In briefLEAP: Learned Block-wise Evidence Retrieval for Long Audio-Video Question Answering
SummaryThis paper addresses hour-long audio-visual question answering, where context limits make encoding whole recordings impractical. LEAP splits a recording into fixed-duration blocks, scores candidate windows in each block with a lightweight localization pass, and answers from the top-ranked windows in one bounded pass. The authors report 4.5–16.8% gains over the Qwen3-Omni-30B-A3B baseline across several AVQA benchmarks, and transfer to MiniCPM-o 4.5 with gains of 3.1–13.0% over its published results.
In briefReliability-Aware Checkpoint Selection for Domain Generalization: Reselecting Models Using Source Reliability Within an Accuracy Tolerance
SummaryThe paper argues that choosing checkpoints in domain generalization by source-validation accuracy alone may not yield reliable probabilities on unseen target domains. It proposes accuracy-constrained reliability selection (AC), which keeps checkpoints near the best source accuracy and ranks them by source-side NLL and class-wise calibration error, without using target data or extra training. Across 360 runs on OfficeHome and TerraIncognita, the reference rule reduced target calibration error and NLL relative to source-accuracy selection, with a small rise in mean target accuracy.
In briefConflictGuide: Making Competing Model Behaviors Visible Improves AutoResearch
SummaryThis paper argues that LLM-based AutoResearch systems, which edit model code based only on a single task score, ignore trade-offs between competing model behaviors. The authors propose ConflictGuide, which identifies competing behaviors and designs probes to measure them, then uses a two-stage evolution that explores with task feedback and later steers edits toward conflict alleviation. Across five model families, it reduces task and conflict-related errors by up to 28% and 14% relative to scalar-only AutoResearch.
In briefTauIndianBankBench: A Benchmark for Procedure-Aware Evaluation of Banking Chatbots
SummaryThis paper introduces TauIndianBankBench, a benchmark of 1,000 Indian retail-banking tasks in which an assistant helps a simulated customer. In addition to checking the final records, it verifies that the assistant performed every required procedural step and made no unrequested changes. When only final records are checked, 6.5% to 12.0% of sessions per model are judged successful even though the assistant skipped a required step or made an unrequested change. The authors argue that scoring both procedure and outcome gives a more accurate picture of language model performance in Indian retail banking.
In briefIs RoPE Reaching Its Limit? Theory, Diagnosis, and Fixes for Long-Context Failures
SummaryThis paper studies why language models using RoPE (rotary position embeddings) fail on long contexts, arguing that RoPE trades off stable token preferences against distinguishing nearby positions. The authors extend prior theory to allow unequal query-key scales across frequencies and introduce RoPE Profiler, a toolkit that reuses cached activations to produce semantic and positional diagnostic scores. Across 49 long-context task settings, reasoning tasks mainly suffered semantic reversal while retrieval tasks mainly suffered positional insensitivity. Targeted high-frequency rescaling, without additional training, improved accuracy by up to 20 points on Qwen3-8B and 25 points on Llama-3.1-8B-Instruct.
In briefAdaGEPA: Adaptive feedback allocation for improving AI prompts through reflection
SummaryThis paper introduces AdaGEPA, a method that adaptively chooses which examples to use as feedback when revising prompts for language models, based on the prompt's performance and task structure. In each feedback minibatch, it replaces at most one example to target an identified weakness. Across six downstream benchmarks under matched rollout budgets, AdaGEPA achieved higher mean validation scores than non-adaptive feedback selection. In an initial Schema-Guided Dialogue study, half-budget prompts outperformed full-budget baseline prompts on joint goal accuracy.
In briefHyperspectral image super-resolution that generalizes to unseen sensors at any scale
SummaryThis paper proposes OmniHSR, a model for hyperspectral super-resolution that aims to handle new sensors and arbitrary upsampling scales with one network. Instead of predicting spectral values directly, it predicts band-shared spatial operators that are composed into a continuous field and applied to all bands. Trained only on ARAD with 0.538M parameters, it outperforms directly transferred baselines on six unseen datasets without target-domain data, and runs inference up to 36 times faster.
In briefCoVisco: A Codec-Native Vision Encoder for Unified Image-Video Understanding
SummaryThis paper presents CoVisco, a vision encoder that works directly with codec-format inputs and compresses visual tokens natively during encoding, for unified image and video understanding. Long videos are split into temporal segments, each with learnable abstract tokens, and segmented attention with abstract-communication layers preserves video-level context. In the four-segment, 64-frame setting, abstract-only inference uses 400 visual tokens and reaches video understanding performance close to, and on some benchmarks above, OneVision-Encoder.
In briefMCD: Causal Distillation of Multimodal In-Context Learning in Vision-Language Models
SummaryLarge vision-language models learn well from examples in context, but this ability weakens as models get smaller. The paper proposes Multimodal Causal Distillation (MCD), which uses structure-preserving token interventions to find causal evidence and transfers how a teacher responds when that evidence is kept or removed. Across three model families and seven benchmarks, MCD improves student performance by 7.23 points on average and beats vanilla distillation by 4.68 points.
In briefStable Scores, Leaking Benchmarks: Measuring Contamination in Public Brain-Tumor MRI Datasets
SummaryThis paper examines whether public brain-tumor MRI classification datasets keep test data independent of training data. It proposes a three-layer contamination framework (duplicate, patient, and source-label leakage) and audits the three most-used corpora using a chest-radiograph control. The authors find severe leakage, including near-duplicate images across splits and patient overlap, and show that removing leaked test images leaves accuracy essentially unchanged. They conclude that stable accuracy alone does not show a model recognizes tumors, and they release contaminated-file lists and deduplicated splits.
WhyProposes a three-layer contamination framework and audits major brain-tumor MRI corpora across nine architectures, finding leakage persists after deduplication, a notable result for benchmark evaluati
Bhanu Prakash Vangala, Sowmya Guda, Latha Peddi et al.
In briefNavHarness: Adaptive Goals for Agentic Vision-Language Navigation
SummaryThis paper presents NavHarness, an agentic framework for vision-language navigation, where robots move according to instructions and observations. It uses four agents: a Goal Agent that sets adaptive goals, a Verify Agent that checks goal completion, a Memory Agent that compresses multimodal history, and a Visuomotor Agent that executes navigation actions. The authors evaluate it on R2R-CE and RxR-CE, and in real-world tests over eight challenging routes, each run three times, it reports 83.3% success and 1.51 m navigation error.
In briefCluster Attention Neural Operator for solving parametric PDEs
SummaryTraditional simulations of parametric partial differential equations require repeated computation for each parameter, which limits high-fidelity design. This paper proposes the Cluster Attention Neural Operator (CANO), which uses a cross-attention mechanism that dynamically clusters queries while keeping full-resolution keys and values. The authors report state-of-the-art results on benchmarks covering fluid and solid dynamics, irregular geometries, and long-term temporal rollouts.
Ming Zhong, Antonio Colanera, Gianluigi Rozza et al.
In briefThe Concrete-Arbitrary Gap: LLM Kinship Reasoning Depends on Presentation
SummaryResearchers tested language models on 500 formally matched kinship graphs, comparing relations expressed in familiar vocabulary with those using explicitly defined nonce predicates. Familiar vocabulary yielded higher accuracy across all four models, with all paired gaps statistically significant. Adjusting reasoning budgets and prompt language substantially narrowed the difference, indicating it is modifiable rather than fixed.
In briefTRACE: Choosing the Best Search Trajectory in Parallel Search Agents
SummaryThe paper introduces TRACE, a lightweight learned selector that ranks completed search-agent trajectories using the evidence behind their answers, instead of relying on majority voting. It links rollouts through shared content or document identity and propagates evidence across them, without extra search or generative aggregation. The authors report gains over voting on six WebQA policies and six long-horizon dataset-backbone combinations, and at least 10x higher throughput than compared aggregation methods.
WhyFrames parallel search consolidation as trajectory selection with cross-rollout evidence propagation, validated across multiple benchmarks against voting and generative aggregators.
In briefA reinforcement learning framework that plans chronic multimorbidity treatment while respecting patient preferences
SummaryThe paper proposes FAHOC, a hierarchical reinforcement learning framework for treatment planning for patients with comorbid hypertension and type 2 diabetes, which incorporates patients' willingness to adhere to recommendations. It is evaluated on longitudinal data from about 50,000 patients across five Southeast U.S. hospitals. The authors report a quality-adjusted life year improvement over observed clinician practice and no violations of patient preferences in held-out testing.
Nafiseh Payani, Soham Das, G. Anthony Wilson et al.
In briefDoGBench: Can AI Agents Write Software Docs to Expert Standards?
SummaryThis paper introduces DoGBench, which the authors describe as the first benchmark for generating and maintaining real user-facing software documentation. It contains 292 items from open source projects such as Helm, PostHog, and Mautic, where agents must decide whether docs need updating and either produce an acceptable patch or abstain. The best-scoring of seven agents reached 47.3 out of 100 on a 117-item held-out split. An audit of failures highlights gaps such as describing interfaces without checking how readers use them and leaving other affected pages stale.
WhyA first benchmark for user-facing documentation generation, with 292 real open-source items and maintainer-validated rubrics, giving solid evidence and useful analysis.
In briefParents' Mixed Views on AI in Children's Pretend Play
SummaryThis study interviewed 10 U.S. parents of children aged 4-15 about AI supporting children's pretend play. Parents saw potential benefits, such as making role-based play accessible and helping them join family play, but also worried about lost imagination and creativity, emotional attachment to AI, and reduced human interaction. The authors conclude that AI's appropriateness and child-AI interactions are key requirements alongside technical safeguards and parental control, and call for longitudinal research.
Sungho Oh (Sander Oh), Mohammad Namvarpour (Matt Namvarpour), Maxi Heitmayer et al.
In briefOSWorld-Science: A Benchmark for AI Agents Using Scientific Software
SummaryThis paper introduces OSWorld-Science, a benchmark and evaluation environment for testing computer-using agents built on vision-language models on scientific software. It contains 146 tasks across 12 models, covering workflows such as molecular drawing, pathology image analysis, statistical computing, and physical simulation. Evaluation inspects application states and generated artifacts and awards partial credit. The authors report that current state-of-the-art models with a strong harness still struggle with key scientific tasks.
In briefTransferable Graph Metanetworks: Generalizing Across Neural Network Widths
SummaryA weight space network takes another neural network's weights as input and predicts its properties. The paper proposes Transferable Graph Metanetworks, modified to be invariant to different widths representing the same function and continuous so similar functions get similar predictions. These changes markedly improve size generalization on all tasks studied. Performance is strongest on networks trained with maximal-update parameterization, remaining robust up to 42 times the training width.
WhyProposes width-transferable graph metanetworks with proved size-generalization guarantees under muP via infinite-width theory, supported by experiments on several tasks; impact is mainly within weight
In briefDo Better Goal Representations Improve Goal-Conditioned RL?
SummaryThe paper studies how much the quality of goal representations affects offline goal-conditioned reinforcement learning. The authors build an exact temporal-distance goal representation in deterministic mazes and deliberately degrade its geometry while keeping the learner fixed. On OGBench navigation tasks, large changes in goal quality barely affect performance, whereas the same interventions on the current state more than double success, identifying the state pathway as the main bottleneck.
Syed Nazmus Sakib, Abdul Monaf Chowdhury, Nafiul Haque et al.
In briefLearning Where to Look: Guided Attention for Cardiac MRI Vision-Language Models
SummaryThis paper studies fine-grained visual question answering on cardiac MRI through anatomical grounding and guided attention. The authors built 128,915 anatomical-grounding and 42,799 clinical QA pairs across several CMR sequences. They propose Cardiac Anatomy-Routed Attention (CARA), which uses predicted anatomical priors to guide decoder attention, and combine it with grounding pretraining to form CARA-VL. Experiments report strengths in clinical assessment and regional localisation, with promising generalization to an external cohort.
