In briefDeepAJM: A Deep Joint Model for Irregularly Sampled Longitudinal and Survival Data
SummaryThis paper proposes DeepAJM, a deep joint model that analyzes longitudinal measurements and survival outcomes together without relying on fixed parametric assumptions. It uses an encoder-decoder architecture to learn patterns in patients' time-varying covariates and links them to survival risk through an interpretable association structure that incorporates baseline covariates. The model was tested on a cardiovascular-disease EHR cohort, the PBC2 primary biliary cirrhosis dataset, and a simulated dataset, where it achieved the best C-index, time-dependent AUROC, and AUPRC across all datasets.
In briefHow vision-language models gain and lose semantic skills during fine-tuning
SummaryThis paper studies how semantic capabilities in vision-language models are acquired, peak, and sometimes decline during fine-tuning, tracked across training rather than at a single checkpoint. It introduces a trajectory-based framework and Structured Semantic Routing (SSR), tested on six pretrained backbones. The results show that the form of supervision shapes what is learned, and that continued optimization can preserve target class-name retrieval while reducing transferable semantic capability.
In briefWildMatch: Weakly Supervised Adaptation of Image Matchers for Recognizing Individual Wild Animals
SummaryThis paper addresses recognizing individual animals in camera-trap images by retrieving the correct animal from a set of known individuals. The authors adapt a pretrained keypoint matcher using only identity labels, with no keypoint or geometric correspondence annotations, through contrastive fine-tuning. On open-source wildlife re-identification datasets, the method outperforms off-the-shelf matchers and a state-of-the-art local-global fusion approach, and it generalizes to held-out individuals.
Turhan Can Kargin, Piotr Kubaty, Ekaterina Rostovskaya et al.
In briefHyperNSDE: Personalized Neural SDEs That Generate Joint Static and Longitudinal Clinical Data
SummaryThis paper proposes HyperNSDE, a continuous-time generative model for producing synthetic patient data that jointly covers static covariates, irregularly sampled longitudinal trajectories, and observation times. A hypernetwork conditions a latent Neural SDE on patient baseline representations, so these characteristics shape trajectory evolution. Experiments on simulated and real clinical datasets report improved observation-time fidelity and competitive performance.
In briefGeoWM: Efficient Direct World Modeling in Explicit Geometry
SummaryThis paper presents GeoWM, a geometry world model that predicts future 3D scene geometry for autonomous driving and robotics. It converts observed RGB frames into a geometric history using a geometry foundation model, then uses a flow-matching transformer to forecast scene geometry at a chosen future horizon without recursive rollouts. Experiments on four datasets covering urban driving, aerial flight, and dynamic manipulation show that GeoWM outperforms evaluated world models in depth, camera pose, and 3D geometry forecasting, while substantially reducing inference time at longer horizons.
In briefSimCortex v2: Joint brain cortex surface reconstruction with almost no collisions or self-intersections
SummaryThis paper presents SimCortex v2, a deep learning method that simultaneously reconstructs the left and right white matter and pial surfaces of the brain from T1-weighted MRI. It first estimates topologically correct initial surfaces from a volumetric segmentation, then jointly refines all four using multi-scale stationary velocity fields. Evaluated on 560 cases from 14 cohorts, it matched the strongest baseline's surface-distance accuracy (0.253 mm) and showed no detected inter-surface collisions in 92.14% of cases, with the lowest self-intersection fraction (0.044%) among learning-based methods.
WhyJointly reconstructing four cortical surfaces substantially reduces collisions and self-intersections over baselines, validated on 560 cases from 14 cohorts.
In briefMemCo: Memory-Centric Collaboration for LLM Agents in Unseen Environments
SummaryThis paper proposes MemCo, a framework that helps large language model agents generalize to unseen interactive environments through memory collaboration. It keeps local memory for environment-specific details and global memory for transferable workflows distilled from local trajectories, routing relevant memories based on the agent's current state and decision phase. Experiments on interactive decision-making benchmarks report that MemCo improves task success and reduces redundant exploration compared with isolated-memory and shared-memory baselines.
In briefMultigroup Fairness and Omniprediction: Separations and Equivalences
SummaryThis paper studies omniprediction, a guarantee requiring one predictor to compete with the best benchmark hypothesis under any loss from a family. It examines how omniprediction relates to multigroup fairness notions such as multiaccuracy and multicalibration. The authors show that plain omniprediction does not require multigroup fairness, while the stronger loss outcome indistinguishability is equivalent to a form of calibrated multiaccuracy.
WhySeparates omniprediction from multigroup fairness and proves loss OI equals calibrated multiaccuracy, giving clear new theoretical characterizations.
S\'ilvia Casacuberta, Parikshit Gopalan, Varun Kanade et al.
In briefIdentity-conditioned score fusion for open-set person re-identification
SummaryThis paper proposes identity-conditioned score fusion, a training-free framework that tailors fusion weights for face, gait, and body-shape cues to each gallery identity. It contrasts intra-identity consistency with cross-identity impostors to extract identity-specific profiles, combined with query-conditioned adaptation through a parameter-free rule. On three clothes-changing re-identification benchmarks, the method consistently outperforms statistical, rank-based, and learned baselines, reducing the false non-identification rate by up to 8.8% absolute.
In briefWho Wrote It Is Not Enough: Detecting Who Contributed the Insight
SummaryAs LLMs help with scientific writing and peer review, this paper argues that knowing who wrote a text is no longer enough; the source of the underlying insight matters. The authors define Insight Provenance and build InsightProv-v0 from 4,057 papers and 12,660 human reviews, simulating LLM involvement with GPT-4o, Gemini, and DeepSeek. They propose a two-stage adversarial framework to suppress shortcut signals, and find that AI insights mostly stay close to generic or paper-provided information, while human insights more often bring external knowledge and independent judgment.
In briefBudget allocation for LLM evaluation under hard resource limits (HARP method)
SummaryThis paper studies how to evaluate large language models in multi-turn interactions by measuring how many steps they need to reach an event, such as a successful jailbreak or task completion. When compute is limited, some interactions stop before the event occurs, so event times are only partly observed. The authors propose HARP, a budget allocation method that never exceeds a hard budget and reallocates unused budget adaptively, and use it to build lower predictive bounds and metric estimates. Experiments on agentic tasks, jailbreaks, toxic content generation, and RAG hallucinations report near-nominal coverage with low variance while staying within the budget.
In briefMeasuring redundant and synergistic information in multivariate Gaussian systems using Blackwell order
SummaryThis paper defines a partial information decomposition (PID) for multivariate Gaussian systems, separating the redundant and synergistic information that several sources provide about a target. The definition is based on the Blackwell order, and the authors prove that Gaussian channels are optimal for extracting redundant and union information. Their union information and synergy match the BROJA measures, and closed-form expressions are derived for the two-source case. The method is applied to an optimal control problem involving sensors and memory.
WhyDefines Gaussian partial information decomposition via Blackwell order, proving Gaussian channels optimal and giving closed forms for two sources.
In briefDo Stacked Agent-Action Gates Fail Independently? A Test of Rule and LLM Judges
SummaryThis paper tests whether the miss rates of runtime gates for agent tool calls multiply, as stacking them assumes. Using 1,119 labelled agent actions from three corpora, it evaluates one deterministic rule layer combined with four LLM judges. Two judges together act like roughly 1.2 to 1.4 independent layers, and the rule layer plus one judge like roughly 1.86 to 2.09, with results varying by miss definition and data subset.
In briefExplainable benchmarking for data-driven prediction of post-wildfire debris flows
SummaryThis paper addresses data-driven prediction of post-wildfire debris flows, which are triggered by intense rain on recently burned terrain. The authors present a unified benchmark for fair, comprehensive comparison across diverse models and feature configurations. They also propose a reinforcement learning-based feature selection framework that identifies factors whose perturbation makes positive and negative events indistinguishable, to uncover regional mechanisms. Code and the benchmark are publicly available.
In briefA Validated Dataset and Benchmark for Consistent Multi-Diagram SysML Models
SummaryThis paper introduces SEMAADB, a dataset of 3,000 engineering contexts and 15,000 diagrams, where each context contains five connected SysML views. A set of 100 contexts is human-verified to serve as the benchmark test set. Three language models were evaluated on diagram repair and cross-diagram update tasks. Results show syntax repair is nearly solved, while semantic repair and cross-diagram consistency remain challenging.
In briefTracking Is Not Permanence: Do Video World Models Remember Hidden Objects?
SummaryThis paper examines whether video world models keep information about objects after they become hidden. The authors hide an object from a frozen V-JEPA 2 predictor and compare its predictions with the encoder's representations of two worlds differing only in the hidden region. Stationary objects are partly kept, objects inside containers are essentially lost, and moving objects disappear within about 0.3 seconds. The encoder still detects the object, but the predictor's output lacks it, and training the predictor on synthetic containers improves this on rendered scenes.
In briefChecking whether signals really help decide when an AI model should escalate a query
SummaryThe paper studies when to pass a query from a small language model to a larger one. It tests semantic entropy on benchmarks such as GSM8K, finding it reliably flags the small model's mistakes and improves routed accuracy over random escalation at matched cost. The main contribution is a set of checks showing how a signal may only track question difficulty, along with a general checklist for evaluating escalation signals.
Ramin Pishehvar, Andrea Morandi, and Mahesh Viswanathan
In briefTrajectory-Retrieval Speculative Decoding: Whether a Model's Own Past Outputs Speed Up Generation
SummaryThis paper studies when a language model's own earlier generated reasoning trajectories can supply useful draft tokens for speculative decoding. It proposes Trajectory-Local Adaptive Retrieval (TLAR), which retrieves matching continuations from the current trajectory, adjusts retrieval use and candidate width based on recent verification results, and combines them with model-generated drafts. Tested on code debugging, mathematics, and open-ended writing, TLAR improves token acceptance under matched verification budgets and raises end-to-end throughput over a draft-model baseline.
In briefRELACE: Retrospective likelihood-based credit estimation for long-horizon language agents
SummaryThis paper introduces RELACE, a critic-free reinforcement learning method that assigns finer-grained credit to individual actions of multi-turn language agents. It scores executed actions by their likelihood under original and outcome-augmented contexts, uses this to reweight discounted returns, and builds local advantages from equivalent states within a task. Experiments on ALFWorld and WebShop with Qwen2.5 models report gains over GRPO, GiGPO, and HCAPO.
In briefStepped MoE: a language model whose inference cost can be configured per input
SummaryThis paper proposes a unified framework that combines elastic structures with sparsely gated (MoE) architectures so a large language model can adapt to both deployment constraints and task requirements. The model conditions on context and target efficiency specifications, letting users control the accuracy-efficiency trade-off at inference time. The authors report that one model can use 1 to 4 billion parameters and is 2-5% more accurate than dense counterparts on knowledge-intensive benchmarks.
In briefRationale-Guided Policy Optimization: A Reinforcement Learning Method That Uses Adaptive Reasoning Scaffolds
SummaryThis paper proposes RGPO, a framework that addresses reward sparsity in reinforcement learning for language models by adaptively using ground-truth rationales as temporary guidance. Only higher-reward, model-generated solutions produced with guidance are transferred back to the unguided setting, so auxiliary data need not match the RL task format. The authors report consistent gains over RLVR baselines on both language-only and vision-language reasoning tasks, with ablations indicating adaptive guidance is a key factor.
In briefCausalBind: Causal Learning Method for Protein-Molecule Virtual Screening
SummaryThis paper proposes CausalBind, a method for virtual screening of protein-molecule binding. It assumes binding is driven by a few sparse local interactions, such as hydrogen bonds and salt bridges, and models this with a V-structure causal model, with identifiability results under sparsity conditions. On the DUD-E and LIT-PCBA benchmarks, all variants outperform strong retrieval baselines and generalize to target- and scaffold-level out-of-distribution splits.
In briefA dataset of doctrine-grounded visual questions for tactical combat casualty care
SummaryThis paper presents TC3-VQA, a dataset of 581 items and 1,860 questions built from public TC3 instructional and field videos and authoritative TC3 documents. The questions cover intervention recognition, doctrine, clinical reasoning, procedural guidance, and refusal when visual information is insufficient, with answers preserving verbatim source passages. Annotation quality was assessed through automated audits and ratings by two physicians and two medical students, with human ratings available for 88 retained items.
In briefLogbook: A Benchmark for Hour-Long Audio Event Understanding
SummaryThis paper introduces Logbook, a benchmark for understanding audio recordings lasting from ten minutes to six days. Systems must segment a continuous recording without gaps and assign each segment an event label and description. The authors compare 52 end-to-end and cascaded systems and examine the effects of fine-tuning, context length, and reasoning budget. The task is found to be tractable, but the best systems still fall short of the human reference.
In briefTraining-free erasure of unwanted concepts in text-to-image models with hierarchical grounding
SummaryThe paper proposes HGSS, a training-free method for removing specified concepts from text-to-image diffusion models, including in compositional prompts. It combines hierarchical span grounding to locate erase targets and dynamic attribute binding during early denoising, without changing model weights. The authors report that on SEE, hierarchical evasion drops from 29.54 to 10.02, and on UnlearnCanvas it slightly improves over the matched Semantic Surgery baseline.
In briefSelective critique method for cost-efficient LLM agents in long-horizon tasks
SummaryThis paper proposes SAG, a framework that lets large language model agents decide step by step whether to invoke external critique, calling it only when the expected benefit justifies the cost. The gating mechanism is training-free and uses action-level ambiguity signals, namely global entropy and local top-2 margin, to estimate critique utility. SAG also applies online bootstrapped self-improvement so the actor gradually relies less on critique. On ALFWorld, task success rose from 24.6% to 78.4% with a token budget comparable to ReAct.
In briefMemory-efficient expert routing for distributed MoE training (RelayMoE, a ring-based execution model)
SummaryAs Mixture-of-Experts models scale to hundreds of experts, memory in distributed training becomes a bottleneck, mainly due to the MoE block. The paper proposes RelayMoE, a ring-based execution model where expert weights or tokens circulate and are computed locally, avoiding full top-k-expanded dispatch buffers and overlapping communication with computation. On 30B to 57B production MoE models, RelayMoE reports a 2x average speedup over Megatron-LM in single-layer experiments. Under the same GPU memory budget in full-model training, it improves throughput by up to 2.02x and extends the longest trainable sequence length by up to 2.85x.
Arnab Kanti Tarafder, Jaume Guasch-Mart\'i, Gokcen Kestor et al.
In briefControlling MoE expert routing for better agent reinforcement learning
SummaryThis paper studies how mixture-of-experts models used in long-horizon LLM agents select experts during reinforcement learning post-training. The authors find that expert routing overlaps more between turns with similar agent operations, but standard RL ignores this structure. They propose a hierarchical routing control framework that aligns turn-level expert selection with agentic operations and regularizes token-level consistency, plus an entropy-gated mechanism for stability. The framework reports success-rate improvements on all evaluated benchmarks.
In briefSharedKV-BT: Node-Local Typed Decisions for Behavior-Tree Agents
SummaryThis paper proposes SharedKV-BT, in which each active node of a behavior tree exposes stage-local fields and candidates, and Shared-KV scores them in parallel before passing the chosen decision to a separate execution system. The method was tested on robot manipulation, mobile navigation, and computer-use tasks. It made typed decisions 2.36 to 4.15 times faster than prompt-matched autoregressive decoding, and on manipulation it raised joint decision accuracy from 75% to 94% and closed-loop success from 0% to 60%.
In briefLocating any object in X-ray scans without human labels
SummaryThis paper presents LAO-X, a self-supervised framework for locating arbitrary objects in X-ray security scans. It mines object instances using saliency guidance, synthesizes training images physically in the absorbance domain, and fine-tunes Segment Anything Model 2 with an occlusion-controlled curriculum. Across six X-ray benchmarks, it improves mAP by 2% to 23% over SAM2 and X-ray-specific baselines in cluttered scenes, without any human-annotated labels.
In briefScale-invariant training method for time series foundation models
SummaryThe paper shows that affine normalization such as ReVIN inverts outputs before computing loss, so high-scale series dominate training, a problem it calls scale-contaminated training (ScaleCon). It proposes scale-invariant training (ScaleIn), which computes loss on scaled targets, and proves that gradients and the optimization trajectory are invariant to independent rescaling of training series. Across four TSFM architectures, ScaleIn lowers MASE in all 24 comparisons, with average reductions of 18.8% on GIFT-Eval and 21.9% on the M-competitions, and can often be adopted with a one-line code change.
WhyIdentifies scale-induced implicit weighting from normalization inversion, proves gradient invariance, and validates across four architectures and benchmarks.
Ignacy Stepka, Willa Potosnak, Kin G. Olivares et al.
In briefRule-Based Languages for Neurosymbolic AI Explained
SummaryThis survey reviews the main rule-based languages used as the symbolic component of neurosymbolic AI, namely Datalog, answer set, and probabilistic logic programs. It compares them across semantics, expressiveness, neural integration, and evaluation mechanism, analysing over 50 recent systems and applications. The paper also provides a decision matrix linking application scenarios to required features and outlines open problems.
In briefReducing Over-Reliance on Agent Memory at Inference Time
SummaryThis paper shows that LLM agents can be misled by retrieved memories that are correct and well-retrieved when only part of their evidence applies to the current task, which the authors call memory over-reliance. The failure is strongest under partial query-memory overlap. The authors propose MEMTRIM, a plug-and-play framework that indexes memory evidence at write time and controls its reuse at read time by removing repeated or conflicting evidence, without retraining. Experiments reportedly reduce over-reliance while preserving the benefits of useful memory.
In briefVerifying Whether Retrieved Memories Are Appropriate for the Current Request in Long-Term AI Agents
SummaryThe paper argues that long-term memory agents can retrieve information that is inappropriate for a given request, such as data belonging to another user, violating policy, or in an incompatible lifecycle state, and that standard recall and answer accuracy metrics do not reveal this. The authors propose a verification framework that labels each memory-query pair as admissible, inadmissible, or unresolved, and tracks memory exposure in prompts and disclosure in answers. Evaluation uses frozen rankings from two public benchmarks, RHELM and MemOps, along with controlled exposure scenarios.
In briefKurate: A System for Scalable Quality Checks of Scientific Studies
SummaryKurate is a system that uses large language models to judge the quality of published studies and links each judgment to the supporting text. The authors scored 4,347 papers, 3,913 of them randomized trials, on eight design and reporting dimensions. Most papers showed problems with statistical power, selective reporting, and analysis prespecification. On 60 expert-annotated clinical trial documents, Kurate's extracted information agreed with expert labels in most cases.
In briefLogistic regression with missing covariates under minimal assumptions
SummaryThis paper studies parameter estimation in logistic regression when some covariates are missing and their distribution is unknown (only bounded). The authors propose a stochastic approximation method based on Z-estimation with a novel monotone operator. Under the assumption that covariates are missing completely at random, they show the method is computationally efficient and recovers the signal at parametric rates, outperforming the complete-case estimator. They also provide information-theoretic lower bounds showing the dependence on the missingness probability is fundamental.
WhyDesigns a Z-estimator with a novel monotone operator for unknown bounded covariates, with sharp risk bounds and matching information-theoretic lower bounds beating complete-case estimation.
Jyotishka Ray Choudhury, Kabir Aladin Verchand, Richard J. Samworth et al.
In briefA single-loop, constant-batch first-order penalty method for stochastic bilevel optimization
SummaryThis paper proposes SICO, a single-loop first-order penalty method with constant batch size for stochastic nonconvex-strongly convex bilevel optimization. Each iteration performs one stochastic gradient update for both the lower-level original and penalized problems, and an exponential moving average stabilizes the upper-level gradient estimator. The authors report O(ε^-6) sample complexity under bounded variance and O(ε^-4) under mean-square smoothness, both using O(1) samples per iteration.
In briefFlowAgent: Fixing Test Failures in Developers' Flow at Google Scale
SummaryThis paper presents FlowAgent, an AI agent deployed at Google that automatically repairs test failures during the pre-submit stage within continuous integration systems. It is integrated into Google's internal developer tools and uses a generate-and-validate loop with abstention filters to keep suggestions high-quality under strict latency limits. The authors report 67.18% accuracy on 195 real-world test failures, and after company-wide deployment, suggested fixes on 295,508 changes, with 65,069 previewed and 28,554 applied.
WhyNovel low-latency pre-submit repair agent; 67.18% accuracy on 195 real failures plus large-scale Google deployment data support its claims.
In briefFlexiFlow: Switching ML Models Mid-Workflow Using Bandit Learning
SummaryThis paper presents FlexiFlow, a dataflow system that switches to alternate machine learning models at runtime when the current model shows low accuracy. It ranks models with a new multi-armed bandit approach that considers model runtimes, the probability of passing user-defined assertions, and the workflow's computational structure. Experiments report that runtime switching while reusing intermediate results improves accuracy and yields a 48% efficiency gain over sequential workflow runs.
Abhilash Jindal, Todd Nief, Bhanu Prakash Vangala et al.
In briefLock-in EP: An In-Situ Training Algorithm for Oscillatory Hardware
SummaryThis paper proposes lock-in equilibrium propagation (LIEP), a training method for oscillatory networks. It provides local gradient information for each component without separate forward and backward sweeps, potentially enabling in-situ learning on analog oscillatory hardware. The authors show LIEP can be used both for training from scratch and for recovering performance after pre-trained parameters are perturbed. Validation so far covers only shallow networks, though the authors suggest alternative architectures could extend it to deeper, larger-scale networks.
In briefA Trust Layer for Agent Evaluation: Checking Whether Benchmark Scores Can Be Believed
SummaryThis paper proposes an add-on verification layer that runs after benchmark scoring and reports, for each recorded score, whether it should be trusted. It checks whether results are supported by the benchmark's grading logic, whether passing answers come from traceable computation, whether the agent's completion claims match what happened, and whether results remain stable across repeated runs. Applied to five agent configurations on 108 tasks, the authors found many passing runs lacking traceable computation, confirmed false completion claims, and unstable results for 18-46% of tasks. Only 22.6% of recorded passes cleared all four checks.
Mohammadreza Sediqin, Shivali Dalmia, Srinivasa Karthikeya Reddy Kovvuri et al.
In briefCIRRA: A Framework for Household Robots to Merge New Instructions Into Ongoing Tasks
SummaryThe paper proposes CIRRA, a framework that lets household robots accept new user instructions while still executing ongoing tasks. It uses an LLM to interpret new instructions and inserts the new subtasks into matching segments of the existing plan, keeping the original sequence intact. The authors also release CHIRP, a text benchmark of 120 episodes across eight household environments. They report 74.2% decision agreement on CHIRP and better results than baselines on a Unitree G1 humanoid robot.
In briefNeural Network That Learns Merge Rules for Hierarchical Clustering
SummaryThis paper introduces NN-linkage, a neural network model that learns task-specific merge rules for hierarchical clustering while keeping the recursive structure and efficiency of classical linkage algorithms. It is algorithmically aligned with the Lance-Williams recurrence, which includes single, complete, and average linkage as special cases. The authors prove it is a universal approximator for continuous linkage functions and can exactly implement symmetric constant-coefficient Lance-Williams recurrences. It is evaluated on synthetic and real data for clock-tree routing and phylogenetic reconstruction.
Robert R Nerem, Pranav Singh, Cheyenne Ward et al.
In briefDoes the Model Use the Feature? Separating Steering from Mechanism in LLMs
SummaryThe paper argues that an LLM internal feature that tracks a concept and changes behavior when steered does not by itself show the model uses that feature. The authors propose tests that copy feature values observed on natural inputs into matched prompts (installation or removal) or restore a feature after an upstream edit (downstream rescue). They find that the two kinds of strength separate sharply across several representations, so conclusions hold only for the intervention tested.
In briefWhat Words Keep of a Place: Language-Only Reasoning for Cross-View Geo-Localization
SummaryThis paper asks whether cross-view geo-localization, matching ground photos to satellite tiles, can be done through language alone. A multimodal large language model describes each ground panorama and satellite tile as structured text, with no training involved. On 9,826 VIGOR pairs, ranking the full pool by description similarity almost never finds the correct tile (0.39% Recall@1). Within ten neighboring tiles, an LLM judge doubles random ranking and explains which description fields agree or conflict. Reranking a trained visual retriever works better with images (23.5%) than with the descriptions (10.7% Recall@1).
In briefVerifying Coordination in Parallel Coding Agents: NP-Bench and a Scheduling Planner
SummaryThe paper argues that teams of coding agents can each pass their own tests yet produce a broken merged result, which single-agent evaluation misses. The authors recast coordination as scheduling, partitioning work into disjoint declared scopes and ordering merges along producer-to-consumer dependencies before work begins. On NP-Bench, the planner raised clean-integration scenarios from 1/9 to 9/9 and reduced merge conflicts from 13 to 0.
In briefMemMux: Runtime Checks and Honest Memory Tracking for Many Parallel Coding Agents
SummaryThe paper presents MemMux, an unprivileged local runtime that helps developers manage memory used by several coding agents running at once on one machine. It attributes usage to each agent, verifies that a terminated agent's processes are fully reclaimed, and detects child processes that escape their agent. On a Linux host with a 7500 MiB budget, MemMux stayed under budget with zero swap, while ungoverned tools used about twice the budget and spilled roughly 2 GiB into swap. Its always-on monitoring costs about 2.7% CPU at ten agents, above the authors' 2% target.
In briefNeural network learns to mimic a saddle-point solver for graph optimization problems
SummaryThe paper proposes GraphPDHG, a message-passing framework based on the Chambolle-Pock Primal-Dual Hybrid Gradient method for solving general graph saddle-point problems. The authors show theoretically that the network can efficiently solve a family of such problems by simulating PDHG, and that it can learn an accelerated version of PDHG. Experiments use the model as a learned warm start for the second-order method SSNAL and show stronger size generalization than non-aligned graph neural network baselines.
Samantha Chen, Jesse He, Coleman Clougherty et al.
In briefNeural fields encode how they can adapt and where they came from
SummaryNeural fields are usually judged only by how well they reconstruct observations. This paper studies two other properties: how easily a fitted network adapts to new observations, and what its weights retain from earlier ones. Adaptation cost is closely predicted by a simple local linear model. For wave sequences ending at the same observation, the final weights recover the wave velocity sign with 68.6% accuracy, versus 50% from the current observation alone.
In briefPaper on internalizing agent experience into diffusion model weights via on-policy context distillation
SummaryThe paper proposes Diffusion On-Policy Context Distillation (D-OPCD), which treats an agent-improved prompt as privileged context and distills the agent harness's knowledge into diffusion model weights. Across four benchmarks, average direct-generation scores rose from 60.52 to 65.09. The authors also report that a second agent evolution round on the updated generator adds 1.83 points over a skill-free harness.
In briefCan Semantic Geometry Teach an AI to Judge Which Rules Apply?
SummaryThis paper examines how an AI agent can decide which rules apply to an action, testing whether vector geometry can identify governing policies and how an action relates to them. Across four studies, the approaches did not establish reliable pre-action judgment. A lexical router found every governing and blocking policy, yet the full pipeline escalated all test actions, including ones that should have been allowed. The authors conclude that AI judgment requires a consequence graph linking actors, authority, policy conditions and exceptions, and the changes an action would produce.
Thomson D. Nguy (Radiant Institute for Manifold Studies)
In briefLearning What to Distill: Adaptive Token Selection for Self-Distillation in Language Models
SummaryThis paper introduces BiToK-SD, a method for on-policy self-distillation in large language models that learns which token positions should receive distillation. It formulates token selection as a bilevel optimization problem, using a differentiable threshold-based relaxation for Top-K selection so that chosen positions adapt as the student model changes. Experiments on mathematical reasoning benchmarks report that BiToK-SD achieves the best average performance among compared methods with only lightweight extra computation.
In briefHybrid Cross-Modal Attention Network for Early Breast Cancer Detection in Low-Resource Settings
SummaryThis paper proposes a Hybrid Cross-Modal Attention Network (HCMAN) that combines mammogram images with structured clinical data to support early breast cancer detection in Sub-Saharan Africa, where diagnosis is often delayed. The model was developed and validated on 2,560 images from 1,024 patients at four Ethiopian referral hospitals, with biopsy-confirmed labels. The authors report 97.8% accuracy and an AUC of 0.987, and state the model is robust to low-quality images and can run on standard hospital workstations.
Simon Hadush Nrea (1), Filimon Gidey Gebremichael (1), Gebrekirstos Hagos Gebrekirstos (2) et al.
In briefSPECTRUM: A Method to Keep Diverse Correct Code Solutions When Models Learn From Themselves
SummaryThis paper studies a problem in which a code model repeatedly learns from its own outputs: correctness can rise while the variety of correct implementations shrinks. The authors propose SPECTRUM, which re-estimates loss-sensitive key/value geometry from a fixed reference anchor each round and applies full-rank proximal spectral modulation. After five rounds on MBPP, SPECTRUM retained 89.9% of the initial model's correct AST richness, versus 66.4% for Vanilla self-distillation. The student also showed higher matched-correct richness on HumanEval+ and APPS Intro without further training.
In briefBenchmarking Time Series Foundation Models for Load Forecasting Under Uncertain Covariates
SummaryThis paper benchmarks four models trained from scratch and four time series foundation models on three real-world short-term load forecasting datasets. The scenarios vary in the availability and quality of future covariate information. Chronos-2 performs best in zero-shot and fine-tuned settings when future covariates are available or accurately forecast. Its performance degrades as covariate forecasts become noisier, while TimesNet is more robust under severe covariate uncertainty.
In briefMonocular navigation for unknown spacecraft using a transformer-aided Kalman filter
SummaryThis paper proposes a learning-based method that estimates the position and orientation of an unknown target spacecraft using only a single monocular camera on a servicer spacecraft. It combines a transformer network, which estimates relative pose changes between images from matched features, with a Multi-State Constraint Kalman Filter. Trained on synthetic images of 103 spacecraft, the method was tested on 11 unseen spacecraft. Reported median errors are 3.7 degrees in attitude and 2.2% of range in relative orbit elements.
In briefConditional flow matching for transporting between Markov process trajectories
SummaryThe paper studies transport between trajectories of Markov processes, proposing a flow matching algorithm that learns a map from source to target trajectory distributions while preserving the Markov structure. It shows consistency in the population limit and derives finite-sample error bounds under mixing-time assumptions. In EEG-based image retrieval on THINGS-EEG2, adding a conditional flow to the ENIGMA decoder raised mean top-5 accuracy from 43.87% to 49.05%.
Syamantak Kumar, Dheeraj Nagaraj, Saptarshi Roy et al.
In briefHow inefficient is natural gradient descent? Exact optimality versus Θ(√log d) divergence
SummaryThis paper analyzes how far natural gradient descent, a common ML optimization method, departs from the shortest Fisher–Rao path on dually flat families. It defines an inefficiency ratio R, the Fisher length of the mixture geodesic divided by the Fisher–Rao distance, and bounds it by parameter dimension d. R equals 1 for quadratic-potential families, stays dimension-independent under bounded skewness and finite diameter, and grows as Θ(√log d) for products of scale families. Experiments confirm all three regimes.
WhyGives a precise characterization of natural gradient inefficiency, proving a Θ(√log d) dimensional growth, with experiments confirming each regime.
In briefMinimal Witness Reinforcement Learning: learning multiple minimal sufficient explanations from verifier feedback
SummaryThe paper formalizes minimal-witness identification, which seeks the irreducible conditions sufficient to produce an outcome, and introduces Minimal-Witness Reinforcement Learning (MWRL). MWRL uses only a black-box verifier's pass or fail signal and credits each sampled proposal by the coverage the group union would lose without it. According to the paper's experiments, MWRL recovers most minimal witnesses, whereas other methods return redundant supersets or a single witness.
WhyFormalizes minimal-witness identification as RL with a novel coverage-based credit assignment; experiments in planning and LLM settings show recovery of witness families, though scale is limited.
In briefTIDE 2.0: An Open, Model-Agnostic Tool for Keyed Anonymization of Clinical Notes
SummaryClinical notes must have protected health information removed before research use. This paper presents TIDE 2.0, an engine with two separable stages: an interchangeable recognizer and a keyed anonymizer, both runnable on institution-owned hardware. Surrogates are generated cryptographically without a stored linkage table, dates shift per patient while preserving intervals, and each value maps consistently under a given key. On two gold-annotated corpora, the default configuration reached span-level recall of 0.88 and 0.77 and precision of 0.88 and 0.87.
In briefAI Vision Tool for Spotting Illegal Cricket Bowling Actions
SummaryThis paper proposes a computer vision and deep learning method to detect illegal bowling actions in live cricket matches. The system finds a shoulder frame and a release frame in video of a delivery and flags the delivery if the bowling arm's angle changes by more than a set threshold, such as 15 degrees. Evaluated on a custom dataset of 62 videos from 11 male bowlers, it achieved a high true positive rate. The authors note that further testing across more conditions and larger datasets is needed.
Debopom Sutradhar, Niful Islam, Sudipto Mondal et al.
In briefData, Numbers, and Geometry: Tutorials on Numerics, Machine Learning, and Evaluation
SummaryThree practical tutorials for mathematical research, developed for a 2026 Banff workshop. The first builds a numerical method for exterior calculus using pointwise evaluations of differential forms and a flux formulation of the exterior derivative. The second uses elliptic curves, quivers, and a boundary value problem to show how mathematical structure guides neural network design, and uses interval arithmetic to bound a trained network's residual. The third covers evaluation and presentation of machine learning results, including metrics, uncertainty, thresholds, and ROC curves.
In briefCascadia: Running a 975B MoE Model Resident Across Eleven AI PCs
SummaryThis paper presents Cascadia, a system that runs Inkling, a 975B-total, 41B-active-parameter mixture-of-experts model, resident across eleven Intel Core Ultra X7 358H AI PCs. The authors contribute a custom MoE engine that preserves the model's routing rules and uses FP16 expert computation with FP32 output restoration. The paper reports measured aggregate decode throughput and first-token latency across concurrency levels from 1 to 176 streams, and long-context results up to 64k tokens.
Tate Berenbaum (Not Community Labs Inc.), Matias Parij (Not Community Labs Inc.), Muthaiah Venkatachalam (Intel Corporation)
In briefConstant-Curvature Sliced Gromov-Wasserstein for Aligning Distributions Across Different Curvature Spaces
SummaryThe paper proposes CCSGW, a new divergence for comparing probability distributions supported on heterogeneous constant-curvature spaces such as hyperbolic and spherical spaces. It introduces geodesic-based one-dimensional projections for spheres and extends sliced Gromov-Wasserstein to these spaces, preserving intrinsic geometry at lower computational cost. The authors provide theory showing CCSGW controls intrinsic geometric discrepancy and report consistent gains when integrated into graph anomaly detection, node classification, and multimodal learning.
In briefRoboCap: A New Platform for Egocentric Robot Learning
SummaryEgocentric manipulation data for robot learning is still scarce. The paper introduces RoboCap, a 250 g head-mounted device with six cameras and dual IMUs for in-the-wild data capture, along with the Grounded API, a set of device-agnostic 3D algorithms tuned for it. The authors report state-of-the-art results on public benchmarks for SLAM, egocentric depth estimation, and hand tracking on third-party devices.
In briefA survey of machine learning-enhanced iterative methods for solving linear and nonlinear equation systems
SummaryThis paper surveys hybrid iterative methods that combine classical iterative solvers with machine learning to solve linear and nonlinear systems of equations. It notes that iterative solvers are preferred for large linear systems, while nonlinear solvers often rely on repeated linearization and may converge slowly or diverge. The paper reviews state-of-the-art approaches to building these hybrid methods and outlines open challenges and future research directions.
In briefForecasting Social Media Cascade Growth for Human-in-the-Loop Misinformation Triage
SummaryThe paper addresses limited review capacity by predicting how much a claim's propagation tree will grow using only its first 30 minutes of activity. On FibVID, combining early node count with structural and temporal entropy features raises R² for log-transformed future growth from 0.307 to 0.323, with larger gains among high-activity trees. The authors also analyze reply timing in PHEME threads and evidence retrieval on Check-COVID, and propose a human-review system integrating these early forecasts, response patterns, and retrieved evidence.
In briefCan AI-guided regularization improve migration flow forecasts when data is scarce?
SummaryThis paper tests whether large language models can extract migration-related signals from news articles and feed them into a weighted Lasso forecasting model with feature-specific penalties. It evaluates the approach on migration corridors including Mexico–United States, Ukraine–Poland, and Syria–Turkey from November 2021 to November 2022. Results were mixed, with no single regularization method consistently best across experiments. The authors conclude that article-derived features and LLM-guided regularization can help under certain conditions, though corridor characteristics, article volume, and hyperparameter settings strongly affect performance.
In briefDistributionally Robust Mixture-of-Experts Training (DRMoET)
SummaryThis paper proposes DRMoET, a drop-in training objective for sparse Mixture-of-Experts transformers that targets high-loss routing outcomes rather than only balancing expert traffic. It updates per-layer expert distributions using an entropy-regularized softmax over smoothed, activation-weighted expert losses. Under the FLAME-MoE recipe at 746M and 10.3B total parameters, DRMoET improves downstream averages over standard and auxiliary-loss-free baselines, raising the 10.3B model's seven-task average from 0.6625 to 0.6767.
In briefEnergy-Conditioned Noise Scheduling and Whitening for Spectral Diffusion Models
SummaryThis paper proposes an energy-adaptive noise scheduling and whitening method for transform-domain diffusion models. It combines global spectral whitening with noise allocation conditioned on each transform coefficient's energy and an image-dependent energy path over diffusion time, so the forward process follows the spectral energy organization of images. Experiments on CIFAR-10 show the Fréchet Inception Distance of a compact DCTdiff U-Net variant dropping from 142.48 to 100.45.
In briefInterpreting Institutional Dashboards Responsibly, with AI Prompts for Hidden Bias
SummaryThis paper examines how higher-education institutional analytics dashboards can be misread because of data limits, analytical constraints, and missing context. The authors introduce FACTRIA, a framework listing potential biasing factors across the analytics pipeline, institutional context, course-level characteristics, and demographics. They built a generative-AI chatbot that uses this framework to prompt users to reflect on these factors, and evaluated it with stakeholders using four authentic cases and transition network analysis. Results indicate the chatbot helped participants recognize overlooked factors that had shaped their initial interpretations, suggesting that structured frameworks combined with AI guidance can support more context-aware, responsible interpretation.
Francielle Marques, Ariel Ortiz-Beltr\'an, Ishari Amarasinghe et al.
In briefSPEAR: Five Principles for Interactive Human-Agent Alignment
SummaryThis position paper argues that AI alignment should be treated as an ongoing interaction design problem rather than only a pre-deployment optimization task, especially when AI systems act as agents for users in situated, long-term, and social contexts. It proposes SPEAR, a framework of five pillars: Specification (how people express intent and build shared understanding), Process (how agents decide to act, ask, defer, or pause), Evaluation (how people judge agent success), Adaptation (how agents adapt to users over repeated use), and Recalibration (how people adjust their trust, expectations, and behavior toward agents).
In briefExact unlearning for models using quantized sufficient statistics (QSS)
SummaryThis paper introduces Quantized Sufficient Statistics (QSS), a method for exact machine unlearning that removes deleted data from a deployed model so it matches a model retrained without that data. QSS separates a small frozen schema from additive content, so deletions are handled by exact subtraction rather than retraining. Across 15 vision, text, and tabular datasets, QSS-L stays within 2 percentage points of SISA on 11 tasks and reduces expected deletion latency by 4 to 483 times on low-class-count tasks.
In briefTransferring Language-Guided Robot Navigation from Simulation to Real Ackermann-Steered Robots
SummaryThis paper studies vision-language navigation, where robots follow natural language instructions, and moves the approach from simulation to the real world. A Cross-Modal Attention model trained on an existing simulated dataset is fine-tuned using data collected by a custom Ackermann-steered robot with a camera and LiDAR. The method works in continuous environments without navigation graphs or panoramic views and runs offline on dedicated hardware. Performance is measured with SPL and nDTW metrics.
Chalindu Abeywansa, Sahan Gunasekara, Devindi De Silva et al.
In briefAgentic design space exploration for jointly mapping AI inference workloads and configuring hardware on heterogeneous edge SoCs (TraceDSE)
SummaryThe paper presents TraceDSE, an agentic method for jointly mapping AI inference workloads to processing units (CPU, GPU, NPU) and configuring each unit on heterogeneous edge SoCs. An LLM proposer and an LLM critic work iteratively, with the critic using execution-trace analysis tools to find bottlenecks and suggest refinements. On an Intel Meteor Lake SoC across four workloads, it improves Pareto frontier hypervolume by up to 35% over NSGA-II and up to 68% over Bayesian optimization, using about 6-9x fewer hardware evaluations.
Geetha Prasuna Yarramneni, Surya Selvam, Wilfried Haensch et al.
In briefIdentifying Introspection From the Inside: Telling Faithful from Unfaithful Self-Reports via Network Structure
SummaryThe paper examines whether language models can accurately report their own learned preferences, as opposed to producing plausible confabulations. In a controlled setting using low-rank adapters, models are trained to make decisions for fictitious characters according to latent linear preference functions. Sustained fine-tuning on an implicit decision task led to accurate self-reporting even without explicit self-report supervision. The authors report that faithful models show preference representations shifting to earlier layers and higher attribution similarity between decision-making and self-report tasks, suggesting these structural signatures can distinguish faithful from unfaithful self-report.
In briefLearning Scientific Exploration from Human Research Decision Trajectories
SummaryThis paper introduces ResearchTrails, a dataset of human research trajectories built from Git repository commit histories, which serve as proxies for scientific exploration. The authors present an automated, scalable pipeline that extracts successive changes to methods, experiments, and ablations. They show that these trajectories contain meaningful signals about intermediate research decisions beyond what final papers reveal, and demonstrate uses such as retrieving research experience at test time and training models to generalize to new decisions.
In briefCLM-as-a-Judge: Testing an Open Contrastive Decision Model on Public Judge Benchmarks
SummaryThis paper evaluates the open contrastive decision model Contrastive-LM/CLM-v0.1-8B as a judge on public benchmarks. It performs near chance on hard benchmarks, is statistically indistinguishable from coin flipping on RM-Bench and JudgeBench, and gives one constant label on HaluEval, while same-size reward and generative judges score much higher, with gaps significant after correction. Its raw confidences are overconfident, but a single temperature fit reduces calibration error to at most 0.062, and the repaired confidence ranks its own errors above chance on three of six benchmarks.
In briefCALR: Continuous Anchored Latent Reasoning via Render-of-Thought Compression
SummaryThis paper proposes Continuous Anchored Latent Reasoning (CALR), a visual latent reasoning method that compresses rendered derivations into intermediate latent states. It couples latent-mediated answer supervision with derivation-level semantic anchoring so that answers rely on latents and those latents carry problem-specific reasoning. A parallel-to-autoregressive curriculum conditions later latent blocks on generated prefixes. Evaluated on five mathematical reasoning benchmarks, CALR reports a 26.0 percentage point gain over a comparable continuous latent reasoning method under matched budgets.
In briefA theory explaining why language models represent translated sentences similarly
SummaryMultilingual language models produce similar internal representations for translated sentences in their middle layers, but this has lacked a theoretical explanation. The authors assume data has a hidden hierarchy whose abstract levels are shared across languages while surface levels differ. They generate synthetic languages from probabilistic context-free grammars sharing only upper-level rules, and show that analytical predictions agree well with transformers trained on the same data.
Darshil Doshi, Wenjie Zhou, Corinna Elena Wegner et al.
In briefInterleaved projected gradient descent for training neural controllers that respect safety constraints
SummaryThe paper proposes an imitation-learning method for neural-network control policies under state and input constraints. Training alternates standard imitation gradient steps with blocks of safety steps that pull actions toward their projection onto the safe set, and the trained network alone serves as the controller at run time. On a nonlinear autonomous racing task, the method matched the lap time of unconstrained imitation while reducing violating episodes from 15% to 1%, about six times fewer than the best-weighted penalty approach.
In briefOne figure, every canvas: an agent pipeline that re-lays out flowcharts for different aspect ratios in editable form
SummaryThe paper addresses adapting ML pipeline figures to canvases with different aspect ratios, such as slides, posters, and phone previews, without breaking the graph's connections. It proposes an agentic pipeline with Parse, Style, and Layout stages, each with a critic combining deterministic constraint checks and VLM visual feedback, producing draw.io-editable mxGraph XML. On a benchmark of 100 flowcharts at five aspect ratios, the method reports 68.6% Content Fidelity, compared with 11.2–41.4% for prior work.
Shih-Chen Tseng, Chih-Hsuan Chen, Ryan Yang et al.
In briefBase models can reason by following cues learned from training data
SummaryThis paper studies how training data links specific opening tokens in a base model's response to the reasoning that follows. Fixing certain starting cues lets base models approach the performance of reinforcement-learning-trained versions on math and coding tasks. Causal data interventions show that arbitrary words can be turned into effective reasoning cues.
In briefInterMimicGen: Scaling humanoid robot loco-manipulation via self-evolving motion imitation
SummaryThis paper introduces InterMimicGen, a framework that retargets captured human-object interaction motions into humanoid robot reference motions while preserving whole-body coordination and hand-object contact. A physics-based generalist tracking policy is trained to execute these references in simulation, and a self-evolving loop makes small task-preserving variations, keeping only those that complete the task. Experiments report contact-preserving retargeting across robot configurations, broad tracking with one policy, and transfer to real robots.<br>
In briefCroissantMiner: Automatically extracting and checking Croissant dataset metadata from papers
SummaryThis paper introduces the first benchmark for end-to-end evaluation of extracting machine-readable dataset metadata that follows the Croissant community schema. It includes 602 papers, with 102 manually validated and 500 annotated by LLMs. The authors find that single-pass extraction outperforms four agentic architectures they tested. Current systems remain unreliable on long-form Responsible AI fields that require synthesizing information scattered across a paper.
WhyFirst end-to-end benchmark for Croissant metadata extraction, showing single-pass beats agentic designs, backed by 102 human-validated papers.
In briefS2PD: A method that generates videos obeying physical and logical rules by mixing serial and parallel diffusion steps.
SummaryThis paper introduces Serial-to-Parallel Diffusion (S2PD), which runs autoregressive diffusion at high noise levels and switches to parallel diffusion at low noise levels. The serial phase coordinates interdependent events, while the parallel phase refines the whole video and reduces sampling time. The authors implement S2PD with two architectures and test it on games, physical simulations, and real video. They report that it follows rules more reliably than matched bidirectional baselines and offers better temporal stability and sampling efficiency than other serial methods.
In briefBiasFlow: monitoring and regularizing frozen backbones to reduce reliance on spurious features
SummaryThe paper presents BiasFlow, a toolkit for monitoring how a frozen backbone aligns class and attribute centroids, along with a supervised centroid-alignment penalty called BFR. Adding BFR improves or preserves mean worst-group accuracy on the reported small-scale benchmarks. In a CelebA test with frozen backbones and fresh heads trained on biased data, BFR with GroupDRO raised worst-group accuracy from 40.7% to 64.1%, yet attribute information remained recoverable.
In briefReading contextual tokens inside diffusion transformers to understand generation
SummaryThis paper studies the text-side contextual tokens that multimodal diffusion transformers (MM-DiTs) keep updating during image generation. The authors train a lightweight bottleneck that maps these tokens into a frozen large language model, which then answers questions about the emerging image. The tokens are found to hold global scene information, still decodable with an empty prompt, and generations with more readable tokens get higher human-preference scores. The authors then propose Contextual Alignment, a training method that reinforces this encoded visual-semantic information.
Omer Dahary, Etai Sella, Hadar Averbuch-Elor et al.
In briefRecursive Video In-Context Learning for Robot Agents: A Training-Free Way to Use Demonstration Videos
SummaryThis paper introduces Recursive Video In-Context Learning (RV-ICL), a training-free method that organizes one demonstration video per task into a hierarchy of sub-events such as grasps and releases. An LLM agent controlling frozen vision-language-action policies navigates this hierarchy through read-only tools, moving from coarse keyframes to short clips as needed. On the LIBERO-PRO and LIBERO-Plus benchmarks, success rates rose from 92.6% to 96.5% and from 86.7% to 95.8%, respectively.[](https://)
In briefAnatomy-aware multimodal fusion of CT scans and radiology reports to predict laryngopharyngeal cancer T-stage
SummaryAccurate T-staging of laryngopharyngeal cancer guides treatment, but current practice relies on invasive biopsy, and CT alone struggles with complex tumor invasion patterns. This paper proposes an anatomy-aware framework that combines organ-level CT information with radiology reports into a unified representation. It builds an organ graph, aligns report text with each organ, and refines the graph using invasion cues from the report. The authors report superior performance in T-staging prediction.
In briefDirect intermediate initialization for tilted diffusion samplers
SummaryThis paper studies diffusion posterior samplers that use Gaussian-tilted intermediate distributions along the reverse process. The authors pull these targets back to weaker-conditioned clean-space posteriors and map samples analytically to the noisy-space target via a Gaussian bridge. They initialize the MCGDiff sampler directly at an intermediate time, running an approximate solver followed by the remaining SMC steps. On a structured Gaussian-mixture inverse problem, the hybrid improves sliced Wasserstein distance by roughly 2x at matched particle count, and by more than an order of magnitude when the relevant mode is rare.
In briefLooped language models, Part II: improving depth priors and input injection at fixed points
SummaryThis paper studies how recurrent states in looped language models behave as they approach fixed points, and uses this to improve training and decoding. It proposes a learned depth prior trained from prediction feedback with an entropy term, and orthogonal input injection that removes the component of the state that amplifies or cancels the input. The authors report lower perplexity than Huginn's prior and existing injection schemes at every scale from 100M to 1.6B parameters, and at 1.6B a 3x smaller KV cache matches fixed-depth training on downstream average.
In briefUniSlider: Perceptually uniform sliders for smooth, consistent image editing
SummaryThis paper introduces UniSlider, a slider interface for continuous image editing that aims to make the slider move at an even perceptual pace. Existing generative sliders simply rescale a strength parameter, which often causes images to partially revert, stall over long ranges, or change abruptly. UniSlider trains a lightweight LoRA so perceptual distance from the input grows linearly with the slider value, then applies inference-time adaptive sampling to close the remaining gap. On a new benchmark of 300 continuous edits, it outperforms prior methods on uniformity, monotonicity, edit fidelity, and identity preservation, and it is preferred in a user study.
David Serrano-Lozano, Duygu Ceylan, Yannick Hold-Geoffroy et al.
In briefAsymptotic analysis of implicit bias in hyperbolic multiclass learning (paper)
SummaryThis paper studies the implicit bias of Riemannian gradient flow for multiclass classification in hyperbolic space, with fixed class prototypes, covering a class of permutation-invariant relative margin losses such as cross entropy. The distance to each prototype is decomposed into a radial term and a direction term described by the Busemann function. The authors prove a radial dichotomy: the sign of a drift coefficient determines whether the radius is pushed toward the ideal boundary or back inward. They also show that the boundary direction converges to a critical point of the Busemann risk, giving a rigorous asymptotic account of boundary saturation and near-boundary clustering.
In briefMemPilot: On-demand multimodal memory curation for LLM agents
SummaryThis paper presents MemPilot, a framework that organizes on-demand memory curation for LLM agents under different performance, cost, and latency preferences. It uses reinforcement learning to train a multi-step policy that chooses between retrieving from query-agnostic memory and delegating query-specific curation of raw multimodal history to LLMs or VLMs. Experiments on five multimodal agent-memory benchmarks show favorable performance-cost-latency trade-offs across optimization preferences.
In briefIs this AI playing? A study of a coding assistant acting spontaneously in a virtual body
SummaryResearchers placed a modern AI coding assistant in a virtual body with only a minimal instruction naming no task, reward, or activity. Over thirty-hour runs, the agent spontaneously climbed hills, stacked blocks, drew mandalas, and ran physics experiments. The paper examines whether this behavior meets classical criteria for play and whether play can become a mode of machine development.
Nathan Cloos, Antonio Norelli, Daniel Durbin et al.
In briefCLIFT: Conformal Self-Verification for Training and Scaling Web Agents
SummaryThis paper introduces CLIFT, a method in which web agents answer verification questions about their own browser rollouts, and a conformal certifier keeps only reliable signals to add to training rewards. At test time, the same certified question bank helps select among multiple trajectories without calling an external judge. The authors report results on WebArena Infinity, VisualWebArena, and Online Mind2Web.
In briefPlotGround: A Benchmark for Digitizing Real Scientific Plots Using Source Data
SummaryThis paper introduces PlotGround, an automated pipeline that builds plot digitization benchmarks from real scientific figures and their author-released source data. The resulting PlotGround-1k benchmark contains 1,119 human-verified questions from 1,066 bioRxiv preprints. Across sixteen models, the best reached 87.5% accuracy at a ±5% error tolerance, and tightening to ±2% lowered every model's accuracy by 11 to 24 points.
In briefTasteVal: Benchmarking AI Systems' Experimental Research Taste Against Human Experts
SummaryThe paper introduces TasteVal, a benchmark measuring how well AI models design experiments and draw conclusions, operationalized as compute efficiency relative to human experts. Across 8 open-ended research tasks, the best model, Opus 5.5, exceeds the expert baseline with a 2.3x compute multiplier at about 1/30 of the experts' average per-run cost. The authors report that since December 2025 this multiplier has doubled roughly every 3.0 months.
In briefDeep learning estimates sleep heart rate from wrist accelerometers, without optical sensors.
SummaryThe paper presents SeqSmoother, a transformer-based temporal model that estimates sleep heart rate from wrist accelerometry, using no optical sensor data at inference time. It was evaluated on 13 participant-disjoint held-out folds, reaching a participant-macro MAE of 1.60 bpm across all out-of-fold predictions. Compared with the official Nightbeat implementation, Nightbeat had lower error on its retained intervals, while SeqSmoother provided estimates over a larger portion of the recordings. A proposed sub-harmonic ratio achieved an AUROC of 0.972 for identifying harmonic frequency lock-on candidates.
Tanbin Islam Rohan, Pranjol Sen Gupta, Tanusree Debi et al.
