In briefPredicting Future OOD Degradation from Source-Side Training Dynamics
SummaryThis paper studies whether persistent out-of-distribution (OOD) performance degradation can be forecast before it is observed, using only signals from source-side training. In a controlled shortcut-learning setup, a simple logistic regression predictor shows a clear forward-looking signal, while training time alone does not. Temporal summaries of these source-side quantities are more informative than their current values. Confidence and entropy dynamics also retain substantial predictive information when transferred from a CNN to an MLP without further training, offering a proof of principle for early warning signals of future OOD failure.
In briefHRIL: Learning Multimodal Synergy Using Higher-Order Tensor Modeling
SummaryThis paper addresses how multimodal models capture synergy, meaning task-relevant information that appears only when several modalities are combined. The authors propose HRIL, which builds a cross-moment tensor over modality embeddings, applies Tucker decomposition to get a core tensor, and adds a regularizer that prevents energy concentration. Experiments on a controlled synergy task and real-world benchmarks show consistent gains over existing multimodal contrastive methods, especially on tasks dominated by synergistic interactions.
In briefGeoReform: A Reflective Framework That Refines How Geometry Facts Are Written for Multimodal Models
SummaryThe paper observes that multimodal large language models often miss geometric relations in diagrams, and that converting them to text can introduce redundant or ambiguous content that misleads reasoning. It proposes GeoReform, which treats formalization as an optimizable policy, diagnosing failed rollouts and revising how entities, relations, and constraints are selected and organized. On Geometry3K, the method raises Qwen3VL-2B accuracy from 42.0% to 56.0%.
In briefFrom Long Text to Predictive Features: LLM-Guided Block-Wise Feature Engineering via Executable Program Search
SummaryThis paper proposes LLM-BlockFE, a framework that uses an LLM offline to turn long unstructured text into executable feature programs, so no LLM calls are needed during online inference. It builds programs by appending code blocks and scores candidate features with a downstream model, using block-level rollback to avoid suboptimal greedy search. Across two public and two private datasets it improves AUC, and across five deployed financial risk-control applications it improves KS.
In briefGenIA: Generative 3D reconstruction aligned to test-time inputs
SummaryThis paper presents GenIA, a method that aligns a generative 3D foundation model (SAM3D) with the observed images at test time, without retraining it. It derives object translation and scale from geometry to improve pose, and aligns appearance using visibility-biased attention, cross-view fusion, and differentiable rendering guidance. The authors report better pose prediction and object reconstruction than recent optimization-based, per-frame, and video-to-4D methods on synthetic and real benchmarks.
Stefano Esposito, Naama Pearl, Polina Karpikova et al.
In briefARC: A Reasoning Recipe for Improving Robot Foundation Models
SummaryThis paper presents ARC, a reasoning recipe that improves the zero-shot task performance of existing robot foundation models without new robot demonstrations or foundation-scale training. ARC combines reasoning traces grounded in the robot's next action, an automatic labeling pipeline that builds ARC-Trace-DROID from DROID, and a method for adapting pretrained models to use these traces for control. The authors report new state-of-the-art results on RoboLab-120 and MolmoSpaces, and an 82.2 percentage point task success gain for π0.5 on real robots.]
Gokul Puthumanaillam, Tao Sun, Elie Aljalbout et al.
In briefWorldAlign: A decoupled reward framework for more consistent 4D video generation
SummaryWorldAlign is a decoupled 4D reward framework that improves world consistency in generated videos, covering both static scenes and moving subjects. Static regions are aligned with a geometric world prior using masked reprojection and a camera-motion reward, while dynamic subjects are evaluated by a vision-language model using sample-specific checklists. Tested on the Wan2.1 and Wan2.2 image-to-video generators, it reportedly improves static and dynamic consistency without suppressing overall or subject motion.
In briefMarformer: a Transformer that predicts distributions of missing data
SummaryThe paper introduces Marformer, a BERT-like Transformer that directly predicts conditional marginal distributions of missing variables given any set of observed values. Unlike generative methods, it does not model the full joint distribution, needs no domain knowledge of the data-generating process, and makes predictions in a single forward pass. Evaluated on three synthetic domains and one real annotation dataset, it matches or outperforms classical missing-data methods and the evaluated generative baselines, while being substantially faster.
In briefLatent Core Tokenizer: Compressing Text Without Losing Meaning
SummaryThis paper proposes the Latent Core Tokenizer (LCT), a language-agnostic method that first finds reusable linguistic units using Minimum Description Length, entropy-based boundaries, and morphotactic constraints, then builds a shared vocabulary. Across 104 languages with a 200K-token vocabulary, LCT yields lower fertility and higher MorphScore than BPE, Unigram, and parity-aware BPE, with comparable cross-lingual tokenization cost disparity. On four multilingual downstream benchmarks, it improves aggregate score by 1.48, 1.83, and 2.00 points over these baselines. The authors conclude that compression alone does not predict representation quality.
Felermino D. M. A. Ali, Millicent Ochieng, Ogbemi Ekwejunor-Etchie et al.
In briefOnTrack: real-time monitoring of LLM agent runs by comparing them to successful ones
SummaryThe paper proposes OnTrack, a streaming monitor that compares each step and dependency of an LLM agent against recorded successful runs, alerting users or blocking the agent in about a millisecond per step. It is evaluated on SWE-bench trajectories, where using only the first 8 steps it ranks failing runs below succeeding ones better than content similarity methods (+0.057 AUROC). With an abort policy, about 18% of compute otherwise spent on failing runs is saved, and 83% of interrupted runs were indeed heading to failure.
Babak Barazandeh, Connor Swanson, Chinmay Kulkarni et al.
In briefAgentGarten: Code-Built Virtual Worlds for Evolving AI Agents
SummaryThis paper presents AgentGarten, a framework that combines simulators, game engines, and a shared neural renderer to build real-time interactive environments. Agents perceive the world through visual observations and distill each round of experience into playbooks that later agents inherit. The authors report that agents learned from just 4 rounds, compared with millions for a conventional reinforcement learning counterpart.
In briefLearning Koopman embeddings from data with kernel dictionaries via bilevel optimization
SummaryThis paper introduces EDMD-kDL, a kernel-based method that learns finite-dimensional Koopman embeddings directly from data. It combines collocation and bilevel optimization to jointly learn a kernel dictionary and the corresponding Koopman approximation. Tested on global sea-surface-temperature forecasting and video data, it performs comparably to or better than neural-network-based approaches. Because kernel matrix size depends on collocation points rather than dataset size, it scales to large datasets.
Joel-Pascal Ntwali N'konzi, Feliks N\"{u}ske, Stefan Klus
In briefEmbodied Turing Machines: code-only robot policies that can improve themselves
SummaryThe paper proposes Code-Only-as-Policy (COAP), which tracks robot, environment and task state from camera images and proprioception in code and makes all decisions from that state, with no VLM or VLA at run time. The authors argue that explicit state, controllable execution and extensibility make it suited to recursive self-improvement, with coding agents developing a shared library. On RoboDojo's 42 bimanual tasks, the library reaches a 70.24% success rate without a model at test time.
In briefWhich Skills to Distill? SGUID selects a compact skill bank for stable model-skill co-evolution.
SummaryThe paper finds that fewer than 25% of retrieved skills give useful distillation signals in on-policy distillation with skill-conditioned teachers. It proposes SGUID, which keeps only skills that consistently yield effective learning signals during training. Across four Olmo and Qwen models, distilling 6 selected skills matches or exceeds full-bank distillation, which can be up to 11 times larger.
In briefHANS: A Handwritten Answer Sheet Dataset for Noisy Hybrid Document Parsing
SummaryThis paper introduces HANS, a dataset built from real student answer sheets containing math expressions, text, hand-drawn tables, and noise such as corrections and strikethroughs, with fine-grained annotations. It also proposes NA-GOT, an end-to-end framework that suppresses noise in two stages using a feature-level module and a noise-aware attention mechanism in decoding. Experiments show HANS is challenging for existing methods, while NA-GOT improves accuracy and stability in answer process recognition. The dataset will be released publicly upon publication.
In briefClosing the horizon gap in adversarial MDP policy optimization
SummaryThis paper studies policy optimization for online episodic tabular Markov decision processes with adversarial losses and bandit feedback. The authors use regularized Q-functions to jointly control the stability of local policy updates across all state-action pairs, reducing a horizon-dependent gap in regret bounds relative to occupancy-measure-based methods. The algorithm achieves high-probability regret bounds for both known and unknown transitions, with the unknown-transition bound matching the best known result. The approach is also extended to adversarial linear-mixture MDPs with the same horizon improvement.
In briefCited but Not Consulted: Testing Whether Legal AI Verdicts Depend on Their Stated Authorities
SummaryThe paper tests whether language models' verdicts actually depend on the statutes or precedents they cite. Holding case facts fixed, the authors swap the cited authority for an unrelated one and decode verdict changes from hidden states. Models name the correct authority in most generations, but the verdict changes far less consistently, varying widely across four benchmarks.
WhySwapping cited legal authorities and tracking hidden states shows explanations often fail to reflect verdict dependence, supported by seven models and four benchmarks.
In briefAccurate but Not Humble: Checking whether AI agents admit uncertainty when evidence conflicts with their beliefs
SummaryThis paper examines how LLM agents respond when retrieved evidence contradicts their prior knowledge, asking whether they revise answers, admit uncertainty, or stick with wrong conclusions. The authors propose measuring epistemic humility through three behaviors: Identify, Solve, and Escalate. They test four agents under controlled and naturally occurring knowledge conflicts. They find that higher task accuracy does not necessarily mean greater epistemic humility.
Kaiser Sun, Bernal Jimenez Gutierrez, Hongjun Liu et al.
In briefSample thresholds for learning functions on the Boolean cube under noise
SummaryThe paper studies Gaussian regression under squared L2 loss in a known m-dimensional subspace of low-degree functions on the Boolean cube. It gives a sharp sample threshold for minimax error, with an exponent set by a binary-entropy-based function, and sharpens the Polyanskiy–Samorodnitsky uncertainty principle. The results imply that with noise the parametric rate can need samples growing like 4^k, whereas noiseless identification needs only O((m+t)2^k).
In briefDistilling Routed 3D Privileged Information to Improve Spatial Reasoning in Vision-Language Models
SummaryVision-language models struggle with spatial reasoning because RGB images do not directly provide geometric information. The paper proposes GPD, which gives a teacher model depth, semantic, and bird's-eye-view cues as compact text and applies a privileged KL loss only to incorrect trajectories to augment GRPO training. The deployed model uses RGB input only.
In briefMine or Yours? Language Models Linked by Memory Claim Each Other's Memories
SummaryResearchers used two copies of one language model, each with a private rule and code word, where the receiving copy could also read its partner's internal memory. When the partner's memory was weighted double, the receiver named its partner's assigned rule in 97% of answers. At equal weight, it took the partner's rule as its own in about half of its answers. The authors argue that systems letting models exchange internal states must test ownership as well as transfer.
In briefSplitJEPA: Learning invariant and variant latent worlds without reconstruction
SummaryThis paper introduces SplitJEPA, a joint embedding predictive architecture that recovers a latent state and its invariant and variant organization directly in representation space, with no reconstruction. Under stationary Gaussian predictive dynamics and a full-rank variation condition, the authors prove the invariant and variant subspaces are identified up to independent block-wise isometries, without an observation decoder. Experiments on synthetic nonlinear systems and robotic manipulation tasks support the theory.
In briefOvercoming Prior Barriers: Fine-Tuning LLMs on Rare Concepts
SummaryThis paper studies how supervised fine-tuning of large language models is affected by uneven pretrained support for different concepts. The authors define a "prior barrier" and find it follows a long-tail distribution, making rare concepts harder to learn. They propose PASS, an adaptive instruction selection method under a limited budget, which reportedly outperforms seven existing methods across four settings.
In briefA benchmark and agent framework for reasoning-based visual editing
SummaryThe paper notes that multimodal models still struggle with complex instruction following, appearance consistency, and flexible inputs in visual editing. It introduces RISEBench++, a hierarchical benchmark covering six reasoning dimensions, 12 subcategories and 65 task types, with 1000 human-annotated cases in English and Chinese. It also presents RISE-Agent, a training-free agentic framework, and reports that the strongest of 58 evaluated approaches reaches only 56.6% accuracy.
In briefCan AI Agents Learn Their Way to the Top? A Study of Learning Game Strategies Over Long Competitions
SummaryThis paper introduces Adversarial Heuristic Learning, where AI agents refine game policies and supporting software from experience while model weights stay fixed. The authors present AAArena, a benchmark of 12 adversarial games and 1,920 archived human programs, evaluated under a protocol modeled on real competitions. The best configuration earns 6 gold medals but does not top the remaining 6 human ladders, and performance is weaker in games with more complex rules.
In briefAmbient Discrete Diffusion: Training Discrete Diffusion Models With Scarce Data Using Out-of-Domain Data
SummaryThis paper presents RefineMix, a framework for training discrete diffusion models when in-domain data is scarce. It uses out-of-distribution data at selected diffusion times to improve generalization without biasing the sampling distribution. Across five domain-shift settings, RefineMix matches or outperforms in-domain finetuning and data mixing. For protein sequence generation, finetuning on just 197 in-domain examples nearly doubles the fraction of proteins that are novel, foldable, and in-family.
Julian Kleutgens, Mauricio Tec, Claudio Battiloro et al.
In briefVFold: A Symmetry-Aware Method to Compress the Value Cache in Language Models
SummaryLanguage models store key-value cache states to speed up text generation, but this cache can use most of the memory when processing long inputs. This paper proposes a value cache merging method that exploits similarities between layers, reducing memory use without requiring changes to the model architecture. The authors report that it avoids significant performance loss during decoding. They also state that it can be combined with existing techniques such as quantization or key cache pruning to reach compression ratios neither achieves alone.
In briefContiLNN: Liquid Neural Networks for Fixing Slice Gaps in Medical Image Restoration
SummaryThis paper introduces ContiLNN, which adds bidirectional closed-form continuous-time modules to 2D restoration backbones to use continuity between neighboring medical image slices. The slice-index intervals adjust the model's gates, so it adapts to sampling variations without numerical ODE integration. Across CT denoising, MRI super-resolution, and reduced-count PET restoration, the authors report higher mean PSNR and lower RMSE than Restore-RWKV.
In briefasdex: A JAX toolkit for automatic sparse differentiation of Jacobians and Hessians
SummaryThis paper presents asdex, a standalone automatic sparse differentiation toolkit for the JAX ecosystem. It exploits sparsity in Jacobian and Hessian matrices through pattern detection, graph coloring, compressed differentiation, and decompression to reduce the number of AD passes. Its asdex.jacobian and asdex.hessian functions serve as sparse drop-in replacements for jax.jacobian and jax.hessian.
In briefRiCo: Neural simulation of rigid-body contacts using local contact reasoning
SummaryThis paper introduces RiCo, a method that represents interactions between rigid objects through sparse neighborhoods of contact surface points, so only nearby surfaces exchange contact information. On the MOVi benchmark, RiCo reduces 100-frame position and orientation errors by 31-35% and about 38% compared with baselines, and reports penetration-time and mean-depth differences of 11.0% and 2.22 mm. It generalizes zero-shot from small training scenes to scenes with 270 objects, and preliminary real-world multi-ball collision experiments provide early evidence of sim-to-real transfer.
In briefCombining small randomized trials with large observational data to estimate treatment effects more precisely
SummaryRandomized trials give unconfounded treatment effects but are often small, while observational studies are large but may be confounded. The paper proposes a framework that fuses the two without special assumptions on the observational data, keeping the unbiasedness of trial-based estimation while borrowing power from the large dataset to improve precision. It builds an average treatment effect estimator, AIPW-Fusion, with closed-form weights and confidence intervals, plus two conditional effect learners, DR-Fusion and R-Fusion. Experiments are reported to support these findings.
In briefOptimal-complexity method for nonsmooth nonconvex composite optimization
SummaryThis paper studies stochastic nonsmooth nonconvex composite optimization, where the objective is a nonconvex Lipschitz function plus a convex regularizer, accessed via stochastic gradients or function values. The authors extend the online-to-nonconvex conversion framework by using new learner losses that include the regularizer itself, with a mirror descent variant achieving low regret. The algorithm finds a Goldstein-type stationary point with O(δ^{-1}ε^{-3}) stochastic gradient queries or O(dδ^{-1}ε^{-3}) function-value queries. These rates match the optimal rates for noncomposite problems.
WhyExtends online-to-nonconvex conversion to composite nonsmooth nonconvex optimization, matching optimal oracle complexity of noncomposite case with rigorous analysis.
In briefSparseDecoding: Decoding-aware pruning to make LLM inference more accurate and faster
SummaryLLM decoding is slowed by memory reads, and pruning can reduce them, but standard pruning calibrates on natural text rather than the model's own generated tokens. SparseDecoding builds its calibration matrices from layer-wise activations collected during dense-model autoregressive generation, excluding prefill, to match the decoding distribution. It also adds an optimized N:M sparse matrix-vector kernel with bitmask indexing. On Llama-3.1-8B, Llama-3.3-70B, and Qwen3-14B/32B, it outperforms fixed-text calibration on long-form generation and reaches up to 1.48x end-to-end decoding speedup on A100 GPUs.
WhyProposes decoding-aware pruning and an N:M sparse matrix-vector kernel, validated on several LLMs with up to 1.48x speedup.
In briefPrior or Feedback? What an LLM Uses When Adapting Neural Operators
SummaryThis paper studies whether an LLM agent choosing fine-tuning settings for neural operators under a limited trial budget relies only on its initial task context or adapts to experimental feedback. In nearly every matched comparison, the LLM reached lower held-out test nRMSE than random search and Bayesian optimisation. Controlled interventions show its first configuration already has a useful initial bias, and that reassigning validation scores changes its next proposal, indicating a mix of task-dependent prior and sensitivity to feedback.
In briefSpatial patterns from agents learning to move in public goods games
SummaryThe paper studies how cooperators and defectors that each learn movement policies with tabular Q-learning form spatial patterns in a public goods setting. Cooperators cluster around resource peaks, and their learning rates combined with defectors' rates shape collective welfare. Under the tested learning-rate conditions, mean welfare falls below random movement because increased crowding outweighs resource gains.
In briefMaking generated character motion more exaggerated and expressive via training and inference guidance
SummaryThis paper addresses motion generative models that produce physically plausible but flat character animation by incorporating the animation principle of exaggeration. The method has two stages: supervised fine-tuning of pre-trained text-to-motion models on a curated exaggeration dataset during training, and, at inference, a formulation of exaggeration based on dynamic movement primitives used as guidance for diffusion and flow-matching models without extra training. The authors report that compared with three strong motion generation models, their methods produce more exaggerated and expressive motions while preserving neutral reference intent and physical plausibility.
Amirhossein Zamani, Arianna Rampini, and Bruno Roy
In briefVerdicts That Ignore the Rule: Auditing How LLM Compliance Systems Respond to Regulatory Rules
SummaryThis paper tests whether large language model compliance verdicts actually depend on the regulatory rule supplied to them. Across five models and 20 regulatory and platform-policy domains, the authors deleted, swapped, or negated the governing rule and measured whether verdicts or internal representations changed. Verdicts were often invariant to substantial rule perturbations, and the guard model was the least rule-sensitive and least accurate, at 51% versus 90-92% for general-purpose models.
Saisab Sadhu, Aadit Sengupta, Vinay kumar Sankarapu et al.
In briefIs In-Domain Training Enough? A Multi-Agent Method for Fine-Grained Industrial Anomaly Understanding
SummaryThe paper argues that a single multimodal large language model struggles to handle detection, localization, description, and reasoning for industrial anomalies, and that in-domain training does not close this gap. On the MMAD benchmark, trained specialists reach at most 75.5% defect localization accuracy, versus 92.3% for human experts. The authors propose SiGMA, a multi-agent framework combining heterogeneous MLLMs with a visual defect expert and a label-free reliability controller, which reaches 85.2% average accuracy on MMAD.}
In briefBudgetPix: Compute-Adaptive Tokenization for Image Generation Models
SummaryStandard image generation models spend the same computation on every image patch, which cannot adapt to different resource limits at inference time. BudgetPix dynamically allocates compute based on visual complexity and spatial layout, so one checkpoint can run across a range of compute budgets. On text-to-image generation, the paper reports matching MiniT2I-L at 512² and PixelDiT at 1024² using 25% of the original compute. On class-conditional generation with a MeanFlow backbone, it uses 60% of the compute with a 0.8-point FID increase.
In briefA structural theory of how minds represent problems, using context, invariance, and a knowledge space
SummaryThis paper proposes a minimal structural framework that makes representational operations for problem solving explicit, including context formation, invariance recognition, representative selection, abstraction, and procedural reuse. A context is formalized as a partition of a subset of a state space, and the framework is realized as a Knowledge Space made of a Concept Graph and a Procedure Graph. Using simple examples and a finite weak-solver demonstration, the authors show that good representational organization can simplify admissible regularities. The contribution is structural rather than algorithmic.<br>
In briefFrom What to Which: Decoding How Frozen Multimodal Models Locate Described Objects
SummaryThis paper asks whether frozen multimodal large language models encode information that identifies which specific instance a generated modifier, such as color or position, refers to. The authors introduce OTTER, a lightweight supervised probe that uses optimal transport to align generated tokens with image regions and produce compact grounding maps. The results report that instance information can be decoded from modifier tokens, most clearly for spatial terms, and that the grounding holds under context perturbations and extends to free generation and cross-dataset transfer.
Barbara Toniella Corradini (AI for Good (AIGO), Istituto Italiano di Tecnologia, Italy) et al.
In briefLooking inside LLMs: small-world wiring as a marker of reasoning ability
SummaryThe paper builds functional graphs from attention-head activation similarities and measures a small-world index (SWI) of their connectivity. A higher SWI consistently correlates with better fluid reasoning performance across models and training checkpoints. The authors also propose Small-World Allocation (SWA), a pruning method guided by core and bridge scores of heads, which better preserves performance across six LLMs and reduces WikiText perplexity by up to 20%.
In briefLearning logic rules from data using gradient-guided language models
SummaryThe paper presents grasp, a framework for learning probabilistic logic programs from data. It treats relational structure learning as functional gradient boosting, where the weak learner is a first-order rule and an LLM proposes candidate rules in place of exhaustive symbolic search. Evaluated on four relational benchmarks, including Tox21, mutagenesis, and Cora citation matching, it outperforms symbolic, neural, and LLM-based baselines while producing interpretable weighted rule ensembles.
Saurabh Mathur, Sahil Sidheekh, Bhavan Vasu et al.
In briefMulti-Agent Egocentric World Model: Synchronized First-Person Video Generation for Fine-Grained Interaction
SummaryThis paper proposes ME-World, a multi-agent egocentric world model that generates synchronized first-person video streams for several agents interacting through fine-grained actions in a shared environment. It jointly denoises the streams in one token sequence, conditions each on all agents' target-view poses, and uses shared environment memory to keep generation grounded. The authors train and evaluate on real and synthetic multi-agent data and report improvements in shared-world consistency, action control, identity preservation, and video quality over existing methods.
In briefMachine learning in high-energy nuclear physics, moving from pattern recognition to physics-integrated discovery.
SummaryThis mini-review surveys recent progress in applying machine learning to high-energy nuclear physics. It describes a shift from early event classification and surrogate models toward workflows that directly integrate physical knowledge. Applications covered include heavy-ion collisions, neutron-star physics, lattice field theory, and holographic or continuum QCD. The paper emphasizes how physical constraints are imposed and how uncertainty quantification and validation determine whether AI-assisted results support reliable physics conclusions.
In briefOne Word Opens the Gate: Attacks on AI Model Guardrails
SummaryThis paper studies typed decision models used as guardrails in AI agent systems to decide whether a tool call or message is allowed. The authors evaluate seven open-weight models and find allow-or-block accuracy between 36% and 72%, against a 50% chance level. Adding a few lines of unrelated server log text, or giving the permissive option a misleading name, sharply raises the rate of wrongly allowed prohibited actions, and every defense tested was defeated. The authors conclude these models should not be the component that makes the final decision.
WhyProposes an option-label channel attack exposing guardrail failures; systematic evaluation across seven models and multiple defenses shows each is defeated, with strong empirical support.
In briefSlot3R: Set-associative spatial memory for streaming 3D reconstruction
SummaryThis paper addresses how spatial memory should store observations during streaming 3D reconstruction. It argues that Point3R's use of spatial proximity both to match new observations and to decide fusion merges distinct states, which can erase complementary evidence. Slot3R is a training-free retrofit that lets multiple states share one spatial address while keeping the pretrained Point3R backbone frozen. It reports lower point-cloud accuracy error and trajectory error than Point3R on the evaluated benchmarks, and it completes 600 to 1000 sampled frames where Point3R and InfiniteVGGT run out of memory at 800 frames and beyond.
In briefARGUS: An evidence-constrained AI agent for interpreting noncoding genetic variants
SummaryThe paper introduces ARGUS, a framework for interpreting single nucleotide variants in noncoding regulatory regions, keeping deterministic biological computation separate from LLM-based reasoning. It wraps 458 DNABERT transcription factor binding models in an investigation loop that chooses evidence sources based on uncertainty. A case study on variant rs6983267 at the 8q24 cancer risk locus shows different investigation paths and verdicts or abstentions for four transcription factors, using real ADASTRA, JASPAR, and ENCODE cCRE data. The authors report that an LLM planner reaches the same verdicts as a fixed-priority planner with fewer tool calls.
In briefHarnessSQL: Training SQL agents inside their execution harness for realistic databases
SummaryThis paper proposes HarnessSQL, a post-training framework that trains SQL agents within the same execution harness used at inference, so they learn multi-turn interaction with live databases rather than mapping questions to static queries. It builds isolated executable database environments with hidden execution oracles, keeps only verified trajectories for supervised fine-tuning, and then applies reinforcement learning with execution rewards. On Spider 2.0-SQLite, execution accuracy rises from 15.5% to 45.2% for Qwen3-8B and from 22.2% to 54.8% for Qwen3-14B, with transfer to BIRD-Interact and LiveSQLBench.
In briefDVD: Faster, Compact Vector Decoding for Multimodal AI Perception
SummaryThis paper proposes DVD, a dynamic vector decoding method that converts 2D bounding boxes, 2D masks, and 3D bounding boxes into 1D vector sequences and then into compact discrete tokens. A lightweight de-tokenizer decodes the model's output tokens back into the original 2D and 3D representations. Experiments on benchmarks including RefCOCO, SUN-RGBD, KITTI, and nuScenes reportedly show better performance while significantly reducing token overhead and inference latency.
In briefAdaptLSTM: Efficient online learning that adapts cloud workload forecasts when data patterns shift
SummaryThis paper proposes AdaptLSTM, a framework for forecasting cloud service workloads so resources can scale elastically. It detects distribution drift using validation-calibrated thresholds and performs selective, targeted model updates instead of updating at every step. On the Alibaba Machine Trace and Container Trace, it recovers a large share of naive online learning's gains at a fraction of the compute cost. The approach is reported to work across LSTM, GRU, and Transformer backbones.
In briefSyn-Omni: A Method for Structured Specialization and Staged Collaboration in Omnimodal Embeddings
SummaryThis paper proposes Syn-Omni, a framework for adapting omnimodal embedding models in a structured way. Its OME-LoRA component separates a shared LoRA path for universal semantics from modality-expert paths for modality-specific features. Its Progressive Synergy Routing lets experts first form modality-specific priors and then gradually interact with other modality experts. The authors report that it outperforms omnimodal baselines across 81 tasks covering image, video, audio, and audiovisual data.
In briefHow Much Audio Remains in Audio Encoder Embeddings: An Inversion Audit
SummaryThis paper examines how much source audio information remains in frozen representations from pretrained audio encoders. Using a shared Stable Audio Open latent diffusion decoder, the authors reconstruct five-second, 44.1-kHz stereo music from VGGish, ConvNeXt, CLAP, and EnCodec embeddings, evaluated on the Million Song Dataset. Reconstructability differs clearly across encoder families, and exposing finer temporal or spectral structure improves recovery. Even compressed task-oriented embeddings support reconstructions that preserve measurable source specificity and high-level musical content.
In briefReal-time motion planning in moving hazard zones: classical planners versus learned policies
SummaryThis paper compares classical motion planners and learning-based methods on a shared benchmark rather than proposing a new planner. The test environments are planar domains with rotating sprinkler-like hazards that create time-varying forbidden sectors. Classical planners succeed almost perfectly with higher-quality paths in deterministic settings, but under stochastic obstacle dynamics their online search is highly sensitive to computation budget. PPO-based policies trained on the same scenarios outperform them in latency, success rate, and path quality in these stochastic regimes. The authors conclude that uncertainty in obstacle evolution, more than partial observability, determines which paradigm works better.
In briefEgoVoice: Proactive Spoken Assistance From First-Person Video and Audio
SummaryThis paper introduces EgoVoice, a framework for training and evaluating wearable AR assistants that decide when to speak and what to say based on continuous first-person video and audio. The authors build clean audio streams from HoloAssist recordings of real instructors and fine-tune an omni-modal LLM, further improved with direct preference optimization. Experiments report that EgoVoice improves intervention timing, content relevance, and human preference over the zero-shot backbone, while existing systems rarely give well-timed proactive guidance.
In briefBatch before lifting: scalable deep learning on large graphs using higher-order structures
SummaryThis paper introduces Cluster-TNN, a framework that makes topological deep learning practical on large, dense graphs such as Reddit by lifting graph structures locally rather than globally. It partitions the input graph, samples clusters at runtime, and builds induced subgraphs as mini-batches on which existing topological neural networks run. Across 21 matched comparisons with full-graph execution, peak GPU memory fell in every configuration, by 83.2% on average, while predictive performance remained competitive.
David Leko, Luka Beni\'c, Guillermo Bern\'ardez et al.
In briefRIFT: A Deterministic Tree-Based Method for Anomaly Detection
SummaryThis paper introduces RIFT (Relative Isolation From Trees), an unsupervised anomaly detection method. It builds a minimum spanning tree and scores each point by summing the apparent sizes of tree edges as seen from that point. For one-dimensional data, RIFT exactly recovers the closed-form infinite-limit score of Isolation Forest, and in higher dimensions it is parameter-free, deterministic, and robust to varying density and clustered anomalies. Experiments show accuracy comparable to Isolation Forest, while the proposed ensemble variant has significantly lower variance across random seeds.
In briefNativeScope: Relation-scoped retrieval using native data structure and a correct anchor
SummaryThe paper proposes NativeScope, a retrieval method that first narrows the search to a scope defined by a known anchor and relation, then ranks only the chunks within that scope. It uses structure already stored in data, such as section membership, session boundaries, and native order. On 200 controlled document and memory records, it improved native-unit recall over instance-wide Dense RAG by 42.75 points for documents and 22.00 points for memories. With automatically chosen Top-1 anchors, memory recall dropped to 35.50 percent, showing dependence on reliable anchor localization.
In briefTokenRouter: A Serving System for Efficient Token-Level LLM Routing
SummaryThis paper presents TokenRouter, a serving system for large language model inference that routes individual tokens across different models. Developers describe routing logic from a single request's perspective, while the runtime launches a subserver per model and dispatches requests asynchronously. Each subserver uses a delayed-batching scheduler whose hyperparameters come from a throughput model. The authors report 2.01 to 64.15 times higher decoding throughput than existing systems across various routing algorithms, workloads, and model pairs.
WhyProposes request-centric programming with model-centric execution and a throughput model for scheduling, reporting 2.01 to 64.15x decoding throughput gains across routing algorithms and model pairs.
In briefAdaCast: input-specific parameter generation for adaptive time series forecasting
SummaryThis paper addresses the limitation that existing adaptation methods for time-series foundation models apply one fixed set of parameter updates to every input. The authors propose AdaCast, which uses a generator to produce input-specific low-rank parameter updates for a frozen pretrained model during training and inference. Across six public benchmarks, the authors report that AdaCast outperforms a static adaptation baseline in-domain and improves zero-shot generalization to held-out datasets.
In briefLa-Ribo: Jointly designing RNA sequence and 3D structure with flow matching
SummaryThis paper introduces La-Ribo, a generative framework that designs RNA nucleotide sequences and three-dimensional structures together using geometry-latent flow matching. The authors built a quality-controlled corpus of 168,561 RNA structures, including 10,631 MSA-supported structures generated in this work. The reported results show La-Ribo improves designability and codesignability over the evaluated baselines across sampling budgets and two refolding models.
In briefLanguage Models as Research World Models: Predicting Experiment Outcomes with LLMs
SummaryThis paper studies language models as Research World Models (RWMs) that predict the outcomes of candidate experiments in AI research. The evaluation uses over 2,600 experimental records from nine research environments covering pretraining, post-training, and inference. Research knowledge from real experiments improves predictions for unseen interventions within the same environment and can be reused across environments.
In briefDelay-aware audit of test-time adaptation for time-series forecasting
SummaryThis paper audits test-time adaptation (TTA) methods for time-series forecasting, where labels for an H-step forecast only arrive H steps later. The authors build a leakage-free harness that releases each label only after its delay and test four recent methods using their own code and checkpoints. Under causal delays, several published methods are significantly worse than the frozen model on most datasets, while simple closed-form correctors often perform better at much lower cost.
In briefFrom Prompting to Composing: A Spatial Canvas Interface for Poster Generation
SummaryThe paper argues that text prompts poorly capture the two-dimensional layout intent needed for posters. It introduces a Spatial Canvas Interface where users place elements using semantic, identity, text, and pixel bindings, plus text specifications for elements and global appearance. The authors build Compo, an adaptation of a pretrained image editing model, which supports both direct canvas construction and an agentic mode that plans a canvas from a high-level request.
In briefVibeEdit: Image Editing with Drawn Canvas Instructions
SummaryThis paper presents an image editing interface where users place spatial marks and short notes on an image to indicate both where and what to change. The editor, VibeEdit, carries out object addition, removal, replacement, attribute changes, and movement from these canvas instructions without a text prompt. On a 419-case benchmark focused on choosing among similar objects, it scores 79.9 on a VLM rubric and 32.8 dB outside-region PSNR, versus 67.4 and 24.0 dB for the best text-instructed baseline, FireRed.
In briefIntuition with Evidence: Adaptive Dual Memory for Multimodal Agents
SummaryThis paper addresses long-term memory for multimodal agents by proposing a dual-memory framework. It combines a model-native latent memory, used as an intuitive prior, with a multi-indexed episodic memory that reconstructs exact evidence through retrieval-conditioned joint alignment. At inference, a training-free routing mechanism based on Jensen-Shannon Divergence sensitivity and confidence proxies adaptively composes the two memories. The authors report state-of-the-art results across multimodal benchmarks, outperforming single-paradigm baselines overall (+4.67% over vanilla episodic memory).
In briefSemi-automatic stride-independent patching for deep learning image segmentation
SummaryThis paper proposes semi-automatic stride-independent patching (SSP), an alternative to automatic stride-dependent patching. SSP uses user or expert input to place predefined patches over objects of interest. Three patch-based datasets were built and DeepLabV3+ models with three backbones were trained on them. SSP generally achieved higher segmentation scores and the shortest training time, while non-overlapping patching did not generally yield better segmentation performance despite greater spatial coverage.
In briefEqual-budget comparison of Monte Carlo Tree Search and randomized-restart list scheduling for DAG workflows
SummaryThis paper compares Monte Carlo Tree Search (MCTS) with a simpler randomized-restart list scheduler, both given the same number of evaluated schedules, for scheduling DAG workflows. Under a preregistered protocol at the 2048-schedule budget, MCTS performed 2.88% worse than the restarts. The authors attribute the gap mainly to the studied selection rule, and textbook UCT removed most of it. Results are limited to a controlled synthetic domain with budgets up to 2048.
In briefISBO: Scalable Spatio-Temporal Bayesian Optimization Using Log-Gaussian Cox Process Models and INLA-SPDE
SummaryThis paper introduces ISBO, the first scalable Bayesian optimization framework for spatio-temporal data. It models log-intensity with a Log-Gaussian Cox Process and performs fast inference using the INLA-SPDE approach. Experiments on synthetic and real-world datasets show accurate peak discovery and intensity recovery, with substantial speedups over an RKHS-based baseline.
In briefDiffuPlex: Faster Full-Duplex Spoken Dialog Models Using Rolling Masked Diffusion
SummaryFull-duplex spoken dialog models can listen and speak at the same time, but fine-grained versions must run their backbone autoregressively at every interaction frame. DiffuPlex uses rolling masked diffusion to predict several future user and assistant frames in one backbone call, consuming only the confident prefix while interaction keeps its original frame rate. Across full-duplex and spoken-language evaluations, it substantially reduces sequential backbone computation while largely preserving interaction behavior and general capability. The two inference variants achieve deployment wall-clock speedups of 1.46x and 1.59x..
In briefHow many samples does a Bayesian neural network need? Adaptive inference time with statistical guarantees
SummaryThe paper notes that Bayesian neural network predictions are often approximated with a fixed number of Monte Carlo samples, without controlling the error from that finite sampling. The authors propose using confidence sequences to decide dynamically how many samples are needed while keeping statistical guarantees. Sampling stops once the relevant decision, such as the most likely class or a probability threshold, can be made with the desired guarantee. Experiments show the method gives more samples to ambiguous inputs and fewer to easy ones, reducing overall latency compared with a fixed Monte Carlo budget while keeping decisions reliable.
In briefSciTBERT: Time-consistent language models for science and technology text
SummaryThis paper introduces SciTBERT, a family of BERT-based language models trained on scientific papers, patents, and educational web text with data cutoffs from 2013 to 2025. The authors also create SciTBERT-CI by post-training on paper and patent citations in a chronologically consistent way. The models generally outperform earlier domain-specific encoders on science and technology tasks. The paper additionally presents PatRepEval, a benchmark of patent-related text embedding tasks at the science-technology interface.
WhyIntroduces chronologically consistent BERT models and a patent benchmark, validated across many classification, regression, and retrieval tasks against prior domain encoders.
In briefSteerablePlex: Can Full-Duplex Speech Models Be Controlled?
SummaryFull-duplex speech models can listen and speak at once, but become harder to control as conversations grow, which can make them unreliable as user simulators. The authors introduce SimIF-Bench, a benchmark testing whether a model stays within a scenario and completes goals in order, and find current open-source full-duplex models struggle. They train SteerablePlex with a GDPO-based recipe to follow text instructions mid-conversation while keeping turn-taking, and pair it with an asynchronous backend model that issues instructions. The paper reports it follows multi-stage constraints more reliably than existing open-source models and GPT-Realtime.
Haolong Zheng, Maike Z\"ufle, Dominik Mach\'a\v{c}ek et al.
In briefHow Data Scales in Agentic Reinforcement Learning: Laws and Synthesis Strategies
SummaryThis paper studies how training data should scale in agentic reinforcement learning, splitting it into task, environment, and trajectory axes plus a reward-density control. The authors fit per-axis scaling laws across several large language models and domains, finding that environment scales most strongly, followed by task, then trajectory. Following the derived compute-optimal allocation recipe reportedly yields up to 1.7× compute efficiency over the common practice of scaling trajectories.
In briefFast change detection using diffusion-integrated scores, a training-free method for spotting distribution shifts
SummaryThis paper proposes DI-SCUSUM, a training-free detector for quickly flagging when a data distribution changes. It adds Gaussian noise to samples to form smooth density estimates, computes Hyvärinen scores exactly, and uses importance-weighted score differences as increments in a CUSUM recursion. The authors establish exponential false-alarm scaling and a first-order delay bound, and compare detection delay against likelihood-ratio CUSUM and score-based CUSUM in simulations and on MNIST and Oxford-IIIT Pet.
Arman Adibi, Mohammadreza Maleki, Sanjeev Kulkarni et al.
In briefDataSense-Bench: A Benchmark for Whether AI Can Choose Training Data
SummaryThis paper introduces DataSense-Bench, a benchmark testing whether frontier AI agents can select useful training data and predict its effect. Agents may inspect data and run analysis code, but cannot train the model or access the evaluation tasks; the chosen subsets are used to fine-tune a small LLM and evaluated under a standard protocol. The results show that gains over random selection are limited, agents rank their chosen subsets unreliably, and ranking ability varies across tasks.
In briefJust Weather Scoring: Efficient End-to-End Precipitation Nowcasting via Diffusion
SummaryThis paper proposes Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model that forecasts precipitation directly in radar space and supports few-step generation. It combines Masked Asynchronous Diffusion with a scoring-rule objective to reduce training and inference cost. The authors report state-of-the-art probabilistic forecasting on the SEVIR and MeteoNet benchmarks, with even the smallest model using fewer parameters and running more than 17 times faster.
Jannik Wiese, Johannes Schusterbauer, Tommaso Martorella et al.
In briefA closer look at LLM agents for black-box optimization, tested on a shared benchmark
SummaryThis paper studies how large language model agents can solve black-box optimization problems, where each objective evaluation is costly and the budget is limited. The authors introduce AgenticBBO-Bench, a cross-domain benchmark covering synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design under a unified finite-budget protocol. Experiments report that agentic methods score higher than direct LLM-based methods in all five domains and beat the best numerical optimizers in four, and the paper also analyzes the effects of tools, task information, and the LLM's role in search.
In briefAI-based onboard ship detection to cut Earth-observation data volume on Versal hardware
SummaryThis paper studies detecting vessels onboard satellites so that only content containing ships is downlinked, reducing data by content rather than per-pixel coding. The authors trained a YOLOX-S detector on 68 annotated Maxar scenes with 43 vessel classes, then quantized it and deployed it on a Versal VC1902 DPU, processing a scene in a few seconds with limited accuracy loss. On dense harbor and coastal scenes, tile-based downlink kept 98% of vessels using 29% of scene volume, and crop-based downlink kept 83% using 3%.
Thomas Goudemant, Aur\'elien Bobey, Omar Hlimi et al.
In briefRecursive self-improvement via multi-agent self-supervision, a method where a model refines its own multi-agent workflows.
SummaryThe paper proposes Multi-Agent Self-Supervision (MASS), where a single base model iteratively proposes, executes, and self-evaluates multi-agent workflows, optimized through constrained evolutionary search. After two MASS cycles with Qwen3.6-27B, the authors report 1.2 to 1.6 times higher performance per output token on four open-ended public benchmarks. They also report that a student trained on multi-agent traces outperforms a single-agent student trained on 1.4 times more tokens.
In briefUnifying Policy Learning and State Prediction via Spatial Language Modeling
SummaryThe paper introduces Spatial Language Modeling, which represents scene contours, goals, action targets, and future states with a shared vocabulary of discrete coordinates and semantic tokens. A single autoregressive Transformer learns both action generation and action-conditioned state prediction through a common next-token objective, after random-play pretraining and joint training on expert demonstrations. On Push-T in simulation and on a real robot, the model achieves competitive results, with higher task success and target coverage than the evaluated real-robot policy baselines.
In briefLearning to Plan by Looking Back: Hindsight Hierarchies for Reasoning Models
SummaryThe paper proposes a self-improvement loop for reasoning models, based on the idea that a supplied solution can help a model extract useful solution ideas even when a problem is beyond its current ability. The same model is trained jointly to predict solution ideas from problems, reverse-engineer ideas from problems and known solutions, and solve problems using provided ideas. The authors give a formal specification and a concrete instantiation for interactive theorem proving in the Lean prover, while empirical evaluation is left to future work.
In briefIs real-world training data needed for generalist graph anomaly detection?
SummaryThis paper introduces AG-FORGE, a tool that automatically synthesizes anomalous graphs to address the scarcity and high cost of real-world anomalous graph data. The authors find that training on synthetic data can match real-world training performance, but existing methods limit further gains. They propose TS-GGAD, which combines topological and semantic anomaly evidence with a curriculum learning strategy for large-scale synthetic training. Experiments on 14 real-world datasets show it significantly outperforms state-of-the-art methods.
In briefScalable graph generation using soft community hierarchies
SummaryThis paper introduces Schema, a method that recursively splits a reference graph into a hierarchy of soft communities, giving each node a membership distribution. Generation runs in three independently trained stages: node attributes conditioned on soft memberships, intra-community edges, and inter-community connections over bridge nodes. On four real-world attributed graphs, Schema reproduces the balance between local and long-range structure while reproducing only a small fraction of reference edges. The authors also test scalability on six additional graphs with up to 10 million nodes.
Ahmet T\"uzen, Helge Langseth, Kjetil N{\o}rv{\aa}g
In briefBayesian causal forests for zero-inflated outcomes: a new method for estimating treatment effects
SummaryThis paper addresses treatment effect estimation for semicontinuous outcomes that have many zeros alongside a skewed positive distribution. It proposes ZIC-BCF-Smear, a two-part nonparametric method that models whether an outcome occurs with a Probit BCF and positive intensity with a Gaussian BCF, then converts predictions back to the response scale using Duan's smearing estimator. Simulations reportedly show it performs stably without parametric assumptions. The authors apply it to the SPECTRUM phase III head-and-neck cancer trial, estimating how panitumumab plus chemotherapy affects days with severe toxicity, with the effect found mainly in the positive-intensity part.
In briefWhen KL Regularization Misfires in Group Policy Optimization
SummaryThis paper analyzes seven failure modes where KL regularization against a reference policy interacts poorly with rewards in group-relative policy optimization, such as residual KL updates after reward clipping or in groups with identical rewards. It proposes Zero-Sum Calibrated Policy Optimization (ZCPO), which uses conditional-KL-measured relative drift to calibrate within-group reward coefficients and folds them into the base surrogate. Mathematical reasoning experiments and ablations are reported as supporting the design in the authors' settings.
In briefConnected Self Forcing: Training Video Autoregression Beyond Local Learning
SummaryThis paper introduces Connected Self Forcing, a training framework for autoregressive long video generation. Standard Self Forcing detaches historical caches to save memory, which cuts backward gradient paths between chunks; this method reconnects them so later predictions can guide how earlier context is generated. The authors also propose shortcut gradient replay to recover cross-chunk gradients without storing the full rollout graph, and combine it with distribution matching distillation. Experiments report better long-horizon visual quality and temporal consistency without changing the inference procedure.
In briefOptimal regret in adversarial MDPs with stochastic hard constraints
SummaryThis paper studies episodic constrained Markov decision processes with adversarial losses, where expected cost constraints must be satisfied in every episode. The authors propose MA-OPS, which combines optimistic search for the Slater margin with pessimistic evaluation of selected policies to safely learn a policy with a large feasibility margin. The algorithm achieves regret of order roughly √T/ρ + 1/(dρ), and a matching lower bound shows this dependence on T, d, and ρ is optimal up to logarithmic factors.
In briefBayesian optimisation that improves designs while keeping key physical states near trusted values
SummaryThis paper addresses engineering design problems where the objective and constraints depend on a PDE solution determined by design parameters. The authors propose pulling state-space constraints back into design space by precomputing controls whose linearised response stays within tolerance, forming an ellipsoid for efficient candidate sampling. The linear response map is refined online and the ellipsoid is rebuilt accordingly. The method is demonstrated end-to-end on Tokamak divertor optimisation under plasma-boundary preservation.
Gabriel Diaz-Aylwin, Vignesh Gopakumar, Omkar Myatra et al.
In briefGenerating healthy counterfactual mammograms via diffusion inpainting to improve breast cancer classification
SummaryComputer-aided diagnosis systems for breast cancer screening can miss cancers (false negatives), which delays detection and treatment. The authors train a diffusion model on healthy mammograms and use RePaint to replace annotated lesion regions with realistic healthy tissue, creating counterfactual images that enrich training data. On the VinDr-Mammo dataset, sensitivity improved across four classifier architectures, especially at 80% fixed specificity.
In\^es Cruchinho Garcia, Mariana Mour\~ao, Francisco Maria Calisto et al.
In briefHow tokenizer choice affects different languages in multilingual language models
SummaryThe study trained 123 language models using 54 tokenizers, holding everything except the tokenizer fixed. It finds that tokenizer choice matters more for languages with less language-model training data, and that excluding a language from tokenizer training raises its bits-per-byte. The authors also show that intrinsic tokenizer metrics can predict pairwise BPB rankings of downstream models, offering a way to screen tokenizer candidates before training.
Clara Meister, G\"ul Sena Alt{\i}nta\c{s}, Antoine Bosselut
In briefInstruction-conditioned spectrum understanding with budget-adaptive signal tokens
SummaryThis paper proposes BATok, a budget-adaptive tokenizer that adjusts the number of tokens to input length and allocates them according to signal content, so raw I/Q signals can be fed to vision-language models. The authors also introduce EMSpec-Instruct, a multimodal instruction dataset linking raw I/Q signals, waterfall images, and language for modulation recognition, structured detection, and language-conditioned signal grounding. Experiments reportedly show BATok learns effective signal representations and achieves competitive performance across all tasks.
In briefQ-Shaped Options: A New Way to Learn Action Abstractions for Hierarchical RL
SummaryThis paper studies how to learn a suitable action abstraction (options) for hierarchical reinforcement learning. The authors identify three desiderata for such abstractions and introduce Q-Shaped Options (QSO), where the low-level Q function treats options as goals to keep distinctions needed for optimal control, and the high-level Q function treats them as actions to discard unnecessary distinctions. In offline goal-conditioned locomotion and manipulation environments, QSO learns semantically meaningful option spaces, outperforms baselines, and achieves non-zero performance on tasks where all other evaluated algorithms fail.
Clarisse Wibault, Antoine Gorceix, Antonio L\'eon Villares et al.
In briefOA-MAP: An Evidence-Grounded Multi-Agent System for Explainable Knee Osteoarthritis Progression Assessment
SummaryThis paper presents OA-MAP, a multi-agent framework that assesses structural and pain progression in knee osteoarthritis using MRI, X-ray, and clinical agents coordinated by a central agent. The system retrieves literature as supporting evidence and lets clinicians review and correct intermediate findings through a human-in-the-loop mechanism. Evaluated on 600 participants from the FNIH Osteoarthritis Biomarkers Consortium cohort, the fusion models reached AUROCs of 0.80 for structural progression and 0.68 for pain progression on a 100-participant test set.
In briefRehearse Everything, Remember Nothing: Attic-KV Rehearses What Will Be Read
SummaryThis paper studies how to compress key-value (KV) caches when the questions they will face are unknown. It finds that under tight budgets, rereading the entire context, which is the prevailing rehearsal approach, performs poorly. The proposed training-free method, Attic-KV, has the model quiz itself with question-answer pairs that quote the context, together with a content-adaptive number of anchor tokens. Reported results show it is the best training-free method across the eight tested settings on RULER and LongBench, and it also improves two other hosts.
WhyReplaces full-context rereading with self-quizzing for KV cache scoring, showing large gains across hosts under tight budgets on RULER and LongBench.
In briefA neural density functional for polymer electrolytes that transfers across system sizes and coupling strengths
SummaryThe paper presents a neural density functional describing ion structure in salt-doped polymer melts, with the permittivity entering through the Bjerrum length. The model is trained on planar density profiles at several Bjerrum lengths and, without retraining, predicts density profiles in larger boxes, under two-dimensional fields, and bulk structure factors. At fixed coupling, it is reported to be more accurate than three neural architectures.
WhyProposes a neural density functional with exact Coulomb energy identity and transferable coupling; force matching shows generalization to larger boxes and unseen coupling and concentration.
In briefChecking Whether Narrowly Finetuned Models Contradict Themselves
SummaryThe paper measures model coherence by resampling answers to 175 questions designed so that ambiguity and indifference cannot explain contradictions. The metric has high specificity and flags models only when inconsistencies are glaring. Narrow finetunes score poorly, and inspected model organisms show problems such as identity conflation, introspection failures, and rationalizations.
Robert Graham, Yariv Barsheshat, Phil Blandfort et al.
In briefLVS: Quickly Generating Nearby Views by Reusing Earlier Renders
SummaryExploring a 3D scene interactively needs frequent view changes, and nearby views share much visible content. This paper proposes a per-scene method that reuses previously rendered RGB-D images for nearby camera poses instead of re-rendering each view. It warps source content using depth and relative pose, then uses a lightweight multiscale network to predict RGB residuals that fix artifacts and fill in missing appearance. On GS-render, residual refinement improves PSNR by 0.72 dB over pure warping, and the method shows low query latency on captured and rendered scenes.
In briefSuperNav: An Agentic Navigation System for Any Task in Any Scene
SummaryThis paper presents SuperNav, a navigation system for general-purpose service robots that handle varied human requests in unfamiliar environments. Instead of fine-tuning a multimodal large language model for navigation, it equips a pretrained model with an agent harness, letting the model interpret requests, understand scenes, and decide while navigation tools execute motion. A unified visual-point interface lets the model specify destinations in images and revise decisions from execution feedback. SuperNav outperforms four baselines on instance-level, multi-object, and demand-driven tasks, and is also evaluated on HM3D and deployed on a real quadruped robot.
In briefUse and Disuse: Intent-Structured Memory Consolidation for LLM Agents
SummaryThe paper proposes Hippocam, a hierarchical memory and continual learning architecture that helps LLM agents turn long-running experiences into reusable knowledge. Work is organized into nested intents, and completed intents are consolidated into the outcomes and state needed later rather than full working details. Long-unused experiences become increasingly abstract through recursive consolidation, while recalled experiences are reinforced and updated alongside new ones. The authors claim agents can learn and evolve capabilities from their own experience without parameter updates.
