In briefSoei: Structured Evidence and Intent Tracking for AI Coding Agents
SummaryThis paper examines how to tell what AI coding agents actually did during long, multi-session software work, and whether it is finished. The authors verify an open-source observability tool and analyze 45 real coding-agent sessions, finding that 23.4% of code-editing episodes lack an independent check after the last edit and 57.8% of completed to-dos had no proof. They propose Soei, a system with a hash-chained local ledger, a PROV-aligned graph, and a verification layer that separates completion claims from proof.
In briefCATP: Calibrated attention-based token pruning to speed up multi-image generation in unified multimodal models
SummaryUnified multimodal models generating multiple images from long visual contexts require heavy computation. This paper proposes CATP, a training-free context-compression method for such models that synchronizes pruning decisions across conditioning branches and corrects pooling-induced bias with a second-order Taylor-based attention metric. A fused scoring operator folds this calibration into tiled online Softmax to lower pruning cost. Experiments reportedly show better quality-efficiency trade-offs than prior token-reduction methods, with up to 2.09× attention-core speedup.
In briefCBE: A complexity-adaptive ensemble method for imbalanced classification data
SummaryThis paper proposes CBE, an ensemble learning framework for classification on imbalanced datasets. It links the sampling weights of boundary samples during under-sampling to data complexity and adjusts boundary preference dynamically. The authors also introduce an overall complexity measure combining multi-level metrics to divide imbalanced datasets into three categories. Experiments on 24 benchmark datasets from UCI and KEEL show better F-measure and G-mean than many existing methods on Easy and Median datasets, without performance loss on hard cases.
In briefCombining heterogeneous echocardiography AI outputs through a learnable late-fusion module built on a supervised large language model
SummaryThe paper proposes a late-fusion module based on a supervised large language model that aggregates view-specific predicted finding statements and predicted measurements directly in language space, learning a case-level fusion function. In patient-level three-fold cross-validation on 859 examinations from 646 patients, the method performed comparably to the best reference strategy evaluated and outperformed a general-purpose LLM without task-specific training.
Ryo Takizawa, Tempei Kabayama, Chiemi Yamazaki et al.
In briefMinimax-optimal first-order and best-of-both-worlds regret in multi-armed bandits
SummaryThis paper studies first-order and best-of-both-worlds regret guarantees for multi-armed bandit problems. The authors propose an efficient, parameter-free algorithm that removes a logarithmic factor from a prior result, achieving the minimax optimal first-order regret O(√(K L*)). The method builds on FTRL with a Tsallis entropy regularizer. For i.i.d. losses with a unique best arm, it also achieves O(∑ log T/Δ_i) regret.
In briefSparse fine-tuning that updates only key parameters for efficient adversarial training
SummaryThe paper addresses high memory costs of adversarial training on large models and the robustness gap of LoRA-based parameter-efficient methods. Gradient analysis shows adversarial optimization concentrates on a small set of Robustness-Critical Parameters. The authors propose SAFT, which identifies these parameters via adversarial gradient saliency and updates only them. According to the paper, SAFT outperforms LoRA-based methods in clean and robust accuracy and approaches or surpasses full-parameter adversarial training while updating about 5% of parameters.
In briefKairos: a planner orchestrator that picks among engines to handle all kinds of PDDL tasks
SummaryThe paper presents Kairos, a system that profiles each PDDL planning task and selects among five planning engines to cover everything from STRIPS to PDDL+. Every candidate plan is checked by a validator independent of the planner's transformations. On 1,756 tasks from nine comparison suites, Kairos solved more tasks at the full limits than the strongest evaluated planner of each suite.
WhyThe authors select heterogeneous planning engines by task semantics and schedule them under one budget, reporting comparisons on 1,756 tasks, ablations, and a VAL cross-check.
Francisco Mart\'in Rico, Esther Aguado Gonz\'alez, Francisco Miguel Moreno Olivo et al.
In briefSmooth action chunking for imitation learning using movement-primitive representations
SummaryThis paper addresses jerky robot motion caused by action chunking in vision-based imitation learning policies. It represents each trajectory segment as a linear combination of smooth basis functions, so trajectories are smooth by construction without changing the policy architecture, training objective, or inference procedure. It also proposes a retraining-free, inference-time correction that passes each new chunk through the previous chunk's last commanded action to reduce boundary discontinuities.
In briefEnd-to-end guarantees for masked diffusion models under conditional mixing
SummaryThis paper gives end-to-end theoretical guarantees for a learned denoiser in masked diffusion models, for stationary sources whose distant tokens have exponentially decaying influence. It shows that the empirical risk minimizer trained under any masking distribution achieves small per-token excess error with a number of training sequences independent of sequence length, and that this exponent order is minimax-optimal up to constants. It also bounds the end-to-end KL divergence of sampling under a dilated schedule, which requires far fewer denoising steps than a fixed random order.
WhyThe authors give end-to-end guarantees for learned masked diffusion denoisers under conditional mixing, with sample complexity independent of length and a minimax-optimal exponent.
In briefSoftmax routing collapse and recovery in mixture-of-experts models
SummaryThis paper studies how softmax routing in mixture-of-experts models concentrates probability on a few experts, suppressing the rest, and how those suppressed probabilities recover. It analyzes fixed-expert loss barriers and routing limits in a square-loss model with trainable expert outputs, deriving logarithmic or power-law recovery times. In paired Transformer-MoE training, removing a class-conditional router-gradient contrast delays recovery and reduces accumulated log-odds on fixed tokens.
WhyThe authors analyze softmax routing collapse and recovery in MoE, deriving recovery-time orders and reporting paired Transformer-MoE ablations.
In briefPathCowork: Building AI Copilots from Real Pathology Practice
SummaryThis paper introduces PathCowork, a framework that turns pathologists' narrated slide examinations into training data for visual grounding and diagnostic AI agents. Gaze, narration and navigation are recorded in sync to link visual findings with how they were interpreted. The data come from 48 sessions by 16 pathologists totaling 1,028 minutes of slide review, used to train server and local models.
In briefTimbre: a label-free multimodal test-time reinforcement learning method using reliability-aware harmonic consensus
SummaryThis paper proposes Timbre, a framework that adapts vision-language models at test time without labels, using rewards derived from the models' own predictions. It builds a 2×2 grid of original and perturbed images with direct-solving and reframing modes, selects pseudo-labels by weighted harmonic aggregation, and adjusts modality weights based on batch-averaged shifts in pseudo-label confidence. Across six benchmarks and three Qwen3-VL model scales, it outperforms TTRV in most settings and transfers to held-out instances.
In briefATLAS-Bench: An Agentic Tool-Use Benchmark for Local Area Search
SummaryThis paper introduces ATLAS-Bench, a benchmark of 1,200 local search queries across three regions for evaluating LLM agents. Each query separates mandatory requirements from preferences and factual conditions from subjective judgments, with evidence for checking each requirement. The authors also present ATLASBox, a standardized environment with search tools and a fixed point-of-interest knowledge base. Evaluation across 14 configurations of 11 models shows that constraint satisfaction and providing usable recommendations do not always align.
In briefVerifiable LLM-guided focal SMT solving for quantified arrays
SummaryThis paper proposes LinguaArray, which uses an LLM to suggest a small set of focal ground terms that guide a conventional SMT solver on quantified array problems. The solver alone remains the proof engine and certifies all results. On 1015 SMT-LIB instances, it solves up to 69.1% of cases and increases solved counts by 105% to 397% over the backend solvers.
In briefFDSML: Formula-driven metric learning that pre-trains vision models on synthetic data using generation parameters rather than class labels
SummaryThis paper proposes Formula-Driven Supervised Metric Learning (FDSML), which pre-trains vision models without real images by deriving supervision from image-generation parameters instead of discrete class labels. It combines FineFractalDB, a refined fractal dataset, with a Parameter-Aware Triplet loss that defines positives and negatives by distances in the Iterated Function System parameter space. On ResNet-50, Recall@1 rises from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars compared with FDSL pre-training.
In briefDiagnosing the Reference Gap: Limits of Visual Scaffolds in Multimodal Reasoning
SummaryMultimodal language models can recognize images but often fail to reason over them across multiple steps, such as losing track of counted objects. The authors argue this stems from a reference gap, where language is not tied to the relevant visual entities, regions, or locations, and test whether external references like numbering can narrow it. They introduce ReGap-Bench, five visual reasoning tasks where only the reference varies. Task-matched references improve accuracy by up to 25%, while mismatched references help little or even hurt performance.
In briefLinQuant: Invariance-Based Low-Precision Quantization for Linear Recurrent Models
SummaryThis paper addresses low-precision inference for linear recurrent models, which are emerging alternatives to softmax attention in large language models. The authors identify seven functional invariances, including rotations, diagonal rescalings, and alternative placements of decay and write factors, that preserve full-precision computation while making operands easier to quantize. They build LinQuant on these invariances and apply it to Mamba-2 prefill, improving perplexity at matched latency on Mamba-2 1.3B, 2.7B, and 8B and Nemotron-H 8B.
Neehal Tumma, Philipp Nazari, Rohan Cherukuri et al.
In briefAre SWE-bench Tasks Clear and Well-Tested? An Experience-Guided Self-Evolving Assessment Tool
SummaryThis paper examines SWE-bench-style datasets used to evaluate LLM-based coding agents, focusing on whether task descriptions are clear and whether tests adequately cover the required behavior. The authors propose ExLabel, a self-evolving, experience-guided LLM assistant that maintains an experience bank to make context-aware quality judgments. They also introduce SWE-TQ, a benchmark covering eight programming languages built from cleaned human annotations, on which ExLabel achieves the best F1 scores across settings.</summary>
In briefDecoding Geometry, Not Tokens: Efficient VLMs for Robot Affordance and Trajectory Prediction
SummaryThis paper addresses robot grasp location and motion trajectory prediction, which the authors argue is a geometry problem that standard vision-language models handle poorly by generating coordinates as text. They show that compact 0.5 to 2B parameter models already encode rich spatial information in intermediate hidden layers. They add a lightweight prediction head on a frozen backbone that regresses affordance boxes and trajectory waypoints as continuous values. On ShareRobot-Bench and AGD20K, the 2B variant cuts trajectory error by over 60% and raises affordance mAP by 85% relative to the autoregressive baseline.
Akila Ayanthi, Darshana Priyasad, Tharindu Fernando et al.
In briefAccuracy Is Not Compatibility: Two Evo 2 Checkpoints Depend on One FP8 Matrix Multiplication
SummaryThis paper examines why two Evo 2 DNA language model checkpoints fail on Blackwell GPUs. The authors attribute the failure to how FP8 matrix multiplication sums products, not to how values are stored. Emulating only the first projection with H100-style FP8 arithmetic substantially raises BRCA1 variant-classification AUROC.
In briefAnatomy-Guided Routing for 3D Lung Tumor Segmentation in CT
SummaryThis paper addresses 3D lung tumor segmentation in CT scans, where tumors are small and look different across anatomical locations. It proposes an Anatomical Prior-guided Routing Module (APRM) that uses regional occupancy priors to select a routing gate for each image patch and perform sparse feature routing over a shared expert adapter pool. The authors report consistent improvements over baselines on two datasets with various backbone networks.
In briefHow well deep research agents handle vague user questions
SummaryThis paper introduces UNDERSPEC-DR, a framework for systematically testing how underspecified user queries affect deep research agents (DRAs). Across three evaluated DRAs, vague queries substantially reduced how well the final answers met users' information needs, although the answers stayed grounded in cited evidence. The analysis of intermediate searches suggests agents run broader searches that rely less on the original query and more on learned priors, which may not match user intent.
In briefDobb: Shard-Local Preconditioning in Block-Diagonal Eigenbases, a distributed optimizer that updates parameters locally for large model training
SummaryThe paper proposes Dobb, a matrix-aware distributed optimizer for training language models whose parameters exceed a single node's memory. Under tensor parallelism or FSDP, it computes updates locally within each parameter shard, requiring only data-parallel synchronization of activation second moments. On GPT-2 XL (1.56B parameters), Dobb with 16 blocks reached a validation loss of 2.476, versus 2.511 for Dion at rank fraction 1/16. Under FSDP on 32 H100 GPUs, the amortized optimizer-call time dropped from 124.1 ms for Dion to 9.2 ms.
In briefKeeping Multiple Repair Hypotheses for Cost-Aware Execution
SummaryMultiple corrections of an executable model can match the same observations yet predict different execution outcomes. The paper proposes a repair-to-execution interface that keeps compatible finite guard repairs, updates their weights from diagnostic and normal-execution observations, and uses their predictions to choose routes, priced tests, or a backup. On 24 synthetic two-stage devices, the AI-weighted belief lowered no-test cost by 2.14 percent relative to committing to the generated repair.
In briefSafe multi-objective reinforcement learning with zero constraint violation and convergence guarantees
SummaryThis paper studies multi-objective reinforcement learning that balances several objectives while never violating safety constraints during learning, covering both cumulative and stepwise constraints. The authors combine optimistic reward estimates with pessimistic cost bounds to learn a safe Pareto optimal policy for a given preference, with high probability guaranteeing zero violation for every executed policy. They also present a preference-free method that explores safely first and then computes a safe policy for any preference without further interaction.
WhyThe authors propose zero-violation safe MORL with preference-free exploration and report Õ(1/√T+1/√K) suboptimality bounds.
In briefYOLOF-MLD: A Multi-Level Distillation Framework for Efficient Object Detection
SummaryThis paper presents YOLOF-MLD, an object detection framework for resource-constrained edge devices that combines a YOLO26n backbone, YOLOF's single-input-single-output neck, and multi-level distillation. On the D-Fire and AFS fire-smoke datasets, it reduces YOLO26n's parameter count by 35.7% and achieves a 62.3% speedup on Raspberry Pi 5. It reaches 71.6% and 59.9% mAP@50, compared with 75.5% and 60.0% for YOLO26n, offering a competitive accuracy-latency trade-off for edge deployment.
In briefSymKL: A Unified Measure of Signal Degradation in Representation Space
SummaryThis paper introduces SymKL, a full-reference measure of signal degradation based on implicit neural representations (INRs). An INR is fitted to the pristine reference, and the symmetric KL divergence between the diagonal Fisher-information parameter-sensitivity profiles of the reference and the degraded signal gives the degradation score. The method needs no knowledge of the corruption process or its severity and applies to images, audio, and 3D data. Experiments on image, speech, and volumetric 3D benchmarks report that SymKL tracks degradation severity and stays monotonic where conventional metrics saturate or fail.
Dhananjaya Jayasundara, Sudarshan Rajagopalan, Vishal M. Patel
In briefACD-GS: Curvature-aware densification method for more compact 3D Gaussian Splatting
SummaryACD-GS is a training method that improves how 3D Gaussian Splatting (3DGS) adds Gaussians during optimization. It uses second-order depth responses as a structural cue to asymmetrically control cloning and splitting, and adds a multi-view photometric consistency constraint and a capacity-aware adaptive regulation strategy. The authors report that it reduces Gaussian count and storage by nearly 60% while keeping competitive rendering fidelity across multiple datasets.
In briefConsistency of Chatbot Judgments on U.S. Presidential Actions and Hypothetical Reforms
SummaryThe paper tests how consistently consumer chatbots judge the long-term effect on the United States of 40 actions from the last four presidential administrations and 15 hypothetical reforms. Over 30 days, five chatbot configurations scored each item twice daily. All configurations agreed on the sign and ranking of the administrations' mean scores, though same-day pairs gave opposite signs about 4% of the time. Judgments shifted across days and at three points where the service itself changed. The authors release a dataset of 16,500 judgments.
In briefCASE: Concept-Aware Segmentation Using Positive and Negative Visual Examples
SummaryThis paper introduces Specified Visual Prompt Segmentation (SVPS), where a target concept is defined by both positive and negative visual prompts rather than positive examples alone. The authors build the CASE benchmark on COCO and PACO, covering six task families such as exclusion, isolation, and union. They find that the frozen SAM 3 model does not consistently treat negative prompts as evidence for exclusion. They propose Concept Prompt Adaptation (CPA), a LoRA-based method that improves positive-negative prompt interpretation on CASE while staying competitive on positive-only few-shot segmentation.
In briefOPERA: An expert-taught AI agent that designs CMOS op-amps from specs to layout
SummaryThis paper presents OPERA, a general-purpose large language model agent that designs CMOS operational amplifiers from specifications through post-layout sign-off using open-source tools. The agent is taught the gm/ID sizing method by a designer and then learns the gLayout framework to write a reusable layout generator. In SKY130, six opamps and a shared bias circuit were designed, with all layouts passing DRC and LVS and meeting every target across 15 PVT corners. The benchmark netlists are publicly released.
WhyThe authors teach an LLM agent the gm/ID method and a reusable layout generator, reporting six SKY130 opamps taken from specifications to DRC/LVS-clean layouts meeting targets at 15 PVT corners.
In briefUMI-Live: An Open-Source Bimanual System for Real-World Robot Data Collection
SummaryThis paper presents UMI-Live, an open-source, full-stack bimanual Universal Manipulation Interface system for collecting robot training data in real-world settings. Its custom hardware integrates cameras, LiDAR-inertial sensing, onboard compute, and power, requiring no external tracking or in-environment calibration. The authors fine-tuned π0.5 and MolmoAct2 on the collected data, reporting success rates of 10/10 and 8/10 on a bimanual pick-and-place task. They also release 40 hours of data across three tasks along with hardware designs and software.
In briefAgent-driven hardware-software co-design of AI infrastructure via progressive formulation and verification
SummaryThe paper presents Formulaic, a workflow in which a language-model agent writes the mathematical formulation of an AI infrastructure design problem, a solver optimizes within it, and verification checks its assumptions and results. The agent refines the problem step by step from the computation graph toward the chip, using program templates with stated correctness conditions. On NVIDIA GB10, agent-written templates ran complete Qwen3-8B layers 1.075 times faster than SGLang.
In briefAgentDecept: Testing Whether Tool-Using AI Agents Lie About Their Actions Under Pressure
SummaryAgentDecept is an executable benchmark for measuring behavioral deception in LLM agents under pressure. It contains 477 pressure variants derived from 257 source tasks across six pressure sources, each paired with a pressure-free counterpart. The evaluation protocol compares an agent's user-facing claims with time-ordered observations and executed outcomes, distinguishing falsification, concealment, and equivocation from honest failure and disclosed limitations.
In briefWhat Makes Medical Exam Questions Hard? Using AI Models to Find Patterns Behind Empirical Difficulty
SummaryThis study used the empirical difficulty of USMLE and PANCE medical exam items to guide large language models in inducing difficulty rubrics, then compared model judgments with and without rubric guidance. None of the induced rubrics treated surface features such as stem length, option count, or images as major sources of difficulty. Patterns including cue convergence, misleading cues, and option competition emerged, but rubric guidance did not consistently improve judgment performance, suggesting its main value lies in generating testable hypotheses.
In briefCord: a lightweight method that learns when to deviate from confidence-based decoding in masked diffusion language models.
SummaryThis paper addresses how the order in which masked diffusion language models reveal tokens affects generation quality. It proposes Cord, a lightweight policy that uses confidence decoding as a reference and learns geometric residuals over generation direction, proximity to revealed context, and two-sided infilling. Across three models and five reasoning, code, and constraint-satisfaction benchmarks, average task success rises from 32.0% to 48.2%. Cord adds 0.53M trainable parameters and 5.39% inference latency overhead.
In briefMore Is Better? Exploring Latent Correspondence in Federated Hypergraph Learning
SummaryThis paper studies federated hypergraph learning, where hyperedges across clients have latent correspondences that can be ambiguous, making recovery of these correspondences challenging. The authors propose FedASTRA, which learns correspondence-invariant representations through adaptive heat diffusion and masked latent prediction while communicating only model parameters. Experiments on eight real-world datasets show it outperforms identifier-free federated graph and hypergraph methods, reaching 50.07 accuracy and 29.73 Macro-F1 on ogbn-papers100M with 59.60 MiB communicated per round.
In briefDiagnosing how policy updates shift the usefulness of agent skills before the next rollout
SummaryThe paper argues that policy updates can change the behavioral effect and marginal utility of an unchanged skill, so attribution based on earlier trajectories may no longer reflect its contribution. The authors compare token preferences between skill and control conditions at fixed pre-update contexts and combine these shifts with advantage estimates to build readouts that flag skills with declining utility. These readouts reuse existing training rollouts and need only forward passes for single-step scoring. Integrating them into skill evolution is reported to improve downstream performance while reducing policy and editor token consumption.
In briefSemantic mismapping in generalized category discovery and how to evaluate and reduce it
SummaryThis paper studies generalized category discovery, which asks models to recognize known categories and discover new ones within the same unlabeled set. The authors find that global Hungarian matching can reassign anchored old-category indices, inflating new-category accuracy by up to 25.63 points, a effect they call semantic mismapping. They propose Semantics-Preserving Evaluation to keep old-category semantics fixed, and Mismapping-Mitigating Optimization with a region-balance loss and a novel-diversity loss to reduce the errors behind this effect.
In briefCoToan: Executable Rules and Representation Contract for Reproducible Game AI
SummaryThis paper presents CoToan, an executable specification and representation contract for Vietnamese Mathematical Chess that links the Game Description Language with OpenSpiel. It represents actions by their state transitions rather than arithmetic witnesses, and treats time and victory configuration as part of the task. Cross-representation tests show agreement in legal actions, successor states, and terminal outcomes, while controlled diagnostics show that uniform witness-level sampling alters endpoint-action probabilities and omitting time or victory settings can merge states with different terminal behavior.
In briefPicturePerfect: Diffusion-based Same-identity Facial Replacement
SummaryThis paper explores using diffusion models to replace a face in a photo with another image of the same person. The method builds on a pre-trained text-to-image model, taking a portrait and a reference image as input while keeping the background intact. The authors also construct a large dataset of same-person image pairs with varying expressions and poses. Experiments reportedly yield more realistic, identity-consistent results than existing face reenactment models.
Manik Dhingra, Yashesh Savani, Mahesh Reddy et al.
In briefMulClip: a table-free activation for microcontroller reservoir computing with monotonic and folding regimes
SummaryThis paper proposes MulClip, a lightweight activation function for reservoir computing on microcontrollers, built only from comparison, absolute value, multiplication, and subtraction. Its clipping threshold sits where the response returns to zero after passing its extremum, giving one function both a monotonic and a nonmonotonic folding regime. On a Cortex-M4, its state-update time was 0.34 times that of tanh and shorter than any of 11 tanh approximations. On NARMA-20 and 8-bit parity, third-order MulClip outperformed tanh and the monotonic-clip variant, and on-device results nearly matched PC results.
Guren Matsumura, Masato Takemoto, and Shinzo Koyama
In briefRetroChessBench: A benchmark for reconstructing and proving chess histories
SummaryThis paper introduces RetroChessBench, a 12-task benchmark for retrograde chess, covering exact-length constructions, unique endings, and exhaustive answer sets, with references, rubrics, and a terminal-reward environment. In two trials per task and condition, GPT 6 Astra High failed to produce the required final answer in 22 of 24 attempts without tools and 18 of 24 with code execution. The authors attribute these failures to faulty deductions, such as rejecting legal histories and treating temporary obstructions as permanent. They question how well frontier models' mathematical reasoning generalizes beyond familiar solution patterns.
In briefDoes the Citation Matter? Google AI Overviews' Medical Information Accuracy
SummaryThis paper evaluates Google's AI Overviews on 3,611 health questions, comparing their cited sources with the top five traditional search results. Only 44.4% of traditional top results were cited, with commercial health publishers overrepresented. Against physician-written rubrics, AI Overviews earned 38.2% of available points, and their accuracy was found to be unrelated to whether their statements were grounded in cited sources.
WhyThe authors compare Google AI Overviews with traditional search on 3,611 health questions, reporting source differences, grounding gaps, and physician-rubric accuracy with direct implications for user
In briefTelling Is Not Seeing: Do Text Personas Make AI Answer Like Real People?
SummaryThis paper examines whether text personas such as "You have low vision" can make vision-language models answer like people with low vision, color blindness, or hearing loss. The authors compare persona-based answers with measured human results and with simple rules that use only the facts stated in each persona. The personas never outperform these rules and erase real differences between groups and individuals. In all 14 models tested, describing blur left recognition unchanged, while physically blurring the image lowered it, as it does for people.
WhyThe authors compare text personas in 14 VLMs with measured low-vision, color-blind and hearing-loss groups and simple rules, finding personas fail to reproduce real perceptual differences and that mod
In briefCan Image Agents Complete Visual Commissions? A Benchmark and Modular Framework
SummaryThis paper introduces Visual Commission Bench (VCB), a benchmark of 600 tasks across 12 categories for testing whether image agents can complete end-to-end visual commissions. It also presents V-Protocol, a reference agent that dynamically composes task, cognitive, and expression modules that exchange structured intermediate artifacts. Under the same controller and image generator, V-Protocol reaches 69.0% strict task success on VCB versus 42.0% for SCOPE, and raises strict accuracy on Mind-Bench from 39.2% to 45.4%.
In briefMRC-MMAgent: Multi-Role Collaborative Reinforcement Learning for Multimodal Tool-Using Agents
SummaryThis paper proposes MRC-MMAgent, a framework that trains multimodal tool-using agents to play multiple collaborative roles: solving, inspecting, and continuing. The agent gathers evidence with tools, evaluates its previous trajectory, and resumes problem solving when its current answer is insufficiently supported. Training combines task success, interaction validity, and successful correction with full-trajectory policy optimization across all roles. Experiments using only the original baseline training set show consistent gains over the base models and leading or highly competitive results across benchmarks.
In briefClass-aware Latent Attention Replay (CLAR): Class-incremental hazard diagnosis for AI in manufacturing industrial internet
SummaryThis paper addresses hazard diagnosis for AI systems in the Manufacturing Industrial Internet, where new hazard types emerge with few samples each, hazards can co-occur, and models run on memory-constrained nodes. The proposed CLAR method combines a dual buffer storing mostly compact latent representations with a few raw samples, an Attention Leverage Score that guides class-balanced reservoir sampling, and an exponential-moving-average teacher that slows latent-space drift. On hazard diagnosis data from an MII testbed, CLAR achieves the highest mean average weighted F1 score among the evaluated replay methods at the same buffer size.
In briefMemoryFusion: A Cross-Temporal Memory Method for Fusing Multimodal Video
SummaryThis paper proposes MemoryFusion, a two-stage framework that fuses multimodal videos by combining cross-temporal memory learning with progressive temporal refinement. The first stage uses short-term and long-term memory banks along with a temporal smoothing module, while the second stage applies a lightweight 3D residual module to correct remaining temporal inconsistencies and enhance structural details. The authors report better fusion quality and temporal consistency than existing frame-based and video-based methods, and the source code is publicly available.
In briefMinimizing satisficing regret in bandits with knapsacks
SummaryThis paper studies satisficing regret minimization in bandits with knapsacks, where a learner picks randomized policies under finite resource budgets to exceed a satisficing reward threshold as often as possible. The authors propose SELECTwK, a general template that uses a BwK learning oracle with sub-linear expected regret to search for and test candidate satisficing policies. The analysis, supported by LP duality, shows constant expected satisficing regret when a satisficing policy exists, with sharper bounds under a stronger regret condition.
In briefBEVLight: Generalizable Traffic Signal Control Using Drone Bird's-Eye Views
SummaryBEVLight is a framework that maps drone bird's-eye-view images directly to traffic signal actions, without relying on fixed roadside cameras or pre-estimated traffic states. It links visual features to individual lanes and combines them along the intersection layout into phase representations. The model was trained on over 5,000 human control decisions collected through a web-based virtual teleoperation platform. Closed-loop experiments report zero-shot generalization across traffic demands, signal plans, and intersections, with 7–9% lower average waiting time than visual baselines.
In briefTaxoMFM: A Taxonomy-Aware Gut Microbiome Foundation Model
SummaryExisting gut microbiome foundation models typically use flat taxon representations that do not exploit known taxonomic hierarchies. This paper introduces TaxoMFM, which integrates taxonomic structure into self-supervised learning, and builds TaxoGut-HPP, a dataset of 100,806 gut microbiome profiles from 231 cohorts. In cross-cohort zero-shot transfer, TaxoMFM achieves AUROCs of 0.838 for colorectal cancer, 0.736 for type 2 diabetes, and 0.722 for inflammatory bowel disease.
WhyThe authors integrate taxonomic hierarchy into a gut microbiome foundation model and report cross-cohort zero-shot transfer results on 231 cohorts.
In briefUnlimited OCR 2: A Document Parser for Ultra-Long Documents
SummaryThis paper presents Unlimited OCR 2, a document parsing model that attends only to the content it is currently transcribing, without revisiting preceding pages. It replaces decoder attention with Sparse Reference Sliding Window Attention (SR-SWA), which caps active reference tokens at a fixed size, and combines this with an unlimited-horizon decoding paradigm. At the hundred-page scale, the authors report per-step speedups of 7.0x over Full Attention and 5.6x over R-SWA, and state-of-the-art end-to-end results on MPDocBench and MarathonDocBench.
In briefCAD-WAM: Anticipating Geometry for Generating CAD Models from Text
SummaryThis paper introduces CAD-WAM, a method for generating editable CAD models from text descriptions. It uses a large language model to produce the next modeling operation while also predicting the resulting intermediate geometry. On the DeepCAD test set, it outperforms task-specific baselines and general-purpose LLMs in geometric accuracy and output validity. Ablations show that next-state supervision reduces mean Chamfer distance by 13.2%.
In briefLearning to Bound: Neural Upper Bounds for MPE Inference
SummaryThe paper introduces Learning to Bound (L2B), a framework that trains a neural predictor to produce upper bounds for Most Probable Explanation (MPE) inference in probabilistic graphical models. Admissibility is obtained through statistical calibration, and monotonic tightening as evidence accumulates is enforced by architectural constraints. According to the authors, L2B yields tighter monotonic bounds at orders-of-magnitude lower cost across PGM benchmarks and improves branch-and-bound on high-treewidth instances.
In briefProving and Improving the Platonic Representation Hypothesis with JEPAs
SummaryThe paper turns the Platonic Representation Hypothesis, which says models trained on different modalities converge to the same world representation, into a theorem. It proves that separately trained Joint-Embedding Predictive Architectures learn the same representation up to one rotation when positive pairs are good views of one scene. It also reframes next-token prediction as a JEPA using bad views from one document, and reports experiments showing text JEPAs align better with image encoders while needing far fewer training samples.
WhyThe authors prove a theorem that JEPAs with good views converge to one representation, reporting a text JEPA beating language models up to 72B parameters.
In briefYoutu-Parsing-Omni: One Encoder and One Schema for Parsing All Modalities
SummaryThis paper presents Youtu-Parsing-Omni, a parser that converts documents, images, audio, and video into structured records tied to space and time. It uses one shared encoder for all modalities and one unified output schema covering ten parsing tasks, trained with a self-evolving method called OSAD. A single 5B-parameter model scores 96.96 Overall on OmniDocBench v1.6 and 75.08 Avg. on OmniParsingBench, and its code and weights are public.
In briefAutoregressive Transformer for Multi-Solution Inverse Design of Pixelated Patch Antennas
SummaryThis paper presents an autoregressive transformer for inverse design of 10 x 10 pixelated patch antennas, generating multiple feasible layouts for a given target response through sequential discrete generation. The model is fine-tuned for dual-band design using step-shaped target responses, allowing bands to be specified without manually constructing detailed S11 curves. Compared with binary particle swarm optimization, it produces valid and diverse candidates more efficiently. Fabricated and measured examples show that distinct pixelated layouts can realize the desired responses.
Young-Mok Yun, Seung-Won Shin, Jaeyong Chung et al.
In briefNo Free Lunch in RLVR: Each pass@1 Gain Costs pass@k
SummaryThe paper studies reinforcement learning with verifiable rewards (RLVR), which raises a language model's first-attempt success (pass@1) while the base model outperforms it when many samples are drawn (pass@k). The authors model RLVR as an exponential tilt of a base policy and prove that, under exclusive specialization, any update that raises pass@1 lowers the success of some prompt, with the number of pass@k crossovers bounded by a sign-variation measure. They argue RL moves along the reasoning frontier rather than widening it, and offer a diagnostic that measures this price from evaluation logs.
WhyThe authors model RLVR as an exponential tilt and prove an exact price for pass@1 gains, with a tight bound on pass@k crossings; the abstract reports theory only.
In briefEnabling any-order image generation in pretrained autoregressive models
SummaryAutoregressive image generators are usually trained in a fixed raster-scan order, which limits their use for inpainting, outpainting, and editing at inference time. The paper proposes adapting pretrained raster-scan models into any-order generators using future-aware positional embeddings and an ordering curriculum. On ImageNet inpainting, the method outperforms MaskGIT and is competitive with MAR-L, adding 13% to 33% compute beyond pretraining.
In briefStrongly Adaptive Online Learning Achieves Uniform Discounted Regret
SummaryThis paper studies discounted regret, which weights past losses with exponentially decaying factors so that recent losses matter more in non-stationary online learning. The authors show that strongly adaptive regret bounds directly yield uniform discounted regret guarantees for convex, strongly convex, and exp-concave losses. For convex functions the bound is O(√(log T/(1-λ))) for all discount factors, and for strongly convex and exp-concave functions it is O(log T(1+log(1/(1-λ)))).]
In briefPriorMaskMap: Robust Online Map Building from Imperfect Prior Maps
SummaryThis paper presents PriorMaskMap, a framework for building online vectorized maps for autonomous vehicles that can use prior maps even when they are outdated or partly wrong. It introduces a confidence estimator that marks map regions as reliable, erroneous, or blank by comparing priors with current visual observations. The confidence mask then guides how priors are used to initialize queries and fuse features for the map decoder. Experiments on nuScenes show higher mapping precision than existing methods and greater robustness when priors contain apparent errors.
In briefWhat the "Now" of a Computer-Use Agent Means: Stale Screenshots, Page Loads, and Checks Before Clicking
SummaryThe paper studies how computer-use agents that plan from screenshots can click unintended elements when a page changes before the click. Measuring 31 public web pages, it finds most of the risk arises during page loading rather than during planning. It compares pre-click checks on accuracy and latency and releases a benchmark generator, a break-even formula, and three rules for developers.
In briefWorldFlow: Joint Video and 3D Motion Prediction via Large-Scale Pretraining
SummaryThis paper introduces WorldFlow, a generative world model that jointly predicts future video and query-based 3D trajectories using a unified Mixture-of-Transformers backbone. It is trained on about 1.44 million object-level samples from simulated and real-world domains and evaluated on the new WorldFlowBench with 1,678 samples. WorldFlow achieves the lowest average and final displacement errors among the evaluated learning-based methods across all three evaluation sets.
In briefEfficient covariance-aware vicinal latent diffusion for generating images from continuous labels.
SummaryThe paper addresses continuous conditional image generation, where data is sparse under imbalanced label distributions, causing artifacts and inconsistent conditions. It proposes CAViLD, a latent diffusion framework that builds a condition-aware latent space with density-adaptive vicinal learning, semantic regularization, and NIQE-guided negative augmentation. It uses an EMA teacher for single-branch conditional inference and covariance-aware progressive distillation to sample in as few as four steps. Experiments on RC-49, UTKFace, and Steering Angle show a favorable balance among fidelity, condition consistency, and efficiency.
In briefMemSteer: Turning user histories into training signals to optimize AI agent memory
SummaryThe paper introduces MemSteer, which converts unlabeled user histories into question-answer pairs with supporting evidence for optimizing agent memory designs. It builds an evidence-grounded answer plan to guide question generation and checks the annotations through auditing and blind answer recovery. The authors also present MemSteer-Opt, which revises memory storage, retrieval, and context construction without updating model parameters. Experiments on ATM-Bench and RHELM show results competitive with optimization guided by official QA.
In briefLLM-Assisted Optimization for Simulation-Based Engineering Redesign
SummaryThe paper develops an approach that combines large language models with numerical optimization solvers to redesign existing engineering solutions when context or requirements change. It compares three ways of using LLMs (warm start, variable fixing, and ReAct) with four black-box solvers on an electric vehicle battery pack cost minimization case. The integrated approach reportedly matches the feasible design discovery of solver-only baselines using 80% fewer simulator evaluations, while tripling the fraction of runs finding near-optimal solutions.
Chan-Eui Song, Jeremy J. Michalek, Christopher McComb
In briefDeCoRL: Decomposed Consistency Reinforcement Learning for Multi-Image Composition
SummaryThis paper proposes DeCoRL, a reinforcement learning framework that improves consistency in multi-image composition for image editing models. It breaks overall consistency into object-level consistency by matching instruction-relevant objects across reference images and scoring each object with a VLM-based reward model. The per-object scores are combined into a multi-object reward that guides reinforcement learning post-training. The authors also introduce DeCo-Bench, covering compositions of 2 to 6 reference images, and report consistent gains on it and the public MICo-Bench.
In briefTraining phone app navigation agents with online reinforcement learning on live iOS apps
SummaryThis paper proposes an online reinforcement learning framework that trains user interface navigation agents directly on live iOS apps. The agent is split into a planning component and a UI grounding model, and an LLM serves as judge to give step-level and terminal rewards, optimized with PPO and a new GiGPPO value estimation method. On benchmarks including UINavBench, iOSWorld, and AndroidWorld, the authors report that these agents outperform supervised fine-tuning and terminal-only RL baselines.
WhyThe authors report online RL on live iOS apps with a new neighbor-conditioned critic (GiGPPO), outperforming SFT and terminal-only RL on UINavBench, iOSWorld, and AndroidWorld.
Peitong Duan, Khashayar Gatmiry, Jeffrey Nichols et al.
In briefEfficient Adaptive Training for GUI Agents: Where to Learn and How Far to Roll Out
SummaryThis paper presents a two-stage adaptive training framework for multi-step GUI agents. It builds reusable intermediate checkpoints from successful expert trajectories, selects key decision points using action disagreement, and applies one-step reinforcement learning there before progressively expanding training scope by task progress. On MobileGym and CUA-Gym-Hub, the method reduces environment interaction by 56-61% relative to Vanilla GRPO while improving task performance.
In briefSocialArena: A Benchmark for Measuring Social Intelligence in Large Language Models
SummaryThis paper introduces SocialArena, a benchmark for systematically defining and evaluating how socially intelligent large language models are. It defines social intelligence across three dimensions: role-playing ability, mental state reasoning, and behavioral reasoning, and uses a role-playing dialogue setup to assess them together. The authors also add a U-TRIMF framework for covering diverse social scenarios and a human-annotated arena-style protocol for more reliable evaluation.
In briefIMRouter: Imbalance-aware routing of temporal patterns for LLM-based time series forecasting
SummaryThis paper proposes IMRouter, a framework that improves time series forecasting with large language models. It notes that existing methods overlook temporal pattern imbalance, where some patterns occur far less often and get higher forecasting errors. IMRouter uses Pattern-Adaptive Temporal Representation Learning and Routing-Guided Semantic Alignment to reduce this bias. Experiments on benchmark datasets show improved forecasting performance, especially on less frequent temporal patterns.
In briefVisually grounding multimodal reasoning with evidence-ablated negative samples
SummaryThis paper proposes EVAN, a multimodal reinforcement learning framework for training large multimodal models. It masks the most discriminative image patches to create evidence-ablated images, prompting the current policy to generate plausible but visually ungrounded negative responses. These negatives are used as bounded auxiliary signals during optimization to penalize shortcut learning. The authors report that EVAN mitigates modality imbalance and outperforms strong RLVR baselines on challenging multimodal reasoning benchmarks.
In briefGenSight: a unified world model for video generation and 4D perception
SummaryThis paper presents GenSight, an embodied world model that combines future video generation and dense perception within one pretrained video diffusion transformer. It represents video, depth, embodiment segmentation, and first-frame-relative 2D flow in a unified spatiotemporal form, predicts all of them with lightweight task modulation, and optimizes them jointly in a single velocity prediction space. It supports multi-step future video generation with aligned dense predictions, and single-step dense perception from observed video. Trained on embodied manipulation data across multiple robot embodiments and human-hand interactions, it shows strong results in video generation and dense 4D perception.
In briefBeyond LLM Wrappers: Context Engineering and a Three-Engine Architecture for Reliable Financial AI
SummaryThis paper proposes a financial AI architecture combining a financial knowledge graph, a deterministic optimization engine, and a multi-agent LLM layer, coordinated by a process called context engineering across ERP systems and tools. Based on observations from 15 deployments over about 11 months, the authors report that single-agent pipelines had hallucination rates of roughly 7-9%, while graph-backed multi-agent validation reduced observed rates below 1%, though they note these findings are observational rather than causal.
In briefXLaw-JP-27B: A small open AI model built for Japanese law
SummaryThis paper presents XLaw-JP-27B, an open-weight model specialized for Japanese law that runs on a single 48 GB GPU under the Apache 2.0 licence. It adapts Qwen3.8-27B using hybrid retrieval over statutes and case summaries, fine-tuning on its own verified bar-exam answers, and four-sample voting. Evaluated on eight Japanese legal tasks, it ranks first on six and is the only system that passes the bar exam in all three years tested.
In briefWhere position information matters: profile-based layer-wise allocation of RoPE in Transformers
SummaryThe paper addresses the uniform use of Rotary Position Embedding (RoPE) across Transformer layers by proposing layer-wise allocation of high and low coverage under a fixed budget, guided by early coverage-benefit profiles. The method extends to physical rotary-dimension allocation in Multi-head Latent Attention (MLA) and uses normalized-depth mapping to transfer profiles from small to larger models. From-scratch experiments report validation-loss improvements over budget-matched baselines.
In briefFast rigid alignment of 3-D density maps using sliced Wasserstein distance
SummaryThis paper presents a fast method for rigidly aligning voxelized three-dimensional density functions over SO(3) using sliced optimal transport. Each density map is converted into a tensor of one-dimensional quantile functions taken from central slices over the unit sphere, so that Euclidean distances between the transformed volumes approximate the sliced 2-Wasserstein distance. This allows the approach to work with existing fast voxel-based Euclidean alignment algorithms. The authors provide theoretical analysis and numerical experiments indicating that the sliced 2-Wasserstein distance is a robust metric for comparing 3-D shapes and computing their rigid alignment.
In briefEvent2Reward: Learning Robot Rewards from Progress Relations
SummaryThis paper proposes Event2Reward, a method that learns task progress from robot demonstrations to provide dense reward shaping for reinforcement learning. It combines absolute progress supervision, anchored by normalized temporal labels, with relative progress relations such as stalls, regressions, and recoveries within trajectories. The authors built an event-annotated dataset from about 100K real-robot execution records. Across 40 LIBERO tasks, Event2Reward improves average success by 18.16 percentage points over a sparse-reward baseline, with gains also shown on real robots.
In briefCOGS: A Diagnostic Benchmark for Testing How 3D Models Use Gaussian Splat Attributes
SummaryThis paper introduces COGS, a large dataset of 15,675 artifact-filtered, appearance-rich objects across 40 categories, built for 3D Gaussian Splatting perception. It establishes COGS-Bench for object classification and attribute-usage diagnosis. The authors find that Gaussian attributes are not uniformly beneficial, with some subsets outperforming the full set and attribute effects varying across architectures. They also propose HALO, a reference model that preserves attribute-specific evidence before hierarchical aggregation.
In briefHow Pair-Dependent Judge Errors Change the Minimax Rate of Semantic Entropy
SummarySemantic entropy measures the spread of meanings among a language model's sampled answers, estimated from a judge that decides whether answer pairs are equivalent. The paper models judges whose errors depend on the specific answer pair, and shows that for three meanings the estimation rate is governed by tree statistics on four answers, while on an open set of channels the consistency threshold is governed by trees on five. For each channel in an explicit open class, it gives the order of the minimax squared risk and the polynomial weak windows where this holds.
In briefAnalyzing condensation on patterned surfaces with deep learning
SummaryThe paper presents TileSAM3, a universal droplet segmentation tool built on the Segment Anything Model 3 (SAM3) that uses hierarchical tiling to detect small droplets on inhomogeneous surfaces. The authors created two datasets, one with homogeneous surfaces and one with patterned, defect-laden surfaces under poor lighting. On the lower-quality dataset, TileSAM3 reached an Average Precision of 0.57 compared with 0.088 for thresholding. Combined with KD-tree tracking, it identified localized nucleation and coalescence on a stiffness-patterned surface.
In briefZero-Complement Extrapolation for Logistic Regression with Very Few Minority Samples
SummaryThe paper addresses binary logistic regression when minority-class observations are fewer than features, a setting where maximum likelihood, Firth regression, and SMOTE struggle. It proposes a Self-Calibrated Zero-Complement estimator that extrapolates synthetic minority samples outward from the majority class's high-density fringe, with theoretical justification based on Stein's Lemma. Experiments on Wisconsin Diagnostic Breast Cancer and Banknote Authentication data, plus Monte Carlo simulations, report better F1-scores and estimation stability than Firth's regression and SMOTE when only 3 or 5 minority samples are available.
In briefGoal-Aware Dynamic Multi-Agent Collaboration for Long-Horizon Tasks (GAD)
SummaryThis paper studies how LLM-based agents divide duties such as planning and verification during long, complex tasks. It finds that verification is most linked to task success but often happens late, and that duty-relevant information shrinks as records accumulate in a shared context. The authors propose GAD, which uses a growing execution graph to organize subtasks and dependencies, keeps each agent's context tied to the original goal and acceptance conditions, and uses verification and recovery to check results and guide repairs.
In briefMulti-level constrained learning for enhancing underwater images
SummaryThis paper proposes MCL-UIE, a framework for correcting color and recovering detail in underwater images degraded by light attenuation, scattering, and noise. It combines a monotonic chromatic transform for global color mapping, an uncertainty-gated wavelet reconstruction for high-frequency detail, and reference-anchored degradation replay used only during training. The authors report ranking first in thirteen of sixteen full-reference comparisons and competitive results on no-reference benchmarks.
In briefReasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models
SummaryThis paper studies how the internal representations of language models change during reinforcement learning with Group Relative Policy Optimization (GRPO), using Tic-Tac-Toe as a finite, fully analyzable logical domain. The trained models achieve high accuracy and robustness across prompt variations. Sparse autoencoder analysis, however, shows that they mainly rely more on information explicit in the prompt, such as turn, game progression and board occupancy, while board symmetries and strategic forks remain weakly represented.
In briefEmbodiExp: Reusing Past Experience Across Tasks for Robots and Embodied AI Agents
SummaryThis paper presents EmbodiExp, a framework that builds transferable experiences and combines them online while embodied agents carry out tasks. It packages shared subgoals as outcome-grounded experiences with applicability conditions, an adaptable policy, and a verifiable outcome. Experiments on several embodied benchmarks show the framework is effective.
In briefUsing edited end images as partial goals for image-to-video generation
SummaryThis paper studies temporal artifacts that appear when edited endpoint images are used as targets for image-to-video generation. The authors propose Partial, which constrains only an edit-relevant region estimated from image-editor attention and leaves the rest to the video model. On TC-Bench-I2V across three backbones, Partial reduces temporal artifacts compared with full-frame conditioning, with no significant difference in transition completion ratio.
In briefPRISM: Personalized item embeddings for user histories in recommendation
SummaryThis paper argues that sequential recommenders use item embeddings shared across users, so the same item gets the same representation regardless of the user. The authors propose PRISM, which reuses upstream ranking computation to produce a personalized embedding for each user-item pair, compressed into one embedding by an autoencoder. Offline evaluation on a billion-scale recommendation surface reports an average normalized entropy change of -0.15%, compared with -0.07% for VISTA. The authors also report a power-law gain with history length.
In briefACAM-J-MARS: Multi-Atlas Riemannian Spectral Representations for Decoding Advanced Meditation States from fMRI
SummaryThe paper proposes ACAM-J-MARS, which computes functional connectivity using multiple brain atlases, maps symmetric positive definite connectivity matrices onto a Riemannian manifold and a log-Euclidean tangent space, and adds low-frequency BOLD spectral energy features to decode jhana-type advanced meditation states from fMRI. Evaluated on 204 segments from 22 practitioners, it improves performance across several backbone models. It reaches a balanced accuracy of 0.757 for early versus late classification and 0.545 for three-class classification, a 10.8-point absolute gain over the strongest baseline.
In briefChecking Kanji Reading Accuracy in Japanese Financial Speech Synthesis
SummaryThis paper addresses frequent misreadings of multi-reading kanji words in Japanese financial speech synthesis. It proposes a reading-level check that uses a CTC hiragana recognizer to score the expected reading against confusable alternatives. On 540 utterances from three voice-cloned voices, the method detected 91.7% of misreadings with a 5.0% false-alarm rate, outperforming a transcript-based quality gate.
In briefEgoContact: A dataset for testing whether robots can predict where to touch objects from first-person video
SummaryEgoContact is a large-scale egocentric video dataset for evaluating how well learned manipulation priors predict future contact points. It contains 22.5K manually verified interaction events, each with a pre-contact observation, semantic intent, and pixel-precise contact locations. The authors find that current methods often struggle to predict precise contacts and sometimes fail to localize the target object in diverse real-world scenes. Using EgoContact as the memory bank of an existing affordance-transfer policy raises grasp success from 14.2% to 31.2% in simulation and from 16% to 58% on a real robot.
Adam Dery{\l}o, Alexey Gavryushin, Daniel Cremers et al.
In briefA new way to test how well evaluation metrics judge multi-output AI predictions
SummaryMany AI tasks have several valid outputs for one input, but training data gives only one correct answer, which makes evaluation hard. The authors propose approximate task-specific invariants, such as object boundaries in colorized images or named entities in translations, and use them to diagnose existing metrics for vision and language tasks. They enforce these invariants in model outputs with an augmented loss function. Distributional metrics improved substantially under this approach, while traditional metrics changed only marginally.
In briefJoint Alignment and Shape Completion of Partial 3D Face Scans
SummaryThe paper proposes a single network that predicts both the rigid pose and a shape-complete point cloud of a partial 3D face scan in one forward pass. It builds on GeoTransformer and maps scans to coefficients of a 3D morphable face model learned from 8,000 identities. The authors report lower per-vertex error than six baselines on held-out identities, and the highest accuracy and speed on real-scan datasets NoW and NPHM.
Eli A. Underwood, Nicolas Rondan, Angel V. Peterchev et al.
In briefVEvo-Harness: Causally Guarded Self-Evolving Harness for Agentic Image Generation
SummaryThe paper presents VEvo-Harness, a continual self-evolving harness for agentic visual generation that goes beyond simple generate-verify-rewrite loops. It converts sparse VLM judgments into root-cause attributions and admits skill updates only after counterfactual dual-arm image generation replay confirms they are causally valid. According to the paper, it surpasses state-of-the-art agentic methods and rivals proprietary generators such as GPT-Image and Nano-Pro.
In briefLearning equal-length erasing relational pattern languages from positive data over binary alphabets
SummaryThe paper studies erasing relational pattern languages under the equal-length relation, where every occurrence of a variable must be replaced by strings of the same length. It shows that this class is not learnable in the limit from positive data over a binary alphabet, resolving an open question. It further shows that patterns with at most two maximal variable substrings are learnable in the limit from positive data for every alphabet of size at least two.
WhyThe authors resolve an open question on binary alphabets, proving non-learnability for three-substring patterns and learnability for patterns with at most two substrings.
In briefSTRATA: Reinforcement-Learning Active Sampling for Efficient 3D Spatial Transcriptomics Mapping
SummarySpatial transcriptomics captures both tissue morphology and gene expression, but 3D acquisition is costly. The paper proposes STRATA, which uses a reinforcement learning agent to select informative spots under a sampling budget, using a 3D graph that combines H&E morphology with single-cell priors. An adaptive predictor then fuses retrieval, regression, and cross-layer imputation to reconstruct dense 3D transcriptomic maps from sparse measurements. The authors report improved sampling efficiency and accuracy, especially at low budgets.
In briefMotion vs. Shape Attribution for 4D Point-Cloud Video Models
SummaryThis paper introduces Kinematic Causal Attribution (KCA), a black-box method that explains 4D point-cloud video classifiers by separating reliance on motion from reliance on shape. It uses freeze and shape-free proxy operations to build a motion-by-geometry cube for each tracked region and decomposes contributions with Shapley values. Tested on PSTNet, P4Transformer and Mamba4D trained on MSR-Action3D, the motion share of an arm's importance follows the task's defining cue, and explanation faithfulness is maintained.
WhyThe authors propose black-box motion-versus-geometry attribution, validated via cue audits on three backbones with reported deletion AUC results.
In briefClean Sheets and Checkpoints: Evidence Complexity for Certifying Sequential Agent Risk
SummaryThe paper studies how many zero-success audit results are needed to certify that an AI agent's chance of completing a dangerous multi-step task stays below a given level. The authors model the task as a chain of milestones divided into segments and derive an asymptotic minimum on the expected number of zero responses. They also show that an evaluator can attain this minimum constant and describe how to allocate responses among milestones.
Rohith Yanambaka Venkata, Don McBride, Tiago Sequeira et al.
In briefUncertainty-guided cross-modal alignment for diffusion-based IoT intrusion detection
SummaryThis paper addresses generating minority attack samples for imbalanced Internet-of-Things intrusion detection. It proposes a reliability-guided alignment rule that uses normalized VAE posterior variance to weight cross-modal influence across tabular flow features, DeepInsight images, and packet sequences. A class-conditional diffusion model generates paired attack latent representations for the modality decoders. On TON-IoT, posterior-variance weighting improved both generative fidelity and downstream detection over uniform weighting.
Mahdi Arab Loodaricheh, Anita Raja, Mohammad Hossein Manshaei
