agents via functional behavior modeling
SummaryThis paper introduces BEHAVE, a framework where agents write executable behavior models alongside RTL hardware designs, without fixing implementation timing beyond the specification. A evaluator checks both artifacts against a hidden golden model using generated stimuli, and provides rewards for reinforcement learning. The authors release BEHAVE-Train and BEHAVE-Eval, with 600 human-reviewed specification-behavior pairs. Self-improvement raises Qwen3.8-27B's RTL pass@1 on BEHAVE-Eval from 55.0% to 75.0%, comparable to RL on a 540-task pool.
WhyProposes behavior-model verification replacing cycle-by-cycle matching, with a self-improvement loop; evidence rests on a single reported benchmark improvement.
In briefProjective Normal Fields: A Convex Method for Building Smooth Unsigned Distance Fields
SummaryThis paper introduces Projective Normal Fields (PNFs), which estimate bidirectional surface normals from point positions alone, without needing orientation. The method relaxes rank-one projectors to a convex set and solves a strongly convex optimization combining local tangent-plane fitting, soft-PCA anchoring, and overlap regularization. Heat diffusion and Poisson integration then build a regularized unsigned distance field approximation. Experiments report reduced sensitivity to neighborhood size and competitive results under noise, outliers, and near non-manifold junctions.
In briefTQTS-Bench: A Multi-Syntax Benchmark for Querying Time-Series Databases in Natural Language
SummaryThis paper introduces TQTS-Bench, a benchmark for testing how well large language models can turn natural-language questions into queries for time-series databases. It contains 6,125 question-answer pairs covering 97 databases, 23 query syntaxes, and 22 application domains, all reviewed by experts. The best model tested reached 48.98% execution accuracy, compared with 87.34% for humans, and the authors attribute the gap mainly to differing query syntaxes, misreading time-specific intents, and incorrect schema linking.
WhyA large expert-reviewed benchmark covering many time-series query syntaxes, a useful extension of existing text-to-SQL evaluation.
In briefCoHuB: A Simulation Benchmark for Multi-Humanoid Collaboration
SummaryThis paper introduces CoHuB, a simulation benchmark for evaluating collaboration among multiple humanoid robots using egocentric visual observations. It provides 10 tasks, eight involving two humanoids and two involving three. Synchronized demonstrations were collected through a multi-operator VR teleoperation pipeline, with each operator controlling one humanoid from its own view. Experiments with representative visuomotor policies reveal substantial challenges in coordinated perception and control.
In briefWhen Vision-Language Models Trust Context: Scene Text Reading Under Misleading Context
SummaryThis paper introduces SceneFaith, a benchmark of 781 generated scene images for testing whether vision-language models read the text shown or rewrite it to fit the surrounding context. Outputs are classified as Literal, Canonical, or Other. Across 15 models from seven families, all models rewrote text even on clear images, at rates from 8.45% to 58.51%.
In briefFailed Replication and Simple Baselines in Road-Level Crash Prediction
SummaryThe paper rebuilds the data pipeline of a recent uncertainty-aware graph neural network and tests eleven of its design choices across three London boroughs. Only four hold up on a second borough, and several reverse sign. The authors compare both networks with a parameter-free baseline that ranks road segments by cumulative past crash counts, finding the proposed model statistically indistinguishable from it at matched history depth. They argue that the apparent advantage of graph networks is largely due to short baseline horizons, and recommend horizon-matched baselines and multi-seed reporting.
In briefLarge Language Models in Clinical Medicine: Recent Trends and Outlook
SummaryThis paper reviews recent talks and articles on applying large language models (LLMs) in medical research and clinical practice. It notes that LLMs offer broader capabilities and flexibility with available data than traditional task-specific deep learning systems. It also discusses concerns about reliability, accuracy, transparency, safety, and their appropriate role in medical settings, weighing both opportunities and implementation challenges.
In briefPage-aware RAG for French PDF questions: a five-variant study
SummaryThis paper studies retrieval-augmented generation for questions about French PDF documents, evaluating both the answer and the supporting pages. Five variants add dense retrieval, rank fusion, reranking, and query decomposition to a BM25 baseline. On 595 challenge questions, the full system reaches 0.4450 MRR@10 and 0.4013 Recall@10, versus 0.3430 and 0.2994 for BM25. The authors conclude that accurate page selection, rather than semantic retrieval alone, is the main opportunity for improvement.
In briefBefore the Token Commits: A Trajectory-Level Benchmark for Visual Hallucinations in Diffusion Vision-Language Models
SummaryThis paper introduces DynaHall, a benchmark that tests visual hallucinations along the generation trajectory of diffusion vision-language models, covering object existence, counting, attributes, and relations. It records the intermediate answer tendency at each unmasking step, not just the final output. Across five models from three architecture families, unsupported answers were already preferred before the answer token was committed and were rarely reversed later. The authors also propose PGS, which steers still-masked answer states to reduce false positives, and report that it transfers to another architecture without degrading general ability.
WhyProposes trajectory-level hallucination evaluation and pre-commitment steering, showing hallucinations settle before commitment across five diffusion VLMs with reasonably broad experiments.
In briefWhen Do Model Internals Help? Representation Engineering for LLM Safety
SummaryThis paper compares behavioral alignment (DPO) with representation engineering methods for controlling and monitoring LLM safety under matched settings. DPO gives the strongest overall control, but its safety can degrade after later benign fine-tuning, while representation steering stays competitive mainly with little data. For monitoring, specialized text monitors detect unsafe content most accurately, while representation probes remain competitive at much lower marginal cost. The authors conclude that representation engineering does not generally replace behavioral safeguards but offers practical advantages in specific conditions and complementary safety benefits.
In briefReference-based data curation for rule-following Thai-English translation
SummaryThis paper addresses the trade-off between following prompt rules and translation quality in instruction-following machine translation. It proposes a two-phase pipeline that selects hard-but-learnable English-Thai examples using IFD scoring, then extracts constraints from reference translations and keeps only generations satisfying all of them, producing a 1.97M-record dataset. Models fine-tuned on this data, called ChindaMT, match or outperform same-size baselines and reach up to a 68.4% win rate against the strongest baseline under length-controlled pairwise judging.
Thodsaporn Chay-intr, Krittapad Harnchang, Mahannop Thabua et al.
In briefLongPuzzleBench: Testing GUI Agents on Long Visual Puzzles
SummaryThe paper introduces LongPuzzleBench, a benchmark of 114 levels across six puzzle games that agents must solve through native GUI actions. Success drops sharply on harder, longer boards: many general-purpose agents solve nothing beyond Medium, and none completes the hardest Bolt Unscrew level. Code-execution agents do not close the gap. The authors attribute the failures to agents judging each move by visible progress rather than by the future options it leaves open.
In briefVehicleArena: A Realistic City Simulator for Multiple Self-Driving Agents
SummaryThis paper presents VehicleArena, a 3D urban-driving benchmark for studying independent agents sharing one dynamic environment. LLM-controlled vehicles must fulfill changing passenger requests in complex traffic, and each driving decision affects traffic flow, delays, risks, and what nearby agents observe. The benchmark includes 112 single-agent and multi-agent tasks.
In briefBalanced Fitting: Post-training quantization for vision-language models that goes beyond reconstruction loss
SummaryPost-training quantization makes large vision-language models cheaper to deploy, but it is usually calibrated on a small dataset and must generalize to many downstream tasks. This paper proposes Balanced Fitting, which measures how quantization affects each layer and component (weights, vision activations, text activations), applying fine-grained fitting to sensitive parts and coarser fitting elsewhere to exploit possible regularization benefits. Experiments on multiple LVLMs show it outperforms prior PTQ methods in both weight-only and weight-activation settings, and that lower reconstruction loss does not reliably yield better downstream performance. The source code is publicly available.
In briefKT-EGO: A Knowledge-Transfer Optimization Method for High-Dimensional Costly Black-Box Design Problems
SummaryMany engineering design problems require optimizing high-dimensional black-box functions that are expensive to evaluate. This paper proposes KT-EGO, an extension of the efficient global optimization (EGO) algorithm for problems with more than 20 dimensions. It splits the design space into low-dimensional subspaces and uses a surrogate-based data fusion strategy to transfer information between them, along with an adaptive variable-range search to exploit promising regions. KT-EGO is compared with state-of-the-art algorithms on 12 benchmark functions and a 28-dimensional compressor blade design problem.
In briefWeaveData: A Multimodal Data Analysis System with Self-Critiquing, Self-Evolving LLM Plans
SummaryThis paper presents WeaveData, a system that answers natural-language questions over tables, text, and images by having large language models generate analysis plans. It critiques each plan step by step before execution and checks whether the result answers the question. Failing or off-target plans are diagnosed with the actual data and repaired, with planning experience accumulated for later questions. The authors demonstrate the system on two public multimodal datasets.
In briefPanoVLN: Panoramic Vision-and-Language Navigation for Better Route Decisions
SummaryThis paper studies vision-and-language navigation using panoramic images, where models predict movement actions from visual views and language instructions. The authors propose PanoVLN, which adds confidence-guided execution of multi-step action predictions, training routes with frequent branch points, and fused semantic and geometric panoramic features. With a 4B backbone and RGB-only input, it improves success rate over the previous state of the art by 11.9% on R2R-CE and 8.7% on RxR-CE Val-Unseen. Real-world tests on a quadruped robot also show faster navigation with fewer pauses.
In briefFine-tuning from pretrained initialization reduces data needs in diagonal linear networks
SummaryThis paper studies, using sparse linear regression and two-layer diagonal linear networks, how fine-tuning from pretrained weights helps downstream tasks. Pretraining provides information through the support and signs of the initialization predictor, which reshapes implicit bias and training dynamics. For a clean initialization with correctly inherited signs, the required sample size is comparable to a weighted Lasso that explicitly uses the pretrained support, reducing the data needed to recover target parameters.
Alexandre Decl\`eves, Etienne Boursier, Nicolas Flammarion
In briefDraft-KV: Learning Useful Latent Communication Between Language Models
SummaryThis paper studies language models passing internal states to each other instead of decoded text, and shows that higher receiver accuracy alone does not prove the receiver used the message content. The proposed Draft-KV method sends the key-value states formed while the sharer drafts an answer, with both models kept frozen and only 1.05M interface parameters trained. With a Qwen3-8B sharer, a frozen Qwen2.5-0.5B-Instruct receiver reaches 78.04% on MMLU-Redux, up from 37.45% alone. The authors report that the communication also transfers to held-out tasks.
WhyProposes draft KV-state latent communication with mismatched-message controls showing genuine content use; strong evidence, moderate subfield impact.
In briefA unified framework for concept-based explainable AI with completeness guarantees
SummaryThis paper proposes a theoretical framework that describes concept-based explanations of neural networks in one mathematical language. Concept discovery is modeled as a concept autoencoder, where an encoder extracts concepts from a model's latent space and a decoder reconstructs the original representation. The authors show that the incompleteness of the concepts, meaning how poorly they reproduce model outputs, is bounded by the reconstruction error. They also establish when individual concept attributions sum to the prediction and bound the discrepancy otherwise.
Vojt\v{e}ch K\r{u}r, Adam Kuku\v{c}ka, Tom\'a\v{s} Br\'azdil et al.
In briefCoDrive: Generating consistent multi-vehicle driving videos with precise camera path control
SummaryMost driving world models generate video from a single vehicle's viewpoint, and simply extending them to multiple vehicles does not ensure a consistent shared world. CoDrive jointly generates multi-view videos for several vehicles in the same dynamic scene, with precise camera-trajectory control. It represents all camera trajectories in a shared world coordinate system and combines real single-agent data with synthetic cross-agent interaction data during training. The authors also introduce CoDrive-Bench, and experiments report better trajectory controllability and cross-agent geometric and instance consistency while keeping competitive visual quality.</summary>
In briefNo Pain, More Gain: A Method That Iteratively Merges Teacher Models for Better Multi-Teacher Distillation
SummaryThe paper studies multi-teacher on-policy distillation (MOPD), which distills several domain teacher models into one student model using the student's own generated samples. The authors find that MOPD can struggle to recover some teacher capabilities, and that the initial benchmark score does not reliably predict how well a merge initialization will perform after distillation. They propose IM-MOPD, which starts from a uniform merge and progressively adds task-vector increments for under-recovered domains during training. In a 5-domain setting, IM-MOPD achieves higher average normalized recovery than MOPD with uniform merge initialization or SFT warm-up.
In briefDo Emotion Concepts Transfer Across Data Sources, Modalities, and Model Architectures in Vision-Language Models?
SummaryThis paper examines whether vision-language models encode emotion concepts in consistent ways across images, text, and architectures. The authors built CMES, a multi-source dataset of emotion-conditioned stories, real facial expressions, synthetic portraits, and synthetic emotion scenes, and extracted six Ekman emotion vectors from three models. They report that image-derived vectors share a low-dimensional geometry with text-derived ones, that cross-modal and cross-architecture correspondences persist, and that a shared emotion subspace can be built and generalized to a held-out architecture.
In briefPredictive dual smoothing to speed up column generation
SummaryThis paper studies column generation, a method for solving large linear programs by iteratively adding variables. Standard dual smoothing stabilizes dual solutions using past iterations but may not steer pricing toward useful new variables. The authors propose predictive dual smoothing, which blends the current dual with a learned prediction of future duals and modifies only the pricing objective, using exact reduced-cost checks and fallback pricing to preserve correctness. Experiments on cutting stock and generalized assignment problems report fewer generated columns and less wall-clock time than standard and existing stabilization methods.
In briefBeyond Token Alignment: Event Completion for Cross-Tokenizer On-Policy Distillation
SummaryThis paper studies knowledge distillation between language models with different tokenizers, where one teacher token may need several student tokens to generate. It proposes Event-Set Completion Distillation (ESCD), which aggregates prefix-related teacher events and supervises the total probability of byte-compatible one-step student completions. The authors report consistent gains in mathematics, code, and scientific reasoning, including large-scale MoE distillation from a 1T teacher to a 35B student.
In briefSeLMRoute: Probabilistic Semantic Evidence for Routing Queries to Language Models
SummaryThis paper introduces SeLMRoute, a framework that selects the most suitable large language model for each query. It first extracts candidate-independent semantic evidence, such as reasoning requirements and external knowledge use, stored as probability distributions, then uses a lightweight router to estimate each model's performance. On LLMRouterBench, it achieves 72.08% ± 0.45 average accuracy, and 72.64% in grouped five-fold out-of-fold evaluation, versus 69.23% for the strongest fixed candidate.
WhyDecouples candidate-independent probabilistic semantic evidence from performance estimation, validated on 11,481 queries with grouped cross-validation beating the best fixed model.
In briefFrom Human Narrative to Harmonic Structure: Studying Algorithmic Music Generation Through Chord Wheel Diagrams
SummaryThis paper compares the original chord progressions of six narrative-driven popular songs by Bob Dylan, Johnny Cash, and Ritchie Valens with harmonizations produced by an explainable computational harmonizer working on the same melodies. The authors analyze harmonic vocabulary, functional persistence, repetition, non-diatonic events, and tension-resolution patterns using Chord Wheel Diagrams and BPMN-based representations. They find that high melody-chord compatibility does not necessarily preserve the original human harmonic decisions. The study proposes narrative-conditioned harmonic structure as a complementary perspective for computational music analysis.
Josef Pavl\'i\v{c}ek, Petra Pavl\'i\v{c}kov\'a, Irena \v{S}trausov\'a
In briefSurgGMF: Causal Gaussian Motion Forecasting for Predicting Future Surgical Scenes
SummarySurgGMF forecasts how deformable surgical scenes will change by predicting future Gaussian motion states rather than future images directly. It uses a full-causal-last rendering protocol to prevent target-frame information from leaking into predictions. Tested on 12 EndoNeRF and StereoMIS video slices, learned forecasting methods outperformed classical dynamics baselines in render space. TKAN achieved the highest accuracy, while GRU and LSTM showed lower module-level latency.
In briefWhat Visual Generators Need from Teachers: Rethinking Representation Alignment
SummaryThis paper studies representation alignment, which speeds up diffusion transformer training by pulling a student block toward a frozen pretrained encoder's features. It introduces a recoverability gap that measures how much of each teacher layer's increment an unaligned student can recover, and uses it to pick the teacher layer. The resulting RARE method reaches FID 18.02 without guidance and 4.46 with it on SiT-B/2 with ImageNet 256×256, using about 14% fewer GPU-hours than iREPA.
In briefInformation-theoretic bounds for next-token prediction on Markovian data, linking context length to generalization
SummaryThis paper develops an information-theoretic framework for how next-token prediction models generalize when training data are temporally dependent. It uses mutual information for algorithmic dependence and mixing properties for temporal dependence, deriving expected generalization bounds for cross-entropy loss and extending them via a rate-distortion formulation. Experiments on TinyStories and ETTh2 show that longer contexts can lower both training and test loss, but often reduce training loss more, enlarging the generalization gap. An ETTh2 analysis points to an effective predictive-memory scale near 24 hours.
Masoud Kavian, Abdellatif Zaidi, and Milad Sefidgaran
In briefEdge-level automorphism in GNNs: a measurement method and designs that improve link prediction
SummaryThis paper studies why standard graph neural networks struggle to predict links between nodes that play identical structural roles. It introduces the edge automorphism ratio (EAR), a scalar metric measuring how well a GNN can distinguish links, and reports that EAR correlates strongly with link prediction performance. Based on this, the authors propose EO-GNN, which uses automorphism-aware dropout and subgraph orbit-biased aggregation while preserving equivariance. Experiments on synthetic and real graphs report gains of up to 42.36% and 28.44% in high-automorphism settings.
In briefDynamic Flow, Static Graph: KV cache reuse for faster on-device LLM serving on mobile NPUs
SummaryThis paper addresses running large language models on mobile NPUs, where statically compiled computation graphs and limited memory and I/O bandwidth make cloud-GPU KV caching designs unsuitable. The authors propose a compute-storage co-design that includes selective KV recomputation within static NPU graphs, an inter-graph scheduler for chunk merging, a hierarchical KV manager, and a pipeline that overlaps data movement with NPU execution. In experiments, the design reduces time-to-first-token by 40-60% compared with no reuse and prefix-only caching.
Zhengxiang Huang, Shengheng Chen, Chaoyue Niu et al.
In briefGeometry as Address: Routing Attention to Visual Memory for Long Camera-Controlled Video
SummaryThis paper addresses long-horizon camera-controlled video generation, where earlier observed content must be recovered from a growing visual history. It introduces GEAR, which uses geometry as a token-level address into frame-latent memory instead of building a global 3D representation, avoiding accumulated geometric errors. Geometric Correspondence Attention injects matched historical features into noisy target patches, and an Invisible Octree filters occluded correspondences. The authors report state-of-the-art visual quality, camera control, and revisit consistency, including minute-long videos.
In briefDBCF: Dual-Branch Fusion of Foundation Models for General Deepfake Detection
SummaryThis paper addresses the poor generalization of face forgery detectors across datasets and unseen manipulation types. It proposes a dual-branch framework: a CLIP-based Global Context Branch captures holistic semantic cues, while a DINOv3-based Fine-grained Cue Branch captures local structural irregularities. A feature fusion module adapts the frozen backbones parameter-efficiently. The authors report strong results on multiple benchmarks, especially in cross-dataset and cross-manipulation settings.
In briefQuality Sets Direction, Length Sets Size: Length Control in Open-Ended RL (QGLAS)
SummaryThis paper addresses how reinforcement learning makes language model responses longer and less token-efficient. It proposes Quality-Gated Length Advantage Shaping (QGLAS), which computes advantages from quality rewards alone and adds bounded bonuses only to shorter responses with positive quality advantage, scaled by within-group quality separation. The authors report that across models, benchmarks, and reward sources, at about 30% compression QGLAS retains 98.4–102.0% of quality-only RL's macro-average gains, versus 68.3–75.5% for baselines.
In briefReMCTS: Reflection-Enhanced Monte Carlo Tree Search for Code Generation
SummaryThis paper presents ReMCTS, a framework that uses an MCTS-style search guided by an LLM, execution feedback, and memory to generate function-level code. It organizes candidate programs as tree states, keeps branch-local debugging context, and retrieves failure experience across branches. On HumanEval and MBPP-Sanitized, it outperformed direct generation in 8 of 10 model-dataset pairs under held-out evaluation. A 30-task C++ pilot on HumanEval-X shows compatibility with compiler-backed execution but is not a broad multilingual evaluation.
In briefFrom Pixel Generation to Topological Inference: Trustworthy Bone Trabecular Structure Learning Across Imaging Domains
SummaryClinical CT cannot resolve individual bone trabeculae, while synchrotron microCT offers high resolution but cannot image living patients. The paper proposes a structural dual super-resolution network (SDN) that predicts invariant microstructure from low-resolution inputs instead of generating pixels, evaluated with morphological parameters. On the test set, SDN's six morphological metrics largely match synchrotron references, with SSIM reaching 0.8. The model trained on 3.2 µm data also generalizes to an independent 3.25 µm dataset.
