In briefHow Communication Affects Group Detection of Hidden Adversaries in Self-Adapting LLM Societies
SummaryThe study examines the game Mafia, where an informed minority hides among an uninformed majority, to test whether communication helps groups identify adversaries. Using 7,416 games with societies of 8 to 100 claude-haiku-4-5 agents, it compares communication protocols against silence. Simultaneous broadcast improved adversary identification in all nine compositions tested, while turn-taking removed most of that advantage. The authors also find that rapid adaptation through inherited strategy notes does not necessarily help, so a protocol's value must be measured together with the adaptation it induces.
WhyExact zero-information baseline and matched experiments reveal how communication and adaptation jointly shape collective inference, though scope is limited to Mafia.
In briefCalibrated weak-supervision method for mapping post-harvest burned cropland when labels are scarce
SummaryThis paper presents a calibrated weak-supervision framework for mapping post-harvest burned cropland in Punjab, India, using Sentinel-2 spectral change, VIIRS active-fire data, and MODIS MCD64A1 as calibration references. The best model, a multilayer perceptron with the full optical feature set, showed modest pixel-level agreement with MODIS and weak cross-year stability. The authors conclude the framework suits district-scale burden assessment and hotspot screening, with limited support for exact scar boundaries or temporally stable annual mapping.
In briefCausal Improvement Graph: a persistent graph that guides automated optimization of agent harnesses
SummaryThis paper proposes the Causal Improvement Graph (CIG), a meta-harness framework that stores the state of ongoing improvements in a persistent graph. The graph links Evidence, Hypothesis, Intervention, and Outcome nodes, so earlier findings directly guide later harness revisions. According to the authors, CIG finds stronger harnesses than previous meta-harness baselines across various agent tasks and stays robust to the choice of task solver and proposer.
In briefOptimal query complexity for solving finite-sum monotone inclusion problems
SummaryThis paper proposes an oracle-optimal method for finite-sum monotone inclusion problems under mean-square Lipschitz continuity. It uses a switching regularization scheme that avoids the extra cost of restarting a variance-reduced solver at each stage. The authors also prove a matching lower bound, indicating the method's expected component complexity is optimal up to universal constants in this oracle model.
WhyRemoves the n log n restart cost via regularization switching and operator reuse, with a matching lower bound proving optimality.
In briefSparseCraft: An AI-Driven Hardware-Software Co-Optimization Loop for Sparse Accelerators
SummaryThe paper presents SparseCraft, which uses a language model in a closed loop to iteratively edit the RTL, memory configuration, and sparse-kernel schedule of a Gemmini-based accelerator. Each candidate is checked through legality, simulation, bit-exact output verification, and synthesis before it counts, with measured results fed back to the model. On a 512×512 GraphChallenge sparse-DNN layer, the reported results versus the block-sparse Gemmini baseline are 2.1x fewer cycles, 9.8x less off-chip traffic, and 22.8% less area.
Rajatabha Chakraborty, M P Samartha, Vedant Pahariya et al.
In briefCode2Games: Coding Agents Generating Playable Game Worlds
SummaryThis paper presents Code2Games, an agentic framework that generates a game world from a natural-language game intent, covering scene layout, gameplay objectives, interactive entities, and executable game logic. It first builds a base world in Blender, then adapts it to Unreal Engine 5, using compilation diagnostics, runtime feedback, and gameplay test results to fix inconsistencies. The authors introduce the GameCode4D benchmark with ten fixed prompts and evaluate results on visual quality, interactive fidelity, multimodal artifact quality, and playable-game quality. They report improvements over direct generation by coding agents and existing baselines.
In briefMask-guided KV cache eviction for faster block diffusion language models
SummaryThis paper addresses the large key-value cache that block diffusion language models keep and attend to at every denoising step, which limits memory and speed. It proposes MaskAhead, a training-free method that uses mask-query-based ranking to decide both which past tokens to use for the current block and which to evict. A quantized variant, Q-MaskAhead, works directly on low-bit KV entries. On long-prompt QA, MaskAhead cuts KV memory by 9.5x on average with a 1.2-point mean F1 loss, and Q-MaskAhead by 20.1x with a 2.3-point loss.
Gleb Molodtsov, Ekaterina Alimaskina, Evgeny Uskov et al.
In briefLLMs Above Diagnostic Tools: Unrealized Gains in Industrial Fault Diagnosis
SummaryThis paper examines whether large language models can reliably use external diagnostic tool outputs when acting as an integration layer across five industrial fault datasets. Conflicting external information often overturned the LLM's initially correct judgments. On the Tennessee Eastman set, the integrated output reached 77.43%, below the specialist alone at 83.33%. The authors argue integration layers should be compared with their stronger standalone component, not only with the unaided LLM.
In briefEVISKILL: An LLM Agent Framework That Grows Skills Using Replayable Evidence
SummaryThis paper introduces EVISKILL, a framework that lets LLM agents keep building reusable procedural skills from their interaction experience without updating model parameters. It organizes execution observations into Replayable Evidence Cards and links each proposed edit to the task contexts that support it, then verifies edits through targeted re-execution replay. Experiments on three interactive benchmarks across six LLM backbones show the approach is effective.
In briefSPACE-CLIPv2: Recovering Local Depth Geometry from Frozen CLIP Features
SummaryThis paper presents SPACE-CLIPv2, a depth decoder built on a frozen CLIP backbone that aggregates fixed local neighborhoods in CLIP token space to recover fine local structure. It predicts aggregation weights at selected decoder stages and injects the result through a gated residual update, with a high-pass branch preserving shallow local contrast. On NYU Depth V2 it improves over a matched SPACE-CLIP baseline, and zero-shot evaluation on iBims-1 further improves boundary and planar-geometry measures.
In briefJoint upper-bound coverage and route-choice utility: an empirical test
SummaryThis paper asks whether more accurate traffic forecasts or higher uncertainty coverage actually improve route decisions. Using road speed data from Beijing and Chengdu, the authors build offline proxy tasks and compare raw and jointly calibrated path time upper bounds. Joint coverage rises substantially, yet route choices under the calibrated bounds increase lateness and mean travel time. The authors conclude that coverage is not a surrogate for downstream route utility, and that the offline results do not establish online or causal benefits.
In briefGeoBridge-VLA: Adding Geometry Awareness to Vision-Language-Action Robot Models
SummaryThis paper introduces GeoBridge-VLA, a two-stage method that learns geometric features from a pretrained robot model's frozen visual encoder and uses them to predict actions. The first stage uses depth supervision, and the second adds a gated residual interface to existing visual tokens without a new image encoder. At deployment it needs only RGB images, robot state, and language, not depth. It reaches 70.9% success on LIBERO versus 60.0% for SmolVLA, and 74.0% versus 54.0% across four tasks on a physical ROBOTIS OMY robot.
In briefDART-ES: A Difficulty-Aware Method for Fine-Tuning LLMs with Evolution Strategies
SummaryThis paper proposes DART-ES, a method for full-parameter fine-tuning of large language models using Evolution Strategies (ES) with only forward computation. It estimates each problem's solvability from its pass rate across perturbations and builds a dynamic difficulty state to guide reweighting and replay of rare solvable samples, without an extra difficulty model or backpropagation. The authors report that DART-ES outperforms standard ES on several models and tasks, and uses less runtime and GPU memory than GRPO.
In briefWhy better traffic forecasts may not improve signal control: a layered diagnostic study of forecast-to-decision value
SummaryThis paper studies whether improved traffic demand forecasts lead to better signal-control decisions, using 29 days of reconstructed demand data from Xuancheng, China. It evaluates point forecasts, conformal intervals, dependence-aware scenarios, and closed-loop controllers in layers. Forecast error decreased, but on held-out test dates both causal forecasts and an event oracle increased queue vehicle-seconds. The authors conclude that forecast value depends on temporal observability, action identifiability, dynamics consistency, and objective alignment.
In briefTUD: Predictive-Uncertainty-Triggered Generalist Reasoning for Dual-System VLA Robot Control
SummaryDual-system Vision-Language-Action models pair a slow reasoning generalist with a fast action expert for real-time robot control. This paper proposes TUD, which measures how much re-predicted actions disperse across steps as a predictive uncertainty signal, and uses it to skip unnecessary generalist calls. On VLA-Arena, the authors report higher success at matched call budgets than other uncertainty baselines. In real-robot experiments, they report 75% fewer generalist calls than the strongest fixed-interval baseline, with higher success.
In briefLightVLN: a lightweight, history-aware framework for efficient drone vision-language navigation
SummaryThis paper presents LightVLN, a lightweight framework for aerial vision-and-language navigation that uses a compact 0.5B language backbone. It compresses each historical frame into a single token and reduces the current observation from 256 to 32 visual tokens. On the public OpenFly dataset it reports 50.93% Test-Seen and 36.14% Test-Unseen success rates, and 25.83% on AerialVLN-S Val-Seen. The authors also evaluate it onboard a DJI M350 RTK with a Jetson Orin NX in a reconstructed campus setting.
In briefLooking where the model says it looks: self-grounded attention for visual reasoning
SummaryThe paper proposes Self-Saliency, a method that trains vision-language models to align their visual attention with the image regions mentioned in their reasoning. A grounding model localizes the objects named in each reasoning step, and those regions supervise the model's attention. Across a unified suite of 25 visual reasoning benchmarks, the method outperforms prior attention-steering methods and image-level text grounding baselines.
In briefState-Aware Interaction MIL for Predicting Rare Joint Molecular Phenotypes in Colorectal and Lung Cancer
SummaryThis paper proposes State-Aware Interaction MIL, a weakly supervised method that predicts joint molecular phenotypes from pathology images. It keeps biomarker-specific histological representations, models their interaction, and supervises all four molecular-state configurations. Evaluated on BRAF+/MSI+ in colorectal cancer and EGFR+/TP53+ in lung adenocarcinoma using frozen UNI2-h and CONCH representations, it achieved higher average precision than the compared approaches.
In briefTopological Concept Representations for Operational Editing of Neural Network Concepts
SummaryThis paper introduces Topological Concept Representations (TCR), a post-hoc abstraction that jointly captures the concepts in a learned representation and how they relate. TCR builds an intermediate concept space from concept recoverability and interaction scores and encodes its organization through topology, with interventions made on this abstraction and propagated back to the model. Used as a preprocessing step for existing erasure methods, TCR improves worst-group accuracy by 21.89 on average at comparable concept leakage. It is also used to transfer concepts from teacher to student models, improving concept recoverability by up to 5.54 while keeping competitive test top-1 accuracy.
In briefPavement repair plan: a weighted constraint method using hazard density
SummaryThis paper addresses how road agencies turn a spatially distributed pavement distress survey into a budget-bounded repair-lot plan. Potholes are always repaired, cracks are included only when their benefit justifies cost, and historical patches serve as a re-degradation signal. The problem is modeled as a Weighted Constraint Satisfaction Problem over chainage rather than time, requiring only a single survey. In the case study, the recommended plan covers all potholes and 91.6% of cracks at 83.7% of the length needed to include every soft candidate.
Takato Yasuno, Keita Kobayashi, Ryuta Sakaguchi et al.
In briefMetaKernelBench: Measuring Cross-DSL GPU Kernel Knowledge Transfer
SummaryThis paper introduces MetaKernelBench, a benchmark that tests whether knowledge an agent distills from solving a GPU kernel problem in one DSL helps it solve the same problem in another DSL. It uses 74 fused-subgraph problems, each posed in CuTe DSL and TIRx variants. Across six models, paired lift over solo attempts ranges from -19% to +29%, with regressions on 16% to 45% of problems.
In briefEMBER-Bench: A Benchmark for Cross-Event Causal Memory in Long-Horizon Robot Tasks
SummaryEMBER-Bench is an egocentric benchmark of 189 household tasks and 699 question-answer pairs for testing whether embodied agents can reason about how past events affect current states and future actions. It evaluates next-action prediction and causal traceback in both directions. The best of 16 models reached 61.2% overall accuracy, compared with a mean of 98.3% for two human evaluators. The authors conclude that turning causal information from past events into constraints on current actions remains a key difficulty for long-horizon embodied agents.
In briefFreeLoc: Online indoor localization from floorplans using diffusion-refined poses
SummaryFreeLoc is an online RGB-based framework that localizes a camera within a building using its floorplan as a directly queryable map, avoiding dense offline databases. It retrieves candidate poses by querying floorplan rays on the fly and refines them into continuous pose estimates with a diffusion-aided method. For sequential localization, it builds online likelihoods that enable histogram-filter temporal fusion. The authors report real-time inference, state-of-the-art single-frame and sequential accuracy, and practical performance in real-world indoor robot tests.
In briefPortraitAes: Intent-based structured scoring of portrait aesthetics
SummaryThis paper presents PortraitAes, a system for scoring the aesthetic quality of portrait photos based on the photographer's intent. It routes each image to specialist assessments across defined photographic intents and dimensions, then verifies and fuses the results, with calibration to make scores comparable across images. On both a standard benchmark and a hard-case set, it reports higher Pearson and Spearman correlations than the general-purpose and specialized baselines evaluated.
In briefA framework using information theory to diagnose knowledge tracing benchmarks and models
SummaryKnowledge tracing models are usually judged by overall scores such as AUC and accuracy, which hide where errors come from and whether a benchmark is nearing saturation. This paper proposes an evaluation framework that uses Context Tree Weighting uncertainty as a shared coordinate and compares model gains across different entropy bands. Experiments on NIPS Task 3/4 and Algebra 2005 show that improvements are highly non-uniform, with modern models gaining mainly in high-entropy regions. The framework also flags apparent gains that need checks for noise sensitivity.
In briefMinimizing the largest movement when asynchronous robots form a uniform circle
SummaryThe paper studies the Min-Max Uniform Circle Formation problem, where point robots in the plane must move to distinct positions on a target circle forming a regular n-gon while minimizing the maximum distance any robot travels. It works under the ASYNC model with Look-Compute-Move cycles and non-rigid motion. The authors give necessary conditions for a deterministic solution and present deterministic, distributed, collision-free algorithms that form a circle and a uniform circle in finite time.
In briefPhysics-augmented graph transformers for forward and inverse patch-antenna design
SummaryThis paper proposes a mesh-native graph-learning framework that predicts patch-antenna radiation patterns faster than full-wave electromagnetic simulation. A graph transformer is trained with an auxiliary objective that predicts surface currents, improving accuracy at no inference cost. On an 80,000-sample CST benchmark, GPS+PAIS reaches MSE 0.17 and PSNR 19.67, and transfers zero-shot to canonical patches. For inverse design, surrogate-filtered diffusion achieves 32% lower relative MSE than nearest-neighbor retrieval.
In briefHow extreme is a corner case? Percentile control for highway scenario generation
SummaryThe paper represents the adversity of a generated driving scenario as its percentile within the conditional distribution of future risk given observed history, enabling calibrated control over how extreme a corner case is. It combines a percentile-conditioned joint diffusion model with sampling-time risk guidance to generate multi-agent futures. On highD, using minimum post-encroachment time (PET) as the risk surrogate, 1,422 of 1,440 requests were realized within a 0.05 percentile tolerance.
In briefEPOCH: Evidence-governed search for more reliable AI-driven discovery
SummaryEPOCH is an architecture for AI research agents that governs how evaluator feedback is interpreted and reused, using task contracts, typed memory, active falsification, admission checks, and independent replay. The authors report state-of-the-art aggregate performance on AlgoTune (mean normalized score 0.65 versus 0.53 for the strongest baseline) and the highest mean score on the internal Math14 suite. They also report concrete advances across ten discovery problems, and argue that evidence governance is needed for more trustworthy discoveries.
In briefVisualErase: Dual-branch visual trajectory redirection for robust concept erasure in text-to-image diffusion models
SummaryThis paper addresses concept erasure in text-to-image diffusion models, arguing that existing methods redirect only text-to-image mappings and can leave visual generative trajectories of target concepts intact, making erased models vulnerable to adversarial recovery attacks. The authors propose VisualErase, which uses structure-preserving image editing to build concept-removed counterparts and a dual-branch redirection loss that aligns both conditional and unconditional predictions with these targets, plus a retention loss matching a frozen pretrained model. Reported results show maximum attack success rates over seven attacks of 0%, 8%, and 0.1% for style, celebrity, and nudity erasure, respectively, while general generation quality is retained.
In briefProvably learning deep input-convex polynomial networks with active queries
SummaryThe paper studies L-layer polynomial networks with even-degree monomial activations and nonnegative higher-layer weights, which make the target network input-convex while keeping the parameter landscape nonconvex. The authors propose ASPIRE, a layerwise sampling-based diagonalization algorithm that uses input convexity and active queries to recover all parameters to a given accuracy. Its sample complexity grows only polynomially with depth, which the authors present as the first parameter-recovery guarantee for deep target networks.<br>
In briefEmoRSS: Reducing Over-Refusal Caused by Emotional Phrasing in Language Models
SummaryThis paper finds that emotional expression can make large language models refuse benign requests more often, causing unnecessary over-refusal. The authors propose EmoRSS, an activation-steering method that adjusts internal refusal-related representations at inference time without updating model parameters. In experiments on two LLMs, EmoRSS achieves a better balance than prior baselines between refusing harmful requests and answering benign ones when emotional expressions are present, while preserving general task performance.
In briefOffline learning of equilibria in games with public and private information, using pessimistic minimax methods
SummaryThis paper studies offline learning in two-player zero-sum contextual games where some information is public and some is private, as in auctions with private valuations. The authors introduce unilateral prescriptive concentrability and show that information asymmetry can change offline data coverage through equilibrium behavior. For finite state-action spaces, a pessimistic algorithm achieves an Õ(1/√n) exploitability rate. For general function approximation, they propose PPA-PMD, which obtains a unified Õ(1/√n + 1/√T) rate.
In briefCIPO: Counterfactual Imagination Policy Optimization for Adaptive Tool Granularity
SummaryThis paper studies how LLM agents should choose between atomic tools and composite skills when both are available. The authors propose CIPO, which mines executable skills from successful tool-use trajectories and trains granularity decisions by comparing outcomes of atomic and skill branches from the same state. Experiments across several benchmarks and model backbones report improved task success and decision efficiency. The analyses indicate gains come from better state-dependent choices rather than simply using skills more often.
In briefCrystalJev: calibrated foundation-model probabilities decide which materials to compute further
SummaryThe paper presents CrystalJev, which uses a single forward pass of a frozen atomistic foundation model on unrelaxed crystal structures to output calibrated stability probabilities, conformal guarantees, and a rule for escalating candidates to relaxation or DFT. Across 65 Matbench Discovery models, ordering slower computation by its decision value rather than probability recovered most of the benefit at a tenth of the cost. In a registered prospective test of 2,100 DFT calculations on ionic substitutions, forecasts calibrated on existing data overstated the stable fraction by 0.9 to 2.2 percentage points because the population had shifted, and a few hundred labels from the new pool removed this bias more reliably than label-free corrections.
In briefHow Long, Not How Close: A learned time-distance metric for planning in latent world models
SummaryThe paper proposes TEMPO, a planning objective that learns a small map on top of a frozen latent world model so that distances reflect the number of environment steps between states. The learned distance is blended into the planner's cost without modifying the world model, and it needs no rewards, policies, or success labels. Across eleven simulated environments with the LeWM and PLDM planners, TEMPO improves both planners at every goal distance, raising LeWM from 36% to 99% on TwoRoom three plans from the goal.
In briefTeleGen: Improving LLM Web App Generation with Runtime Telemetry
SummaryThis paper proposes TeleGen, a framework that instruments LLM-generated web applications, collects runtime telemetry during task execution, and compresses logs into concise briefs to guide code repair. On WebGen-Bench, task success rose from 67.7% without telemetry to 76.2%. On Web-Bench, cumulative Pass@2 improved from 21.7% to 29.8%. The authors report telemetry is especially helpful for failures involving navigation, form workflows, and frontend-backend coordination.
In briefGeometry-aware preference optimization for text-to-image diffusion models
SummaryThis paper revisits DPO-style preference alignment for diffusion models through the manifold hypothesis, arguing that Diffusion-DPO optimizes in the full ambient space and ignores the low-dimensional structure of natural images, which can gradually degrade quality and diversity. It proposes Anisotropic Geometry-Aware Preference Optimization (APO), which uses an anisotropic metric derived from the reference model to strengthen regularization where the reference denoiser is sensitive and relax it where safe semantic changes are allowed. The authors report an average win rate above 60% against various existing alignment methods, with fewer training steps and preserved generation diversity.
In briefRobust Tensor Completion via Reflective Convolution Nuclear Norm Minimization
SummaryThis paper addresses recovering full multidimensional data from partial observations contaminated by sparse gross errors. The authors propose RCNNM, which replaces circular shifts with endpoint-nonrepeating reflection to avoid artificial wrap-around neighborhoods in finite nonperiodic data. They establish high-probability exact recovery under random sampling and sparse corruption, along with stability under bounded dense perturbations. Experiments on synthetic tensors, BSDS color images, and CAVE multispectral images show gains over the circular-lifting method, especially near image boundaries.
In briefWhat post-training fixes: per-problem gains are shared across RL runs, and existing checkpoints predict them better than a priori signals
SummaryThis paper tests whether one can tell before training which problems a model will improve on. The authors evaluated eighteen post-training runs on two base models across up to 1532 competition math problems. Independent runs agree strongly on which rarely solved problems improve, but the two base models agree only weakly. Per-problem gains from an existing checkpoint of another model family predict a new run better than a priori signals such as base pass rate.
In briefDo LLMs Understand Sequential Structure? Controlled Tests of Inference and Generation
SummaryThis paper examines whether large language models can recover hidden sequential structure from context, rather than just matching surface action frequencies. The authors test this with controlled two-player Rock-Paper-Scissors games and a one-player stochastic n-gram continuation task. They find that longer context does not improve strategy identification, correct recognition does not ensure faithful simulation, and higher-order dependencies substantially hurt rule recovery.
In briefPriced Guidance: Testing Whether Language Models Can Generate Future Research Ideas
SummaryThe paper introduces Priced Guidance, a framework that measures how many extra bits of hints a language model needs to recover a target research idea, as a way to lower-bound the chance the model generates it unprompted. The authors evaluate five models on core ideas from 87 recent deep learning papers, using an LLM judge to check matches in research object and mechanism. Fable 5.1 had the lowest median compression cost at 69.9 bits, compared with 5,712 bits for gzip, and a uniform ensemble lowered it further to 55.8 bits.
In briefIREA: Aligning Embeddings via Intermediate Representations for Ethical Retrieval-Augmented Generation
SummaryThis paper proposes normative RAG, which lets large language models use external normative knowledge to support ethical judgment. It addresses the asymmetry between narrative situation queries and general normative statements by introducing IREA, a bidirectional method that maps both into a shared situation-behavior representation. Experiments reported in the paper show improved normative retrieval and downstream ethical judgment across multiple settings.
In briefTrajLong: Co-designing agent and long-context training data for mid-training
SummaryThis paper studies how to turn LLM agent trajectories into mid-training data that builds long-context skills. It introduces TrajLong, which compiles trajectories into long-context tasks with dense supervision for evidence grounding, cross-evidence aggregation, and temporal state maintenance. Qwen3-14B-Base and Qwen3-30B-A3B-Base were mid-trained on this data and then fine-tuned, yielding gains across 6 long-context and 12 agent benchmarks and outperforming raw and masked trajectory baselines.
In briefOne token may be enough: linking prompting and activation steering via prefix steering
SummaryThe paper asks whether activation steering applied over a short span at the start of generation can produce effects similar to prompting. Under fixed-state attention assumptions, the authors give sufficient conditions for single- and multi-token steering to match prompt-induced attention-head outputs. They propose Prefix Steering, which intervenes only over a short span beginning at the final prompt token. Across four models and five tasks, short interventions, even a single token, often retain much of full steering's control while better preserving general capabilities.
In briefBACAM: Behavior-Aware Continual Agent Merging for Multi-Turn Interaction
SummaryThis paper proposes BACAM, a method for sequentially merging incoming specialized agents into one model without keeping previously merged experts. It learns parameter-wise merging gates from candidate trajectories using expert-guided behavioral supervision, and uses task-level stability-plasticity control and tensor-level conflict-aware update budgets to limit interference. Across four interactive tasks (web shopping, tool use, information retrieval, and embodied interaction), BACAM reaches an average success rate of 62.82%, 21.69 points above the strongest evaluated merging baseline.
In briefMAGIC: A Topology-Aware Analytic Method for Few-Shot Class-Incremental Graph Learning
SummaryThis paper proposes MAGIC, a replay-free framework for graph few-shot class-incremental learning, which must learn new classes from very few labeled nodes without forgetting earlier ones. It combines a frozen graph representation backbone with closed-form analytic continual learning, uses a topological prior from the base graph with Potts Markov random field inference to reduce novel-class overfitting, and applies drift-aware analytic distillation to address representation drift from cross-session edges. Experiments on five datasets and eight baselines report improvements in mean accuracy, final accuracy, and performance drop under the 5-shot setting, along with reduced training time.
In briefBuilding LLM Agent Systems Like Deep Neural Networks: Modular Design and Architecture Search
SummaryThis paper proposes building large language model (LLM) agent systems modularly, like constructing deep neural networks. Retrieval, memory, and prompting strategies are treated as analogues of deep learning modules, and prompt optimization from feedback is likened to backpropagation. A search algorithm, analogous to neural architecture search, is used to find the best system configuration. The reported experiments show performance gains, including at least 5% from automatic prompt optimization and 11% from architecture search compared with randomly designed architectures.
In briefTrinity: one differentiable physics model for training, refining, and scoring generative chip floorplanners
SummaryThis paper studies floorplanning, which arranges chip blocks and their shapes under objectives like short wirelength and small outline, subject to constraints such as non-overlap. The authors re-implement four diffusion-based placers under one recipe and propose Trinity, a flow-matching floorplanner that uses six differentiable functions in its training loss, post-sampling refinement, and soft costs. On the FloorSet validation set, the pipeline reaches a mean hard cost of 1.014 at 1.63 s per case.
In briefFrom Overloaded to Guaranteed: High-Throughput Multi-SLO Enforcement for On-Premise LLMs
SummaryThis paper addresses how resource-constrained on-premise LLM servers, used by sectors such as hospitals, can meet heterogeneous service level objectives while serving multiple LoRA-adapted services. It proposes HALO, which spatially multiplexes GPU streaming multiprocessors to overlap base and LoRA computations and uses an SLO-aware scheduler based on request-level slack. The authors report that HALO reduces SLO violations and improves throughput compared with state-of-the-art baselines.
In briefKapture: Koopman-based learning to reconstruct ECG from radar more efficiently
SummaryThis paper presents Kapture, a framework that reconstructs electrocardiograms without contact using millimeter-wave radar signals. It models the evolution of radar hidden states in a low-dimensional space with a Koopman operator, and uses a temporal contrastive objective and reconstruction loss to suppress predictable interference such as respiration and body movement. Experiments on about 80 minutes of quasi-static radar-ECG recordings show consistent accuracy gains at matched backbone widths. The compact configuration uses 68.1% fewer parameters and 91.3% fewer FLOPs while approaching full-width accuracy.
In briefEDISCO: A Symmetry-Preserving Diffusion Model for Euclidean Routing Problems
SummaryThis paper introduces EDISCO, the first discrete diffusion model for Euclidean combinatorial optimization problems such as the Traveling Salesman Problem and Capacitated Vehicle Routing Problem. Its generated solutions are exactly invariant under rotations, reflections, and translations, using an E(2)-equivariant edge-score network with a categorical Markov chain over edge variables. The authors report that EDISCO outperforms previous learning-based solvers on TSP and CVRP with fewer training instances. It also performs better than baselines when trained only on uniform data and tested under spatial distribution shift and CVRP constraint-tightness shift.
WhyFirst discrete diffusion solver with exact E(2)-invariant distributions, beating prior learned solvers on TSP and CVRP with less data and validated under distribution shift.
In briefHow Teachers Should Be Prepared: RL on Student-Induced States for Distillation
SummaryThis paper proposes Prep-OPD, which applies reinforcement learning to a teacher model before on-policy distillation so it can adapt to a student's existing reasoning and correct errors. Training optimizes the teacher's continuations from fixed student prefixes, rewarded by final-answer correctness. On eight math reasoning benchmarks with a 4B teacher and 1.7B student, Prep-OPD improves average accuracy over standard OPD and Relay-OPD by 8.28 and 2.30 percentage points.
In briefUnified Dynamics Framework for RL and Classical Control of a Six-DOF Pipeline-Tracking ROV
SummaryThis paper presents a pipeline-tracking system for a six-degree-of-freedom underwater ROV, using one USD scene to supply BlueROV2-Heavy physical parameters to both a NumPy implementation of Fossen's marine-craft equations and an NVIDIA Isaac Sim deployment. Five reinforcement learning algorithms and six classical controllers are compared under the same environment and guidance geometry. Under Coriolis-enabled dynamics, PPO, TRPO, and feedback linearization reach 100 percent success with competitive accuracy, while DDPG and TD3 show specific failure modes.
Cheng Siong Chin, M. Venkateshkumar, Jianhua Zhang
In briefTransferring EHR Representations Across Countries Using Asymmetric Contrastive Learning
SummaryThis paper addresses the difficulty of transferring longitudinal electronic health record representations across institutions and countries with different clinical coding and workflows. It proposes Asymmetric Supervised Contrastive Learning, a pre-training objective that clusters patients sharing a positive outcome without pulling negative trajectories together. Temporal Transformer encoders were pre-trained on records from 3.98 million patients in Taiwan's NHIRD and transferred to MIMIC-IV and EHRSHOT. The method achieved higher mean AUPRC than direct supervised transfer on all four tasks and than Standard SupCon on three, with a 0.003 deficit on readmission.
In briefTempoBridge: Source-conditioned flow matching for predicting how cell populations shift under time, gene, or drug changes
SummaryDestructive single-cell measurements give unpaired population snapshots, not the same cells across conditions. TempoBridge uses a source-conditioned flow-matching model that transports source cells under temporal, genetic, or chemical transitions. Minibatch optimal transport supplies training couplings, while inference needs neither target expression nor optimal transport. The paper reports better held-out performance than several baselines across unseen donors, gene combinations, and compounds.
In briefNear-optimal query complexity for finite-sum nonconvex-strongly-concave minimax problems
SummaryThis paper characterizes, up to logarithmic factors, the minimax expected query complexity of finite-sum nonconvex-strongly-concave optimization under mean-squared averaged smoothness. A dual-chain construction gives a lower bound that matches the existing Catalyst upper bound when κ≥√n. For 1≤κ≤√n, the authors propose recursive proximal descent-ascent (RPDA), which achieves a matching-order upper bound. Under individual smoothness, a lower bound is also proved for all κ≥1, while the intermediate condition-number gap remains open.
WhyCharacterizes near-optimal complexity for finite-sum nonconvex-strongly-concave minimax with new small-condition-number upper bound and matching lower bounds, rigorously proven.
In briefZero-shot time-series QA by separating perception from reasoning
SummaryThis paper presents TSHarness, an agentic framework for time-series question answering that works zero-shot across datasets. It separates numerical perception from semantic reasoning, extracting statistical and temporal features into a structured Time-Series Perception State. The answering agent reasons over this state and triggers re-perception when evidence is insufficient. The approach requires no target-side training or answer feedback.
In briefSoftware World Models: Predicting Which Components a Code Change Will Break
SummaryThe paper introduces the Software World Model (SWM), which predicts the blast set, meaning the downstream components a code change will break. It works in three stages: executing candidate changes from restored system states, fine-tuning a language model on the outcomes, and converting predictions into per-consumer break probabilities for ranking and migration decisions. On held-out synthetic systems, blast-set F1 rose from 0.431 for static reachability to 0.571 ± 0.037. Experiments on real libraries also indicate that predicting structured failure outcomes matters for decision quality.
In briefA graph-based tool for checking whether a physics model's internal structure stays meaningful
SummaryThe paper introduces GIIT, a tool that maps the governing physics of a scientific neural network onto its internal components and tests that mapping through targeted interventions. When evaluated on temporal extrapolation beyond the training domain, the internal layer correspondences drift even before the training domain is left. This drift occurs before conventional output error metrics show severe degradation.
Adwait Patkhedkar, Alifaraz Lakhani, Prathmesh Mohite et al.
In briefD-DOIT: Training-free Adaptation of Discrete Diffusion via Doob's h-Transform
SummaryThis paper proposes D-DOIT, a training-free method for adapting discrete diffusion models to generic reward functions. It uses a discrete Doob's h-transform that reweights reverse transition probabilities and needs only reward values, not gradients. On regulatory DNA design and protein inverse folding benchmarks, the authors report that it outperforms training-free guidance baselines.
In briefDynamically Selecting Level-K Policies Against Unknown Opponents in Partially Observable Games
SummaryThis paper studies how to choose among a library of pretrained level-K policies when an opponent's reasoning level is unknown during a repeated, partially observed game. The authors frame deployment as dynamic orchestration and compare classification-based orchestrators with a reinforcement-learning orchestrator trained to maximize expected discounted return. In pursuit-evasion experiments, the reinforcement-learning orchestrator achieved higher return than the classification-based orchestrators and reached performance comparable to a directly trained policy with fewer training timesteps.
Addison Kalanther, Sanika Bharvirkar, Daniel Bostwick et al.
In briefLearning when minority cases are missing from specific regions of the data
SummaryThe paper argues that standard class-imbalance methods assume minority cases are uniformly undersampled, whereas in practice they may be under-observed in particular regions of the feature space, a problem called localized minority imbalance. The authors propose a tree-based stratified ensemble that recursively partitions the feature space to reduce within-stratum distortion. Experiments on simulated benchmark tabular datasets show it outperforms popular and state-of-the-art methods, and that hold-out evaluation on the same biased data can substantially mislead results.
In briefDiVeR: Decision-critical verifier learning for test-time scaling of robot action models
SummaryThis paper proposes DiVeR, a method for verifier-guided test-time scaling of Vision-Language-Action (VLA) robot policies. It estimates how decision-critical each state is from the dispersion of sampled action representations and uses this to reweight verifier training toward states where action choice matters most. The authors report consistent task success gains on LIBERO, RoboCasa, and real-world experiments with a Franka Research 3 robot, with negligible verifier inference overhead.
In briefAutomatically Choosing Better Thumbnails for Billions of Short Videos Using Bandit Learning
SummaryThis paper presents a real-time system that optimizes thumbnails for short-form videos on a major global platform, where many uploads have no creator-chosen cover image. It uses a multi-stage candidate pipeline and an online multi-armed bandit framework, initialized with image quality priors, to dynamically select thumbnails. The authors report statistically significant gains in user discovery and engagement after global deployment.
In briefStudy of prompt choice and noisy rewards when training a 1B model with GRPO on math problems
SummaryThis paper runs controlled experiments training the 1B-parameter OLMo-2-0425-1B model on GSM8K with GRPO, a reinforcement learning method. It finds the prompt format is the main factor, with a 3-shot prompt reaching 18.3% while zero-shot leaves the base model at 0.08%. In off-policy training, clipping recovers the lost performance, and a weak verifier hurts reinforcement learning far less than test-time selection.
In briefAssembling Reusable Insights to Improve Agentic Machine Learning Engineering
SummaryThis paper studies how insights drawn from related machine learning engineering tasks affect AI agents that build ML solutions. The authors introduce MLE-InsightBench, a benchmark of 160 Kaggle competitions across 12 domains with a held-out target per domain to prevent data leakage, and MLE-InsightForge, an agent that builds domain, competition, and improvement insights. Injecting these insights improved the normalized private-test score by 123.9% over a no-insight baseline, and the three insight types were found to be complementary.
In briefTSAE: a structured sparse autoencoder for interpreting time-series forecasting models
SummaryThis paper introduces TSAE, a structured sparse autoencoder that splits a forecasting model's hidden states into separately inspectable features, organizing cross-variable structure with shared and variable-routed private dictionaries and constraining changes in sparse codes between neighboring time steps. In three-seed experiments with a frozen PatchTST on ETTh1, ETTh2, and ETTm1, TSAE achieved the lowest hidden-state reconstruction error, NFRE, and feature transition rate among five sparse autoencoders. Tradeoffs in selectivity, physical correlation, and calibration vary by dataset, so the authors note that independent semantic validation is needed.
SummaryThis paper studies how large language model (LLM) agents can work more reliably within HomeAssistant, a popular open-source smart home platform. The authors combine supervised learning with optimized prompting and test several LLMs of differing reasoning ability and cost on the hardest tasks in a smart home benchmark. They find that ReAct and Reflexion provide only marginal gains over fine-tuned LLMs for multi-device control, but show promise for resilience tasks such as failure response.
In briefTRACE: Training-free diffusion style transfer using time-adaptive attention and content-style separation
SummaryThis paper proposes TRACE, a training-free diffusion framework for style transfer that applies a reference image's visual style to a content image while keeping its structure. It first separates content and style directions using offline CLIP subspace analysis, then removes cross-image leakage during inference. Style is injected through residual cross-attention, with guidance adjusted at each denoising step. The authors report a better balance between style fidelity and content preservation than the compared baselines.
In briefComplex Agents, Shallow Tests: How Well Agent Harnesses Are Tested
SummaryThis paper presents the first empirical study of test adequacy for agent harnesses, the software infrastructure supporting LLM-based agents, in real-world systems. It finds these harnesses are substantially undertested, with LLM-dependent harness code having less than half its lines and branches covered. The authors propose HarnessTester, a test generation technique that uses explicit agent-harness contracts, and report that it outperforms general-purpose methods and detects 122 real-world bugs.
WhyFirst systematic study of harness test adequacy, with a harness-oriented test generator showing large coverage gains and 88 previously unknown bugs found.
In briefPWM: Personalized World Models via Online Reinforcement Learning
SummaryThis paper introduces Personalized World Models (PWM), a method that customizes interactive world models using short scene videos and online reinforcement learning. It learns a specific scene's visual identity while preserving action-conditioned generation quality. The authors also release PWM-Bench, with 150 tasks across Indoor, Outdoor, and Gaming domains. The GRPO and DiffusionNFT versions improve over native Yume in 71.3% and 65.3% of evaluated scenes, respectively.
In briefResidual Visual Credit Optimization: Conserved Evidence Routing for Multimodal Reinforcement Learning
SummaryThis paper proposes Residual Visual Credit Optimization (RVCO), which treats per-token credit in multimodal reasoning as a conserved routing problem. A controlled visual intervention produces token-level evidence responses, and a budgeted entropic router distributes a fixed sequence utility according to perceptual dependence, with a residual path ensuring positive credit at every valid position. Across four model families and seven reasoning benchmarks, RVCO reports higher accuracy than strong RLVR baselines while keeping late-stage stability and competitive training cost. Rewards, rollouts, and the group-relative advantage estimator are unchanged.
In briefPreAct-Nav: Agentic Reasoning Before Acting for City Navigation
SummaryThis paper proposes PreAct-Nav, an agentic navigation framework that gives frozen policies anticipatory reasoning for urban navigation. A navigation memory keeps a medium-horizon subgoal and turns distant goals into intermediate objectives, while an action-conditioned world model predicts the outcomes of candidate moves in a sandbox. A vision-language model uses these predictions to retain or revise actions, and after execution the memory is updated with real observations. The authors report that gains are larger on longer routes and routes with more turns.|
In briefAre AI Scientists Really Reasoning From Data? A Test of Evidence Grounding
SummaryBenchmarks for AI agents doing data-driven science assume correct answers come from the supplied data, but agents may use prior knowledge or elimination instead. The authors edit data files to keep, withdraw, or reverse the evidence and measure evidence-grounded accuracy. On single-cell questions, Claude agents scored 95% accuracy but only 41% evidence-grounded accuracy.
WhyProposes evidence-withdrawal and reversal tests showing high agent accuracy often stems from prior knowledge, a novel and well-supported evaluation critique.
In briefWhether reasoning models willingly make themselves monitorable and self-report misbehavior
SummaryThis paper studies whether large reasoning models are willing to make themselves monitorable during inference by self-reporting their misbehavior through tool calls. Four models were evaluated on sycophancy, reward hacking, and bias under varying monitors and tool-use pressure. When tool use was optional, models self-reported in only about 16% of warranted cases, and high-severity misbehavior was never self-reported.
In briefVideoResearchAgent: Training AI Agents to Research Videos from the Open Web
SummaryThis paper presents VideoResearchAgent, a framework for training agents that autonomously find relevant videos online, navigate their timelines, and answer questions based on visual evidence. It combines synthetic multi-hop task generation, a local video simulator aligned with real search and watch interactions, and a training method (RDR-GRPO) that randomizes retrieval conditions to reduce overfitting. On Video-BrowseComp, an agent trained from Qwen3.5-4B reaches 40.48% accuracy, comparable to Gemini-3-Flash-Preview, while cutting cumulative API-token use by 74.9% relative to the untrained model.
In briefWhy subliminal learning needs so much data: a noisy inverse view via steering vector recovery
SummaryThis paper studies subliminal learning, where a student model inherits a teacher's behavioral trait from semantically unrelated data, which usually requires tens of thousands of examples. Using a known residual-stream steering vector as a testbed, the authors compare hard token-level NLL supervision with soft full-distribution KL supervision. Soft supervision recovers the trait almost exactly from a few hundred carriers, while hard supervision stays well below it even with tens of thousands. The authors frame this gap as a noisy linear inverse problem and report experiments on Qwen2.5-7B and Gemma-2-9B supporting their predictions.
WhyAttributes subliminal learning's data needs to noisy Fisher inversion, testing four predictions via direct steering-vector recovery on two models.
In briefScaling Verifiable Training Environments for Long-Horizon Work Agents
SummaryThis paper introduces WorkForge, a framework that automatically builds verifiable training environments for work agents from real-world files. It starts from expert workflows, assembles relevant files into a workspace, extracts checkable facts from it, and derives tasks with programmatic and semantic verifiers. The authors built 16.7K environments across 40 professional domains and report that post-training Qwen3.5 models improved their GDPVal and APEX Score results.
In briefScopeSAE: Choosing the Right Model Subspace to Find More Useful AI Features
SummarySparse autoencoders are a standard tool for interpreting neural networks, but existing versions are trained layer by layer. This paper proposes ScopeSAE, which uses gradient-based attribution to pick, for each token, the most influential token-layer states and learns features over that prediction-relevant subspace. The authors report that reconstructions from ScopeSAE give lower next-token cross-entropy than the original activations, and interventional analyses and a KL fine-tuning control suggest this comes from the subspace itself. ScopeSAE also improves effective feature count, interpretability, utilization, and dictionary redundancy over layer-based baselines.
In briefSelf-Evaluating Recursive Agents (SERA): Teaching One Model to Judge Its Own Subtasks
SummaryThis paper studies recursive language-model agents that split tasks and delegate subtasks to copies of themselves. It proposes SERA, which trains a single set of weights to decompose problems, solve subtasks, and evaluate outcomes, with the parent writing weighted success rubrics for each child and training them against verified results. On TextCraft-Synth and TextWorld-Sync, SERA improves over strong recursive-agent baselines by 5.38 and 13.14 points on average.
In briefUnsupervised Transmission System Monitoring Using End-of-Line Test Data: A Ford Motor Company Case Study
SummaryThis paper proposes a multivariate monitoring framework for high-dimensional data that preprocesses and temporally aligns the data, then applies nonlinear dimensionality reduction and control-chart-based Phase I anomaly detection. It can be used in unsupervised or supervised settings. Evaluated on real production data at Ford Motor Company, the method raised accuracy, recall, and F1 score from 0.50, 0.30, and 0.429 to 0.625, 1.00, and 0.769 compared with the company's existing model.
Mohammad N. Bisheh, Mehrdad Moradi, Parinaz Farajiparvar et al.
In briefAdaptive multilingual query rewriting that improves reasoning via agentic reinforcement learning
SummaryThe paper notes that semantically equivalent queries in different languages can lead language models into different reasoning paths, causing performance gaps. The authors propose mRewriter-R1, an agentic framework that treats multilingual query rewriting as a multi-turn decision process and adaptively selects rewriting operators. Experiments show it outperforms strong multilingual rewriting baselines across several large reasoning backbones, and the learned policy generalizes across reasoning tasks and works plug-and-play with different models.
In briefHow Well Language Model Surprisal Predicts Chinese Reading Times: A Systematic Analysis
SummaryThis paper examines how well token-level surprisal from language models predicts Mandarin Chinese reading times. The authors propose the Shortest Matching Sequence (SMS) alignment scheme to reconcile eye-tracking word segmentation with subword tokenization. Using Chinese-Pythia models from 14M to 1.4B parameters, they find surprisal predicts Chinese reading times, but whether predictive power scales with size and training is corpus-specific, with inverse scaling appearing in HKP and MECO.
In briefSFlexRCA: A Lightweight, Scalable Root Cause Analysis Method for Industrial IoT Edge Systems
SummaryThis paper proposes SFlexRCA, a root cause analysis framework for industrial Internet of Things systems that does not require topology information and suits resource-constrained devices. It converts multivariate sensor data into compact orthogonal representations and uses shared lightweight linear modeling, avoiding explicit graph construction and parameter growth. The method is evaluated on the BATADAL, SWaT, and WADI datasets against 10 baselines, measuring accuracy, training efficiency, inference latency, and memory use, with deployment tests on Raspberry Pi devices.
In briefForkPilot: a self-updating policy for deciding when agents should search backward
SummaryThis paper studies how language-model agents running many tool calls can misuse retrospective search, since delayed outcomes make it hard to judge which search decisions helped. The authors introduce Search Value Dynamics (SVD) to describe how the gain and cost of such search change over time. They propose ForkPilot, a two-stage framework that learns a search-value policy offline from completed trajectories and then keeps updating it as new trajectories arrive. Across 6 benchmarks and 7 LLM backbone families, it reports comparable state-of-the-art performance while cutting token usage by up to 59.2%.
Xinyue Zeng, Shivam Shandilya, Guilherme Potje et al.
In briefLeakage in LLM Fact-Checking: Measuring Future Information in Misinformation Detection
SummaryThis paper examines why large language model fact-checking scores on social media claims may be overstated. The authors rebuild the evidence available at each claim's time on the AVeriTeC and QuanTemp++ benchmarks and probe model representations for outcome information. They measure two leakage channels: outcomes memorized during pre-training and retrieved evidence published after the claim. They report substantial parametric leakage that aggregate accuracy can hide, and that allowing post-claim evidence raises zero-shot accuracy on AVeriTeC, while the effect is negligible on QuanTemp++.
In briefReflection-robust 6DoF object tracking using light fields
SummaryThe paper presents a light field based method for tracking the 6DoF pose of moving objects, treating reflections on shiny surfaces as a pose cue rather than noise. Per frame, it robustly estimates depth and surface normals, then splits view-dependent appearance into diffuse albedo and a reflected environment map to form a relightable surface light field, and optimizes pose with a photometric loss. Evaluated on a dataset re-rendered at four reflectivity levels and on two captured sequences, it trails the strongest baselines on diffuse objects but is the only method that keeps its accuracy on fully reflective objects.
In briefResOPD: Tail residualization for sparse on-policy distillation
SummaryThis paper addresses high gradient variance when training a student model with sparse teacher signals in on-policy distillation. Sampled-token estimators are unbiased but noisy, while Top-k objectives are biased. The authors propose ResOPD, which groups unobserved vocabulary into a coarse tail event, computes its exact aggregate gradient, and samples only the within-tail residual, without extra teacher queries. The paper reports substantially reduced gradient variance, more stable training, and improved downstream performance in the evaluated settings.
In briefSpecFold: Folding Multi-Branch Redundancy to Speed Up Speculative Decoding in Diffusion Language Models
SummaryThis paper targets speculative decoding in diffusion large language models, where multiple draft branches are verified together and their hidden states are highly similar. It proposes SpecFold, which uses token-level residual gating and reuses parent computation via folded attention and FFN, implemented with a Triton kernel for sparse multi-branch execution. Across two model families, five models, and five benchmarks, it reports up to 1.64x throughput over Spiffy and up to 1.99x over vanilla decoding, with comparable task performance.
In briefWhen the noncommutative AM-GM inequality holds
SummaryThe paper proves that the noncommutative AM-GM inequality holds whenever n ≥ 2⌈m/2⌉². It is motivated by counterexamples from De Sa's NeurIPS 2020 work. The proof builds a vertex measure on the Boolean cube from Chebyshev nodes to extract distinct indices, and its main difficulty is that this measure is not positive at non-integer nodes.
In briefTemporal Link Predictors Miss Source Activity; a Transferable Fix
SummaryThe paper argues that temporal link prediction focuses on which destination is chosen and ignores whether the source node is active, because standard negative sampling cancels this factor. The authors propose Source Node Activity Modeling (SNAM), a self-exciting intensity model with fewer than 20 parameters. Added to the frozen scores of several existing models, it raises average precision on most backbone-dataset pairs under the reported settings.
WhyProves standard negatives cancel source activity, then shows a tiny transferable activity model improves several backbones across many datasets.
In briefFrom Memory to Guide: A Spatio-Temporal Composer for Procedural Coding Memory
SummaryThis paper proposes a paradigm that turns an agent's procedural memory from passive text prompts into active, inference-time policy adaptation. The authors implement it as the Spatio-Temporal Composer, which adapts retrieved knowledge to current environmental constraints and controls when it applies. On 13 long-horizon software engineering tasks in EngramBench, it raised the pass rate on the most complex tasks by 7.2 percentage points and cut the main agent's token usage by 32.2%.
In briefGreedy local learning for language model pretraining: performance gaps and objective design
SummaryThis paper studies greedy block-wise local learning, which splits a language model into gradient-isolated blocks trained with local auxiliary losses, and compares it with end-to-end training at 125M and 400M parameters. The gap to end-to-end training more than doubles from 2 to 4 blocks but shrinks at 400M. Replacing an MLP auxiliary with a Transformer-based one recovers 22-41% of the gap, and a multi-token-prediction auxiliary helps only at the first block boundary.
Jihwan Moon, Sheir A. Zaheer, Jinmyoung Lee et al.
In briefImproving long-video VLM embeddings by accumulating query-relevant evidence across clips
SummaryThis paper introduces QASLR, a post-training method that builds fixed-size video embeddings by processing selected frames clip by clip and accumulating query-relevant evidence. It aims to avoid the miss-prone uniform sampling and the memory cost of encoding many frames at once. Under the full training recipe, HourVideo retrieval Hit@1 rises from 54.2 to 70.7 for a 2B backbone and from 57.4 to 72.8 for an 8B backbone. The authors report gains on moment retrieval and video QA, and transfer to a second Qwen-family embedding model.
In briefGitSwarm: decentralized system that lets AI agents build on each other's work
SummaryThis paper proposes compounding inference, where intermediate results from AI inference persist and can be inspected, extended, or challenged later. The authors implement GitSwarm, in which homogeneous agents work asynchronously in a shared, branchable Git repository with atomic commits and recorded semantic dependencies. The paper reports results on IMOProofBench-Advanced, ProgramBench, and three neural architecture research tasks, and measures how much contributions are subsequently built upon.
In briefOne Tile, Multiple Instances: Rethinking MIL for Sparse Diagnostic Evidence
SummaryThis paper addresses a limitation in weakly supervised whole slide image classification, where each tile is compressed into one global embedding, hiding fine-grained sub-tile evidence from attention-based aggregators. The authors propose DI-MIL, which clusters dense spatial tokens from a frozen foundation model within each tile to create multiple independently weighted instance embeddings, without re-encoding. Evaluated on cytopathology datasets, DI-MIL improves performance in most metric-level comparisons.
In briefCURIO: Curiosity-Driven Test-Time Learning for Open-Ended Discovery
SummaryThe paper presents CURIO, a framework that adds curiosity-based intrinsic rewards to task feedback, using a world model that predicts hidden-state transitions to encourage exploration of low-reward but potentially promising directions. Tested on six mathematical discovery tasks and single-cell denoising with Qwen3 models from 8B to 235B parameters, it reports gains over a task-only reinforcement learning baseline on five mathematical objectives and improvements in denoising Score and MSE at every tested scale. The authors also report greater structural diversity among generated programs.
WhyProposes curiosity-driven world model giving prediction-error bonuses, outperforming matched task-only RL on six math tasks and single-cell denoising across scales.
In briefPIT-GCL: Topology-Aware Contrastive Learning for Predicting Protein Binding
SummaryThis paper presents PIT-GCL, a dual-tower model that predicts whether two proteins bind, using each protein's sequence, C-alpha 3D coordinates, and a persistent homology summary of its global structure. Each protein is encoded independently by a structure-aware Transformer, and a bidirectional cross-attention module combines the two representations, trained with binary cross-entropy and contrastive losses. On three benchmarks, it outperforms the compared baselines on general PPI and is the only method above chance on PPB-Affinity, while on TCR-pMHC it is outranked by a task-specific sequence model. Because per-protein embeddings can be precomputed and reused, the approach suits large-scale screening.
In briefWhich preference mixes should a single AI model be trained on? A minimax method for multi-objective alignment
SummaryThe paper addresses multi-objective alignment of large language models, where goals such as helpfulness and harmlessness trade off against each other. It proposes MAESTRO, which frames the problem as a minimax optimization over preference distributions and trains one prompt-conditioned policy with reinforcement learning against an adversarial Dirichlet distribution updated by online mirror descent. The authors report that on HH-RLHF, BeaverTails and a summarization task, MAESTRO achieves the best Pareto front on most tasks in a single training run at the lowest training cost among compared methods.
