In briefVirtual Head Attention: a new attention design that cuts KV cache and compute in long-context LLMs
SummaryThe paper addresses the trade-off between KV cache size, training and prefill compute, and model quality in long-context language model attention. It proposes Virtual Head Attention (VHA), which uses two lightweight linear operations, Q Premix and Linear Postmix, to recover head diversity and can be folded at inference into a standard GQA-2 architecture. Reported experiments across four scales show VHA outperforming the GQA baseline with smaller KV caches and 7.9% lower training and prefill FLOPs at 1.7B.
In briefLearning Distributions from Samples and Their Quality Labels
SummaryThis paper studies estimating a parameter of a continuous distribution when each sample carries a binary quality label governed by the same unknown parameter. Imposing hard feasibility constraints on deterministic labels speeds up estimation from the classical n^(-1/2) rate to n^(-1). The authors also show that pointwise qualities and pairwise preferences are incomparable, and that the assumptions hold for Gaussian, Laplace, and exponential families.
In briefAgent-Reflex: Training language agents to reflect and recover from errors
SummaryThis paper proposes Agent-Reflex, a training framework that teaches language agents to proactively reflect on noisy tool feedback and user corrections. It combines skill discovery from success-failure pairs, reflective supervised fine-tuning, and on-policy reflective policy optimization. Built on Qwen3-8B, the method scores 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, outperforming strong baselines. The authors report improved robustness to noise and stable training.
In briefPhonetic-Expert Mixture Low-Rank Adaptation for Accented English Speech Recognition
SummaryThis paper proposes MoPE-LoRA, a parameter-efficient adaptation framework for accented English speech recognition. Instead of allocating experts per accent, it organizes six low-rank experts by manner-of-articulation phonetic categories and routes frames using phoneme supervision combined with learned acoustic gating. Experiments on L2-ARCTIC show gains over single LoRA and full fine-tuning, and in zero-shot evaluation with one accent held out, it yields a 12.3% relative improvement over single LoRA.
In briefFETALFusion: Resolution-aware state-space model for fetal head segmentation across ultrasound types
SummaryThis paper presents FETALFusion, a CNN-Mamba model that segments the fetal head in 2D ultrasound across transabdominal and transperineal scans, to support automated head circumference measurement. It adds a landmark shape prior and domain-specific channel gates. On 471 held-out test images it reaches a Dice score of 0.963 and an HD95 of 2.10 mm. The authors report that a ResNet-34 UNet++ and an ImageNet-initialised ResNet-34 U-Net perform better, and that removing the state-space branch does not measurably change accuracy.
In briefHaloBlock: A Faster, More Accurate Decoding Method for Diffusion Language Models
SummaryDiffusion large language models decode text in blocks, which makes the rightmost tokens of each block hard to generate and causes discontinuity at block boundaries. HaloBlock measures each block's decoding difficulty using bidirectional cumulative entropy and adaptively allocates halo regions on both sides of the block to smooth context across boundaries. Evaluated on LLaDA-8B-Instruct, Dream-7B-Instruct and LLaDA-1.5 on math reasoning and code generation benchmarks, it reports consistent accuracy gains and higher inference efficiency.
In briefAI-Enhanced Hyperspectral Imaging for Detecting Low-Grade Glioma During Surgery: A Clinical Study
SummaryThis clinical study examines whether hyperspectral imaging (HSI) can help detect low-grade gliomas during surgery, comparing it with standard RGB imaging. Tumor regions showed lower reflectance, which may relate to differences in hemoglobin concentration and oxygenation. A 3D CNN combining spectral and spatial features achieved the highest median ROC-AUC of 74.5%, versus 67.9% for spectral-only XGBoost and 64.0% for RGB.
In briefSea-Net: A Small, Interpretable Neural Network for Time Series Classification
SummaryThis paper introduces SEA-Net, a lightweight neural network for classifying time series that also indicates which time points or intervals drove each prediction. It builds on the MILLET approach but uses small multi-scale depthwise-separable convolutional encoders and improved multiple instance learning pooling heads. The bottleneck-encoder variant with Top-k pooling reports 0.931 accuracy and 0.765 NDCG@n with only 41K parameters, suiting resource-constrained edge devices.
In briefMERIT: Reasoning over table images using visual evidence
SummaryThis paper proposes MERIT, a three-stage framework that teaches models to reason directly over table images rather than serialized text, which can lose cell content and structure. It adapts the visual encoder through reconstruction, localization, and color-conditioned supervision, then fine-tunes on verified reasoning trajectories and applies reinforcement learning with separate rewards for evidence, calculation, answers, and format. The method is evaluated on table question answering and fact verification, with tests on additional datasets with different table structures and visual styles.
In briefPUGS: Physics-guided uncertainty modeling for underwater dynamic 3D Gaussian splatting
SummaryThis paper presents PUGS, a method for reconstructing dynamic underwater scenes using 3D Gaussian Splatting. It models each Gaussian's motion as a probability distribution that captures uncertainty from light scattering, refraction, and low visibility, rather than a single deterministic path. The approach uses stochastic deformation optimization and a dual-path self-consistency distillation to produce deterministic reconstructions while keeping temporal coherence. The authors report PSNR gains of 1.16 dB on UVEB and 3.99 dB on DRUVA over the strongest baseline.
In briefHierarchiRetina: A staged system for screening and grading diabetic retinopathy using lesion segmentation
SummaryThis paper presents HierarchiRetina, a cascade that screens, segments lesions and vessels, and grades diabetic retinopathy, with each stage's output inspectable. End-to-end evaluation on 58,689 held-out images gave a quadratic weighted kappa of 0.804, and the screening gate caused 66% of errors. Ablations found the grade ignored the lesion masks.
In briefLearning How to Remember: Choosing Memory Modality for GUI Agents
SummaryThis paper studies how GUI agents built on vision-language models should represent retrieved memory, either as text or as visuals. The authors find that the preferred memory modality depends on the interaction step. They introduce SwitchMem, a routing framework that dynamically chooses between textual and visual memory at each step, and report that it outperforms static-memory baselines on three GUI benchmarks.
In briefGLST: a memory framework for robots whose navigation goals change while moving
SummaryThis paper studies robot navigation where goals change during execution, either by redirecting to an existing goal or by introducing a new one, without explicit signals. The authors propose GLST, which decomposes latent beliefs into global-local spatial and long-short temporal representations, and add goal-conditioned rewards to balance exploration and exploitation. Simulation experiments across diverse terrains show better chained-goal navigation than baselines, and real-robot tests on the Spot quadruped further support the memory decomposition approach.
In briefCineMem: Attention-aligned fine-grained memory retrieval for consistent multi-shot long video generation
SummaryThis paper presents CineMem, a framework for autoregressive multi-shot long video generation that retrieves memory at a finer granularity. It organizes historical tokens with an online-clustered KV cache guided by attention analysis, and adds a query-level retrieval mechanism to access more history under a fixed budget. The authors also introduce VLM-assisted metrics for assessing inter-shot consistency, and report experiments demonstrating the framework's effectiveness.
In briefStein Variational Distributional RL: Score-Guided Particles for Joint Reward Distributions
SummaryThis paper proposes SVDRL, which represents the joint distribution of multi-dimensional returns with Stein particles transported by Stein variational gradient descent along a learned score field. The same field gives a reverse-diffusion sampler that answers tail-risk queries below the particle resolution. Empirically, it is more accurate than particle critics with the same number of particles on rare events and performs on par on MuJoCo continuous control.
In briefBatch-Paced: fixed-budget batch decoding for diffusion language models
SummaryThis paper addresses batch inference for discrete diffusion language models (dLLMs) and proposes Batch-Paced, a training-free method that jointly allocates decoding compute across requests under a fixed budget. Tested on two dLLM backbones and five benchmarks, it improves batch performance at matched budgets, with gains generally growing as batch size increases. The authors attribute the gains largely to higher committed-token confidence for lower-confidence requests, and note that adding a decoding-depth allowance can further reduce batch decode time at a modest performance cost.
In briefLinMorph: Learning to Morph Weights for Faster Video Diffusion Transformer Linearization
SummaryThis paper presents LinMorph, a method for accelerating pretrained Video Diffusion Transformers by converting their softmax attention into linear attention. The authors argue that the main bottleneck is initialization, since random starts discard pretrained knowledge and manual weight reuse is too rigid. LinMorph learns a cross-parameter weight transformation shared across layers, trained on short low-resolution clips and reusable at larger scales. According to the paper, it achieves faster recovery and better final generation quality than random initialization and heuristic weight reuse.
In briefNBCLT: Parameter-free local training that clips updates to norm balls
SummaryThe paper proposes NBCLT (Norm-Ball Clipped Local Training), a unified local-training framework for federated and decentralized learning. It combines a predict-and-correct local update with Euclidean projections onto flexible norm balls, keeping the implementation simple. The authors report that in the federated setting's parameter-free regime, with local computation scaled appropriately to the horizon, it achieves O(ε^-2) communication and O(ε^-4) stochastic-gradient complexity without assuming bounded heterogeneity.
WhyProposes a unified norm-ball clipped local training framework with parameter-free complexity bounds for FL and DL, a substantive theoretical advance with limited empirical validation.
In briefText-only versus image-aware AI judges for evaluating manga translation
SummaryThis paper compares how the same vision-language model judges manga translations when it sees only text versus when it also sees the page images. Using a Japanese-English dataset of official translations aligned by page and speech bubble, the authors found that adding page images helped the model detect deliberately injected, visually dependent errors. The two judging modes produced different error counts and error-category distributions, but human verification showed no consistent preference for either.
In briefSpecCloak: Probing Harmful Code Generation via Engineering Specifications
SummaryThe paper examines how large language models refuse malicious programming requests and finds that phrasing coding requirements as engineering specifications substantially reduces refusal. Building on this, the authors propose SpecCloak, a black-box red-teaming framework that achieves an average execution-verified attack success rate of 28.91% across eight target models on RedCode-Gen. They also present SpecReveal, a defense prototype that reports a 96.88% attack detection rate in a separate prompt-detection evaluation.
WhySpecification-style prompts substantially reduce refusal, supported by representation analysis, interventions, execution-verified attacks, and a detection prototype.
In briefA method for privately continual LoRA using stable and plastic subspaces
SummaryThis paper studies how to let pretrained models learn a sequence of tasks with LoRA while keeping differential privacy and avoiding forgetting old knowledge. The authors propose DP-SPA, which builds a stable subspace from previously privatized updates and a plastic subspace from pretrained weights in its orthogonal complement, then privately optimizes only the LoRA factor. Experiments on five datasets, eight settings, and multiple privacy budgets report higher average accuracy and reduced forgetting.
In briefDriveWLA: A Unified World-Language-Action Model for Autonomous Driving
SummaryThis paper proposes DriveWLA, a unified model based on Mixture-of-Transformers that jointly handles semantic understanding, future scene prediction, and trajectory generation for autonomous driving. It uses three experts (world, language, and action) linked by Cross-Expert Attention, and refines the action expert with Flow-GRPO reinforcement learning. The authors report strong results on NAVSIM v1 and v2 and zero-shot generalization to HUGSIM, nuScenes, and Bench2Drive.
In briefVisual Dependence Is Not Answerability: Testing Signals in Frozen InstructBLIP
SummaryThis paper systematically tests signals from a frozen InstructBLIP model for detecting whether a visual question can be answered, using a 10K stratified subset of UNK-VQA. Signals covered include uncertainty, CLIP alignment, visual dependence, and internal representations. Discrimination is weak, with test AUROC mostly between 0.50 and 0.54. The authors argue that relying on visual evidence differs from that evidence being sufficient to answer.
In briefFlashMask-3: Efficient, Expressive Mask-Aware Distributed Attention for Long-Context Training
SummaryThis paper addresses how structured attention masks become a bottleneck when training large language and multimodal models on long contexts across multiple GPUs. It presents FlashMask-3, which adds a mask-aware kernel for NVIDIA Hopper GPUs and a distributed extension with zero-communication mask localization, a sparsity-aware load balancer, and communication pruning. The authors report higher TFLOPs than several baselines on a single H100 and up to 3.71x throughput gains in context-parallel settings, plus a 7.3–11.0% end-to-end throughput improvement when training a 118B-A14B MoE model at 128K context on 128 H100 GPUs.
In briefPaceReasoner: Learning to Answer Briefly from Long Reasoning Without Labels
SummaryPaceReasoner is a label-free test-time reinforcement learning framework that uses a model's long reasoning responses to improve its short-mode accuracy. Both modes share one trainable policy, and learning relies on progressive, confidence-weighted supervision with a ratio controller balancing the modes. Across three model backbones and four benchmarks, it improves average short-mode accuracy over baselines by 7.4 to 13.5 percentage points while using fewer than 25 reasoning tokens per response on average. Ablations and cross-dataset results support its key components and transfer beyond the adaptation set.
In briefTesting Whether SHAP Reflects True Causal Effects for Sparse Autoencoder Features
SummaryThis paper examines whether SHAP-based attribution accurately reflects the causal influence of sparse autoencoder (SAE) features in language models. Using GPT-2 Small, a layer-7 SAE, and SST-2 sentiment classification, the authors compare attribution scores with results from feature ablation and activation steering. They find that agreement among attribution methods does not indicate causal faithfulness, and that importance and directional faithfulness should be evaluated separately.
In briefInteraction-as-a-Judge: Having computer-use agents operate generated apps to evaluate them, rather than relying only on code checks
SummaryThe paper argues that current LLM evaluations of coding agents focus on code analysis and unit tests, which can miss failures that human users would notice immediately. The authors propose Interaction-as-a-Judge, where computer-use agents directly interact with interactive applications to evaluate them. They release two datasets: 6,858 application pairs with human preferences and agent traces, and 222 applications with human-verified requirement labels. The method improves prediction of human preferences and requirement verification over non-interactive approaches, and it also helps improve generated artifacts at inference and training time.
In briefSwiss4Arg: A Parallel Multilingual Dataset for Argument Mining in Swiss Languages
SummaryThis paper introduces Swiss4Arg, an argument mining dataset built from official Swiss voting brochures in German, French, Italian, and Romansh. It covers 53 federal voting issues from 2019 to 2025, annotated for argument components and support or attack relations by three native speakers per language. The authors release adjudicated labels and individual annotations and report baseline results using GPT-5.4 and SwissBERT.
Sophia Conrad, Fatma-Zohra Rezkellah, Christina Niklaus et al.
In briefTraining LLM agents to collaborate using verifiable information rewards
SummaryThis paper introduces CoSIR, a post-training method that rewards language model agents for sharing evidence, requesting missing information, and using teammates' input. Information is scored by how much it reduces team surprisal under a Bayesian evaluator, and the models Qwen3-4B, Qwen3-8B, and Gemma-3-4B-it are trained through self-play. Evaluated on seven unseen natural-language collaboration tasks, CoSIR achieves the highest aggregate score while staying competitive on general tasks.
In briefDeep Mesoscopic Dynamics: learning field equations for magnets from microscopic trajectories with neural networks
SummaryThis paper introduces Deep Mesoscopic Dynamics (DeepMesoD), which uses neural networks to learn an effective mesoscopic field equation from microscopic trajectories at a chosen coarse-graining scale. It first recovers the known hydrodynamic limit in the Ising-Kac model. For a two-dimensional Heisenberg ferromagnet and layered CrBr3, where no closed-form mesoscopic description exists, the learned models match atomistic references and cut wall-clock time by three to four orders of magnitude.
WhyProposes learning mesoscopic field equations from microscopic trajectories with neural networks, validated on three systems with substantial speedups, useful for magnetic simulation.
In briefDiagnose-to-Generate: Fixing Local Structural Flaws in Flow Image Models
SummaryThis paper addresses localized structural errors, such as malformed hands and limbs, in flow-based image generators. It proposes D2G, which separates visual diagnosis from generative optimization, using a region-annotated dataset (Achilles-600K) and the ARGUS model to predict anomaly categories and bounding boxes. These diagnoses become spatial rewards for patch-level Flow-GRPO, and a sampling method (TPSS) limits spatial drift across rollouts. Experiments on FLUX.1-dev and SD3.5 report improved structural fidelity with FID and CLIP scores preserved.
In briefCross-lingual transfer of “evil” vectors is asymmetric by direction
SummaryThis paper examines whether the internal direction representing an “evil” persona in large language models works across languages. The authors test four open-source instruction-tuned models using English and Spanish as high-resource languages and Swahili and Telugu as low-resource ones. They find that vectors are hard to extract reliably from low-resource languages and steer poorly, while vectors from high-resource languages transfer effectively to low-resource inputs and largely preserve MMLU performance.
In briefLoopREG: Representation supervision lets diffusion Transformers gain depth at test time by reusing blocks
SummaryThis paper introduces LoopREG, a diffusion Transformer that repeatedly applies a shared block, with its intermediate states supervised by a pretrained visual encoder. On ImageNet 256×256, it lowers SiT-B/2 FID at 400K steps from 15.08 to 4.91 relative to the local REG baseline. Adding one extra inference visit with unchanged weights further improves FID from 14.71 to 5.34 in a checkpoint trained with two visits.
In briefFinding What Matters: Agentic Data Exploration for Open-Ended Multi-Hypothesis Discovery
SummaryThis paper introduces open-ended multiple hypothesis discovery, where a system receives only a large tabular dataset and an outcome and must find several data-supported hypotheses without knowing the relevant variables, functional forms, or number of mechanisms in advance. The authors build a benchmark from structural causal models with 351 planted hypotheses across 15 tasks and five domains, keeping the causal model hidden from the discovery method. They propose HypAgent, an iterative search method that removes explained signal, reaching 0.49 F1 versus 0.35 for the strongest baseline.
WhyIntroduces open-ended hypothesis discovery with a 351-hypothesis causal benchmark validated by interventions; HypAgent improves F1 over the strongest baseline.
Sanchit Kabra, Nurendra Choudhary, Jiangshu Du et al.
In briefECMR: Evidence-aware control of multi-agent LLM deliberation
SummaryMulti-agent LLM systems usually fix their discussion protocol in advance, which can mistake repeated support for independent confirmation. This paper proposes ECMR, a controller that uses answer concentration, support diversity and independence, contradictions, and remaining budget to decide whether to propose, challenge, verify, arbitrate, or stop. On a held-out evaluation of 9,000 records from MedQA, CaseHOLD, and FinQA, ECMR-Full raised pooled accuracy from 62.1% to 84.9% and lowered false-consensus rate from 40.7% to 17.0%, while using the largest token budget.
In briefPrefilling-dLLM v2: Learned Chunk Routing for Long-Context Diffusion Language Models
SummaryThis paper proposes Prefilling-dLLM v2, which continues training a diffusion language model to route attention across chunks using its native mask token, reducing the cost of long contexts. On three RULER tasks from 16K to 128K, it outperforms other sparse attention methods. Decoding at 128K length is 172 times faster than full dense attention.
In briefRT-J: Large-scale pretraining of relational Transformers for example-efficient predictions
SummaryThis paper introduces RT-J, an 86M-parameter relational foundation model for predictions on tables and relational databases. The authors pretrain it on Join, a collection of 639 web-crawled databases with 6,023 LLM-generated forecasting tasks, using long contexts and a lightweight random-walk retriever. On RelBench in-context regression tasks, RT-J reportedly needs about 10 times fewer examples on average to match strong feature engineering pipelines.
Rishabh Ranjan, Vignesh Kothapalli, Harshvardhan Agarwal et al.
In briefMicroCLIP: A vision-language model linking text and fluorescence microscopy images
SummaryThis paper introduces MicroCLIP, a vision-language model that learns cellular and subcellular image representations from natural language for open-set classification and retrieval. The authors build MicroPool-2M, about 2.5 million fluorescence microscopy image-text pairs, and align images with text at category, knowledge, and detail levels. Experiments report leading zero-shot classification and text-image retrieval performance on microscopy images, including retrieving specific cellular characteristics from thousands of candidates.
In briefTStruct: Learning Shared Temporal Structures for Long-Term Forecasting
SummaryThis paper studies long-term time series forecasting and finds that samples from different time positions and variables share dominant temporal relationship structures. The authors propose TStruct, a network that directly parameterizes a set of shared temporal dependency matrices and generates combination weights from dynamic sample information and static variable priors. According to the paper, this structure-aware design consistently outperforms several strong baselines across all datasets while maintaining high computational efficiency.
In briefSocratic Skill Learning: Teaching and Practice Combined to Build Reusable AI Agent Skills
SummaryThis paper proposes Socratic Skill Learning (SSL), where an educator agent guides a student agent that tests and writes its own reusable skill procedures. The authors build SkillTransferBench, a benchmark of 35 task families and 408 executable cases, to measure reuse on held-out sessions. On the main cohort, Fixed-Window SSL reports 0.801 utility versus 0.723 without a skill. An Issue-Guided variant reduces educator token usage while keeping similar utility.
In briefThermal-to-visible face translation using frequency routing and identity conditioning
SummaryThis paper addresses generating visible-light face images from thermal images, useful for nighttime biometrics and low-light surveillance. The proposed FreqID-Control framework assigns low-frequency structure and high-frequency detail to separate low-rank adaptation subspaces, and adds compact global identity tokens that are independent of spatial thermal tokens. Experiments on MCXFace and SpeakingFaces show strong image fidelity and identity preservation, and ablations confirm the contribution of both components.
In briefMMTA: A Benchmark for Analyzing Time Series, Text, and Vision Together
SummaryThis paper introduces MMTA, a benchmark testing whether multimodal large language models can combine time-series data with real-world visual evidence across nine domains. It includes video-plus-time-series and image-plus-time-series layouts, with a 3,524-sample test split covering classification and forecasting. Evaluating eight models, the authors find that representing time series as text produces very long token sequences and weakens alignment with visual evidence, especially in long videos. They also propose TimeOmni-v, which adds a dedicated time-series encoder and related design changes that improve long-video classification and inference efficiency.
In briefNavalia: Training Cowork AI Agents with Verifiable Tasks for Long Multi-Step Work
SummaryThis paper presents Navalia, a framework for improving Cowork Agents, which are language model agents that plan, use tools, and carry out complex tasks. It generates financial analysis tasks from corporate reports, linking each step to traceable evidence and verified answers through LLM review and programmatic checks. The verified tasks are used for supervised fine-tuning and reinforcement learning. Starting from Qwen3.6-35B-A3B, Navalia outperforms the base model on all five benchmarks.
In briefDVTR-KD: Distilling Vision-Language Models by Reconstructing Intermediate Visual Tokens
SummaryLarge multimodal models are powerful but costly to run on edge devices. This paper proposes DVTR-KD, a knowledge distillation framework that reconstructs the teacher's dense visual tokens from the student's compressed tokens. Intermediate-layer reconstruction is supervised through the teacher's response predictions, with an auxiliary attention-readout loss aligning how context is integrated. Evaluations on diverse benchmarks show improved performance for token-compressed students while keeping the inference efficiency of visual token compression.
In briefPersistence Positional Encoding: A Simple Transformer Method for Persistence Diagrams
SummaryPersistence diagrams encode multiscale topological features as sets of birth-death pairs, and this paper examines how to feed them into Transformers. The authors change only the point representation, mapping each pair through fixed random Fourier features while keeping the Transformer backbone unchanged, which they call Persistence Positional Encoding (PPE). On eight TU graph benchmarks, changing only the point representation shifts macro accuracy by 5.49 to 6.47 percentage points, more than changing the downstream module. Under a shared evaluation protocol, PPE achieves the highest mean on most benchmarks.
In briefGeoNIRS: Geometry-aware pretraining for fNIRS brain signal models across different probe layouts
SummaryGeoNIRS is a pretraining framework for functional near-infrared spectroscopy (fNIRS) data that uses source-detector positions as a shared spatial reference, so datasets with different channel counts and probe layouts can be trained together. The base model is pretrained on eight datasets and the large model on 22, without using downstream data or labels. On eight held-out tasks it achieves the highest task-averaged balanced accuracy among the compared baselines, and pretraining improves accuracy over random initialization, with larger gains when labels are scarce.
WhyExplicit probe geometry in heterogeneous fNIRS pretraining is a meaningful advance, backed by eight held-out tasks and ablations useful for the subfield.
In briefWhat Makes a Good Visual Representation for Spatial Understanding in MLLMs?
SummaryMultimodal large language models usually learn from 2D visual data, so their visual tokens often lack 3D structure. This paper studies which visual representations help spatial understanding, comparing latent features and explicit depth or point maps from 3D geometry, self-supervised, vision-language, and generative models, transferred by distillation or fusion. The authors find that a representation's 3D structure correlates strongly with its benefit, and that its form determines which transfer mechanism works best. Distilling both latent features and explicit geometry into one MLLM, using only RGB video input, improves over the base model on static 3D and dynamic 4D spatial reasoning benchmarks.
In briefStatic skill inconsistencies don't automatically mean runtime risk: a behavioral study of LLM agent execution
SummaryThe paper examines inconsistencies among documentation, scripts, and configuration files in external skill packages used by LLM agents. The authors audited 575 community and benchmark skills and manually validated 1,182 contract inconsistencies. Using 49 executable tasks, three LLMs, and three experimental settings, they analyzed 441 paired execution trajectories to measure the actual performance impact of resolved defects.
In briefFASHION-RQ: An automated method for evaluating fashion search using hierarchical product structure
SummaryThis paper introduces FASHION-RQ, a framework that uses hierarchical clusters learned by Residual Quantized Variational Autoencoders (RQ-VAE) to automatically generate user-style queries and relevance sets from catalog structure, without manual annotation or predefined taxonomies. The authors build FASHION-RQBENCH, a scalable benchmark with four tasks covering ranking, relevance classification, sibling discrimination, and paraphrase robustness. Experiments show it distinguishes retrieval models across query types and attributes, and a human study finds its queries are about as natural as real user queries with clearer intent.
Tripti Shukla, Shwetha Somasundaram, Akshat Gupta et al.
In briefLearning adaptive control for multi-agent AI models to improve answer reliability
SummaryThe paper presents ADMA, an adaptive controller that learns how multiple AI agents should propose, challenge, arbitrate, and coordinate while the underlying model stays frozen. A controller trained on legal and financial reasoning transfers without policy updates to four textual and multimodal benchmarks and two later model versions, improving accuracy by 6.2 to 14.2 points over direct answering. The gains come with much higher token use and latency, and transfer to the newer model version depends on the task.
In briefHow far visual token pruning can go: a study of its performance ceiling when offline search cost is relaxed
SummaryThe paper studies visual token pruning for vision-language models, which keeps only a subset of image tokens under a fixed budget to cut computation. The authors select one shared token subset per image, approximating utility with questions and answers generated by a frontier VLM, and propose QA mini-batch greedy (QA-MBG) to build nested subsets. Experiments on 900 images from nine benchmarks with Qwen3-VL-4B and Gemma 4 E4B show substantial headroom over evaluated pruning baselines, and the searched subsets substantially narrow the gap to full-token inference.
Janghoon Cho, Shubhankar Borse, Reza Pourreza et al.
In briefAgentic-MatriBO: Agent-Based Bayesian Optimization Compiled from Reusable Capabilities for Materials and Chemistry Design
SummaryThe paper introduces Agentic-MatriBO, a framework that represents Bayesian optimization components as composable capabilities, with a language model proposing plans that a compiler checks before execution. Compared with eight baselines on four offline candidate pools, it reports the best mean endpoint quality across tasks. On COF, it reached the tabulated optimum in all five seeds. Ablations indicate that removing the capability compiler or evidence gate coincides with lower cumulative quality and fewer exact-optimum hits.
In briefDiff Diffusion: Using Diffusion Models to Predict Code Edit Paths Efficiently
SummaryThis paper proposes Diff Diffusion, a diffusion-inspired language model framework for code editing that predicts edit paths from an original snippet to its revised version. It represents denoising as structured edit operations and combines multi-stage training with parallel decoding so computation focuses on changed regions. On public edit-completion benchmarks it matches or beats autoregressive baselines while running 2.78 times faster.
In briefQK-Edit: A training-free method for text-guided image and video editing with MM-DiT attention
SummaryThis paper proposes QK-Edit, a training-free framework for text-guided image and video editing that exploits the attention dynamics of Multimodal Diffusion Transformers (MM-DiTs). It uses a query-key manipulation strategy to adjust key attention components, balancing prompt fidelity and structural consistency. The authors report edits such as object addition, removal, replacement, and style changes, tested on FLUX and HunyuanVideo.
In briefTraining and Testing Reward Models That Check Whether Search Agent Answers Are Backed by Multiple Independent Sources
SummaryCurrent methods for training and evaluating search agents mostly judge final answer quality and pay little attention to how answers map to retrieved evidence. The paper proposes CSEG, a framework that uses a six-layer decision tree to assess full search trajectories for source independence and handling of conflicting evidence, along with the CSEG-Bench benchmark of 1,160 verified trajectories. The trained reward model CSEG-RM reaches 83.9% exact accuracy on the benchmark, and using it to supervise an answer policy raises the independently judged mean label from 2.83 to 4.19.
In briefChecking the Checkers: Verifier Noise and Robustness in RLVR
SummaryThis paper surveys 58 studies on the errors made by verifiers used to score language model responses in reinforcement learning with verifiable rewards (RLVR). It finds that the structure of an error matters more than its rate, and that false positives a policy can learn to trigger may cause plateaus or collapse. The authors compile comparison tables, reconcile conflicting findings, and propose a reporting protocol to make future results comparable.
In briefFollow-the-Perturbed-Leader for linear bandits achieving √T regret on general action sets
SummaryThe paper presents the first adversarial linear bandit algorithm based on Follow-the-Perturbed-Leader (FTPL) that achieves √T regret over general action sets. Using only linear optimization oracles it attains Õ(d^{5/4}√T) regret, and near-optimal Õ(d√T) when an optimal-design exploration distribution is also available. The approach is simpler to implement and generally more computationally efficient than prior oracle-efficient methods based on continuous exponential weights.
WhyFirst FTPL-based adversarial linear bandit achieving √T regret on general action sets, with a Fisher-information stability analysis that substantially improves theory.
In briefEdge Agents: Analyzing Edge-Cloud Collaboration in LLM Agent Execution
SummaryThe paper introduces EC2-Bench, a benchmark for evaluating edge-cloud collaboration in long-horizon LLM agent tasks, covering task performance, resource usage, and privacy exposure. Collaboration offers a favorable balance among these factors compared with edge-only or cloud-only execution. As edge models improve, collaborative performance approaches cloud-only levels, while privacy exposure stays strategy-dependent. The authors also identify cloud-invocation decision-making and instruction following as key edge-side abilities.
In briefPerturbation, Not Provenance: Why Activation Steering Fails as a Text Watermark
SummaryThis paper tests whether linear semantic directions in language model activations can serve as a robust text watermark by steering generation and probing re-encoded text. Across two model families, four directions including a random control, and two paraphrasing attacks, activation-steering watermarks were outperformed by the logit-based KGW watermark in detection accuracy. The authors found that a random direction was as detectable as a meaningful one, concluding the signal is a global, content-independent shift rather than a recoverable semantic mark.
In briefSPARCL: Active learning that spreads chest X-ray labeling across regions
SummaryThis paper proposes SPARCL, an active learning method for chest X-ray lesion segmentation that balances prediction uncertainty with cumulative spatial coverage, directing annotation toward under-labeled areas. It also introduces two metrics, Spatial Inconsistency and Performance Coverage, to measure anatomical fairness across active learning rounds. On an in-house calcification cohort and public CXR benchmarks, SPARCL reaches 95% of fully supervised performance using 24% of the data, versus 29% for the best baseline.
In briefLearning from Fracture: Teaching Language Models to Design Through Physical Experience
SummaryThis paper introduces FracLearn, a framework that lets language models learn from simulated fracture experiments in a Tetris-inspired game. The model abstracts observations into conditional principles about reinforcement, crack evolution and load transfer, then combines them in new architectures. The best composite score is 11.0% above the experience-stage record, with analyses linking gains to crack deflection and temporary trapping. Learning occurs without updating model weights.
In briefTeaching Robots to Use Their Whole Arm, Learned from Human Demos
SummaryRobots usually manipulate objects with their end-effector, while humans often use their whole arm. This paper proposes Whole-Arm Contact Learning (WACL), a hierarchical framework that describes tasks with semantic keypoints, a reference contact trajectory, and safety path constraints generated by a vision-language model, then solves for contact link, location, and joint configuration with a quasi-static surface-contact optimizer. WACL learns from 10 bare-hand human demonstrations per task without teleoperation. On a 7-DoF arm across four whole-arm tasks, it achieved 78% success under end-effector constraints versus 33% for a strong link-pose baseline, and generalized to new constraints without retraining.
In briefSemantic segmentation fusing RGB with complementary sensors via a bidirectional polarity-aware module
SummaryThis paper addresses semantic segmentation that combines RGB images with auxiliary sensors such as thermal, LiDAR, depth, event, polarization, or light field data. The authors propose a bidirectional polarity-aware cross-modality fusion (BPCF) module to capture complementary cues and improve feature alignment. The framework is evaluated on six modality pairings and reportedly achieves state-of-the-art results on eight public datasets. Code and trained models will be released upon acceptance.
In briefGIFT-Eval-CTX: A Benchmark for Evaluating Contextual Forecasting Models
SummaryThis paper introduces GIFT-Eval-CTX, a benchmark testing whether forecasting methods can turn verified future information, such as planned events, into accurate numerical forecasts. Built on GIFT-Eval, it contains 10,000 instances across 20 datasets, annotated with macro and micro structured effects derived from the realized future. The authors report that structured effects lower aggregate relative MASE for GPT-5.6 Sol and Claude Opus 4.8 compared with history-only prompting, and that the MoiraiAgent system performs best overall.
In briefMHVF: Training-free GUI element locating via verified multi-hypothesis search
SummaryThis paper proposes MHVF, a training-free method that helps web agents locate the interface element matching a natural-language instruction. It uses one pretrained vision-language model to generate several spatial hypotheses from full-screen predictions, re-localize them in zoomed regions, and verify candidates against the instruction in marked local views. Experiments with five models across eight benchmark settings show accuracy gains in most settings, such as Qwen2.5-VL-7B rising from 23.6% to 53.6% on ScreenSpot-Pro.
Minghua Chen, Xingming Liao, Weiqiang Zhong et al.
In briefWhen, Where, and Why: Causal Explanations for Video Face Understanding with MLLMs
SummaryThis paper presents VideoFace-X-LLM, a framework for explainable facial video understanding, including the ExplainFace weakly supervised video-text corpus and the H-TCC method. H-TCC identifies key temporal moments (when) and facial regions (where) and tests their causal necessity (why) through targeted occlusion. Fine-tuning on ExplainFace raises appearance Macro F1 from 19.48% to 32.87%.
In briefDexStitch: Learning and composing dexterous robot skills from human videos using contact-aware inpainting
SummaryDexStitch is a framework that splits long human videos into manipulation skills and the transitions between them, then converts them into robot-compatible observation, action, and skill data. It uses per-finger human-object contact to guide dexterous hand retargeting and robotic inpainting, and stitches learned skills into requested sequences. On four household skills, it reports 80.0% single-skill success and 77.5% sequence success.
Byeonggyeol Choi, Seongmin Jung, Unghui Nam et al.
In briefPosterior-Averaged Bellman Flow Matching for Distributional Critics
SummaryThis paper proposes Posterior-Averaged Bellman Flow Matching (PA-BFM), a method for training distributional critics that model full return distributions. It samples a pool of Bellman endpoints, computes a closed-form posterior over a latent anchor, and regresses to the posterior-mean velocity to reduce target variance. Controlled studies report a 79.7% reduction in fixed-query target mean-squared error, and across 50 OGBench tasks and four offline-to-online scenarios, PA-BFM achieves higher aggregate performance than Value Flows.
In briefAgentic4PDM: a framework where a host governs which evidence counts in industrial fault diagnosis
SummaryThe paper treats industrial fault diagnosis as classifying sensor time-series windows into health states or fault types, where labels are scarce and operating conditions change. LLM agents investigate signals with reusable analysis tools, while a deterministic host controller checks requests and assigns evidence roles based on comparisons with labeled references. The authors report higher aggregate macro F1 than the evaluated diagnostic and agentic baselines on selected bearing and motor challenges and an unfiltered industrial test set.
Nam V. Nguyen, Hoang Van Truong, Hieu D. D. et al.
In briefWhere Words Begin and End: Locating Word Boundaries for Forced Alignment in Speech
SummaryThis paper examines forced aligners, tools that locate each word of a known transcript in audio. Comparing nine configurations against human word boundaries on English read, conversational and telephone speech, the authors find the Montreal Forced Aligner most precise, while CTC-based aligners consistently place word starts late. The authors fine-tune Whisper with LoRA to predict word start and end timestamps, training only on human-placed boundaries across 87 languages, and report it outperforms MFA on every English test condition.
WhyFine-tunes Whisper with LoRA for timestamp prediction, comparing against MFA and CTC aligners across English benchmarks and telephone audio with solid evidence.
In briefDiffuplex: A Discrete-Diffusion Spoken Model That Listens and Talks at Once, Balancing Intelligence and Speed
SummaryThe paper introduces Diffuplex, a full-duplex spoken language model built on discrete diffusion that separates deciding when to speak from deciding what to say. An autoregressive policy tracks the user's speech while parallel diffusion generates and refines the response. With thinking mode, it reports 13.0 tokens per forward pass and about 1872 tokens per second. Without thinking, it matches existing full-duplex models' first-audio latency while showing stronger linguistic ability.
In briefParallel BesselKAN: a new network design that merges spline and Bessel basis branches without extra parameters
SummaryThis paper addresses limits of Kolmogorov-Arnold Networks (KANs) on spatial data, where default B-Splines are restrictive and mixing basis functions causes numerical ill-conditioning. The proposed Parallel BesselKAN runs B-Spline and low-order Bessel branches in separate pipelines and joins them through a parameter-free quadratic interaction. The authors report a 35% reduction in parameter footprint and 97.99% accuracy on MNIST, while noting that standard MLPs keep a hardware speed advantage on pure vision tasks.
In briefGENDIS: Gender Discourse and Stance in Multilingual YouTube Comments from India, Kenya and Nigeria
SummaryThis paper introduces GENDIS, a dataset of 9,600 human-annotated multilingual YouTube comments from India, Kenya and Nigeria, with 3,200 per country. The comments are labeled in two steps: first whether gender is referenced, then whether gender-related content is negative, positive or neutral. In a uniform random sample, 23.8% to 35.0% of comments were gender-related, and 82% of those with a resolved stance were neutral or positive.
Luise Koch, Manuel Tonneau, Niyati Malhotra et al.
In briefRethinking Higher-Order Theory of Mind Tests for Large Language Models
SummaryThe paper argues that existing tests treat belief-nesting depth as reasoning order, so some higher-order questions can be answered with lower-order shortcuts. The authors introduce HopToM, a benchmark with certified full-order questions and controls, and report that four frontier LLMs' mean accuracy drops from 73.5% at first order to 20.0% at fourth order under direct prompting. They also propose HopTrace, an inference-time method that improves fourth- and fifth-order accuracy on certified narrative questions.
WhyFormalizes effective reasoning order, exposes order inflation in benchmarks, and proposes certified HopToM and deterministic HopTrace with substantial experiments and clear gains.
In briefPatient-specific deformable image registration using clinical context and deformation history for adaptive radiotherapy
SummaryThis paper presents CAP-DIR, a registration method that integrates source contours, dose, clinical text and prior-registration features to track anatomical changes during radiotherapy. It was evaluated on 16,361 CT examinations from 3,473 patients across four institutional cohorts and five public benchmarks. The authors report higher organ-at-risk region similarity than the strongest learning-based comparator, improved contour acceptability in a blinded longitudinal review, and a 34.5% reduction in global dose-mapping error.
In briefMany Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders
SummaryThe paper argues that when training LLM recommenders with reinforcement learning on Reddit recommendation threads, the current Rank-GRPO method rewards each list against only one reply, so its advantage mostly reflects which reply was sampled (85% of variance) rather than list quality. The authors propose ReMaRk, which scores each list once against the union of all replies' titles and samples each request equally. Across seven backbones and three datasets, ReMaRk reaches the validation recall of 2,000 per-reply steps in 25 to 53% of the steps and leads at that budget.
In briefMore Reader Data or a Better Teacher? Updating Crowd Highlight Models in 2026
SummaryThis paper compares three ways to improve a model that predicts which sentences readers collectively highlight on a social highlighting platform: using newer language models directly, using them as teachers, or adding more reader data. Adding training pages with reader data gave a clear gain, while a stronger teacher model that breaks ties produced no material improvement. The authors also report that frontier models used as direct rankers still trail a small model trained only on reader highlights.
In briefLearning when to reason efficiently: Root-token policy optimization for adaptive thinking
SummaryLarge reasoning models often spend the same heavy computation on easy and hard questions, a problem called overthinking. The authors propose RPO (Root-token Policy Optimization), which trains only the initial root token, such as whether to invoke the <think> tag, using group relative rewards and group-wise advantages. According to the paper, RPO reduces training overhead and VRAM usage, learning difficulty-aware adaptive thinking at about 2% of the training compute of prior methods.
In briefMemory Echo: Repeated Old Records Skewing AI Agent Answers, and a Fix
SummaryThis paper studies a problem in long-term memory for AI agents, where the same historical information from one source is repeatedly retrieved and written back, influencing current answers without independent evidence. The authors define this as "Memory Echo" and introduce MEchoBench to measure it. They also propose the FPA framework, which keeps one representative per source family, uses verified corrections, and grounds answers in authoritative evidence. On 190 fixed questions, FPA raised accuracy and reduced contamination effects across two models.
In briefAutomated 3D Visual-Thermal Inspection of Solar Farm Components Under PV Panels
SummaryThis paper presents Tricorder, an automated system that inspects components under photovoltaic panels, such as connectors, cabling, bolts, and structural parts, which aerial surveys struggle to cover. It combines LiDAR-inertial SLAM, geometrically calibrated RGB and thermal imagery, and machine-learning detection and segmentation to localise assets in a 3D map. The authors demonstrate the approach on utility-scale solar farm data for MC4 connector detection and connection-state recognition, thermal anomaly identification, and detection of bolts, nuts and torque tubes.
Lachlan Tychsen-Smith, Mark Cox, Brett Wood et al.
In brief"Connecting the Dots": Persistent Performance Gaps in Speech Recognition Between Standard and Minority English Accents
SummaryThis paper evaluates 16 prominent automatic speech recognition (ASR) systems released from 2021 to 2025 across 8 datasets covering diverse English accents. It finds that the performance gap between Mainstream US English and minority accents does not reliably shrink over time, with minority-accent Word Error Rates sometimes worsening by up to 65% absolute between system generations. The authors argue that standard benchmarks implicitly favor Mainstream US English, reflecting systemic issues in language technology development.
Alexander Metzger, Aruna Srivastava, Ruslan Mukhamedvaleev et al.
In briefGaussian process bandits that check AI system feedback selectively
SummaryThe paper studies how to optimize AI systems such as prompts, retrieval systems, and agent workflows when cheap model-based evaluations can disagree with trusted task outcomes. The authors build a fixed audit design and derive a finite-time regret bound that accounts for a hard audit budget, and they propose a safe adaptive extension. Synthetic experiments show the adaptive allocation reduces aggregate regret under the exact dictionary, though its finite-budget error certificate is numerically loose.
In briefInstructions as Evidence: Measuring Whether AI Agents Meet Real Needs
SummaryThe paper argues that user instructions are often incomplete, so agents should infer the true requirements from the instruction and environment rather than treat the instruction as the goal; it calls this outcome alignment. The authors build TiGanEval, a benchmark of software engineering and research tasks where requirements are discoverable but partly unstated. Evaluating eight frontier models, they find success rates drop substantially when moving from full specifications to partial instructions, and consequence feedback does not close the gap.
In briefIntermediate coordination settings for LLM multi-agent systems, between centralized and decentralized control
SummaryThis paper proposes Degrees of Freedom (DoF), a continuous framework that measures how much decision-making is delegated to downstream agents in LLM-based multi-agent systems. Experiments across several task domains show that an intermediate DoF configuration achieves higher mean accuracy than both endpoints in five of six evaluated states. The best-performing configuration depends on the task domain.
In briefI3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation
SummaryThis paper presents I3L, a system combining a low-cost open-source whole-body teleoperation interface (JoyLo+) with iterative policy refinement to close the reliability gap between offline training and deployment. It was tested on three robot embodiments and eight simulated and real tasks, including four long-horizon whole-body mobile manipulation tasks. Reported results include a 35.6-point average gain over demonstration-only training and a 72% reduction in takeover duration over four rounds.
Wensi Ai, Christopher Luey, Karthik Dharmarajan et al.
In briefAuditing Failure Attribution: Label Predictability and Judge Behavior
SummaryThis paper studies benchmarks where LLM judges identify which step or agent caused a failure, noting that surface cues can produce high scores without genuine cross-step checking. The authors propose four tests (input restriction, cue following, residual accuracy, and aggregation) to separate real checking from cue use. On three benchmarks, simple predictors match or beat GPT-4.1, and the authors recommend reporting these tests when developing such benchmarks.
In briefAttenIR: Aligning Dense Attention with Mesh Warping for Label-Free Image Retargeting
SummaryThis paper proposes AttenIR, an unsupervised framework that resizes images without manual labels, using content-aware mesh warping. It combines dense visual attention with sparse mesh constraints and guides mesh deformation with saliency priors, jointly optimizing scale, structure, and geometry preservation losses. A coarse-to-fine projection strategy aligns the attention priors with the mesh so that deformation is content-aware at the pixel level. The authors report competitive or superior results compared with traditional and state-of-the-art learning-based methods.
In briefGeometry Predicts Per-Example Neighborhood Loss Under PCA Truncation
SummaryThis paper studies how PCA dimensionality reduction affects text embeddings unevenly across examples, even when average performance stays similar. Using static word embeddings and modern sentence encoders, the authors find that each example's neighborhood retention can be predicted from the representation's geometry and its alignment with the retained PCA subspace. They argue that aggregate metrics hide this per-example variation.
In briefJoint evolution of tools and model for discovering scientific equations
SummaryThis paper proposes a framework that improves both the executable tool library and the tool-use policy of an LLM used to discover scientific equations from data. The agent grows its tools from basic numerical operations by diagnosing its own failures, while a small open model is fine-tuned on successful tool-use trajectories collected with the frozen library. Ablations indicate the two components are complementary, and the full system reports accuracy on several benchmarks..
WhyProposes co-evolving tool library and tool-use policy via verifiable fit-error feedback; ablations show complementary gains across several benchmarks.
In briefHiSTOR: Learning structured representations to predict treatment outcomes from high-dimensional longitudinal data
SummaryThis paper proposes HiSTOR, a framework that predicts potential outcomes under different treatment sequences when covariates are high-dimensional and confounding changes over time. It combines nonlinear dimension reduction with longitudinal counterfactual prediction, learning low-dimensional representations of current patient states and evolving histories. Experiments on CVSim and CancerSim report that it outperforms state-of-the-art methods across varying covariate dimensions, confounding strengths, and prediction horizons.
In briefDuplexAct: How Interactive Are Frontier Full-Duplex Spoken Language Models?
SummaryThis paper introduces DuplexAct, a benchmark for evaluating whether end-to-end full-duplex spoken language models take appropriate interaction actions during conversation. It defines a three-level taxonomy of eight interaction actions and 24 interaction conditions, and measures success through immediate execution and subsequent follow-through. The benchmark contains 4,691 test instances across casual chat, meetings, customer service, and interviews, with 2,531 drawn from real-world conversation recordings.
In briefTraining for multiple candidates, updating with one sample: a new method to improve multi-sample reasoning.
SummaryThe paper addresses candidate coverage in multi-sample reasoning, the probability that at least one of K candidates is correct, which is not directly optimized by minimizing average per-attempt loss. The authors formulate finite-K posterior learning and propose a Fenchel-Bregman method that needs only one posterior draw per update, instantiated as Fenchel-Bregman IVON (FBI) using a variational Gaussian optimizer. Across diverse recursive and autoregressive reasoning tasks, FBI improves candidate coverage, and in a matched comparison it reaches comparable Pass@10 using 12.3% of the direct Monte Carlo baseline's training time per update.
Sungjun Lim, Adrian Robert Minut, Nico Carlsen Daheim et al.
In briefExploring with Memory, Learning without It: A Self-Distillation Method for LLM Reasoning Exploration
SummaryThe paper proposes LEAF, a method to improve exploration efficiency in reinforcement learning with verifiable rewards (RLVR). Before each rollout, LEAF places a dynamic, problem-specific memory into the context, containing the latest response, its reward, and other environment feedback, and updates the policy via self-distillation. On four benchmarks and two open models, it explores about 4x faster than a random sampling baseline. On unseen test sets, it improves accuracy over the strongest baseline by 1.59 and 1.83 percentage points on the two models, using less compute than any RL baseline.
In briefCoCA: Criterion-Aware Credit Assignment for Better Rubric-Based Reinforcement Learning
SummaryThis paper proposes CoCA, a method for rubric-based reinforcement learning, which trains language models by scoring responses against explicit criteria. Standard approaches compress criterion feedback into one scalar reward and give all tokens uniform advantages. CoCA instead derives criterion-level credit from group-relative outcomes and distributes it across tokens based on criterion-induced distribution shifts. The authors report an average gain of 4.2% over standard rubric-based RL on medicine and science benchmarks.
In briefMachine Taste in K-pop: Do AI Models Agree on What Is Good?
SummaryThis paper studies how foundation models evaluate creative work, which the authors call "machine taste," using K-pop songs and music videos. The authors treat aesthetic evaluation as a pairwise ranking problem and measure agreement between models, group-level consensus, and alignment between visual and audio preferences. Models agree more on visual content than on audio, and preferences across the two modalities are only weakly aligned. Visual consensus rankings show positive but varying correlations with historical YouTube viewership across artist groups.
In briefDT-LoRA: Going Beyond the Tangent Space of Low-Rank Adaptation
SummaryThis paper studies why low-rank adaptation (LoRA) often trails full fine-tuning, arguing that its updates are limited to the tangent space of the current low-rank factors. An analysis of ten LoRA variants finds that the leftover gradient components outside this space hold most of the gradient energy and are spectrally concentrated. The authors propose Dynamic Tangent LoRA (DT-LoRA), which splits a fixed rank budget into a stable branch and a dynamic branch that captures the main residual gradient directions. The method adds no trainable parameters or inference cost, and the paper reports consistent gains over LoRA and its variants on vision and language tasks.
WhyAnalyzes LoRA's residual gradients via tangent space and proposes a dual-branch dynamic low-rank update, validated against many LoRA variants.
In briefSafety for Whom? A benchmark testing whether first-person AI assistants understand harm to users and others, not just refuse
SummaryThis paper examines safety in egocentric multimodal assistants that see through the user's eyes, arguing that existing evaluations only measure refusal and ignore who would be harmed. The authors introduce STOU, a benchmark of 2,902 video-audio and video-text queries whose potential harm targets the user, bystanders, or society. Testing 10 open and closed models, they find refusal varies systematically with the harm target and that models often recognize a harm yet still assist. The authors conclude refusal rates alone cannot certify safety.
In briefTSRL: A trigger-based method that trains AI agents on new tasks while limiting forgetting of old skills
SummaryLarge language model agents can forget previously learned skills when they learn new ones. This paper proposes TSRL, which uses reinforcement learning as the main training method and applies expert supervision only on states where the agent's policy fails. Experiments on ALFWorld, WebShop, and Deep Search at 7B and 32B scales report that TSRL retains up to 9.6 points more prior-task capability than SFT-to-RL baselines while matching or exceeding target-task performance.
In briefEvidence-First Agentic Verification (Sauron)
SummaryThis paper introduces Sauron, a verifier for long-horizon, semi-verifiable agent tasks that examines both environment state and execution trace, and accepts a claim only after validating it against its source. Bounded queries and joint checking of related criteria reduce context use, and each verdict includes an auditable record of supporting evidence. On 2,586 human-labeled criteria from 100 personal-assistant episodes, Sauron achieves the highest F1 of five verifiers across eight backbones, and it detects 97% of injected faults while flagging only 2% of clean controls.
Reza Sanayei, Razvan-Gabriel Dumitru, Sunjie Hou et al.
