In briefLEV: A looped decision model that thinks deeper without generating text
SummaryThe paper presents LEV, a decision model that answers typed questions with one probability per option in a single pass, and deepens its reasoning by looping its layers instead of generating tokens. On three looped backbones, accuracy rises with loop depth using the same weights, mostly on hard questions. A per-question stopper that chooses depth saves 15% to 24% of computation with under one accuracy point lost.
WhyApplies looped layers to decision models with per-question depth selection, validated across several backbones, benchmarks, and dense-model comparisons.
In briefRanked well but sized wrongly: ensemble spread quietly fails at extreme streamflows
SummaryThis paper tests whether ensemble spread from deep learning streamflow forecasts, such as LSTM models, reliably signals uncertainty. At higher flows, prediction error is roughly 2 to 4 times the corrected spread, and for observed extremes beyond the training record the gap reaches 18 times. Spread ranks errors nearly as well as an ideal model, yet it systematically understates risk, most severely for the most extreme events.
In briefMAI-Gym: An open platform for evaluating and reinforcement-learning-training GUI agents on mobile and computer
SummaryThe paper presents MAI-Gym, an open platform that unifies GUI agent evaluation, data collection, and online reinforcement learning across mobile, computer, and web environments, scaling to 20,000 parallel instances. The authors train MAI-Gym-9B, a 9B-parameter agent, which reaches 78.4% on MobileWorld and 76.2% on OSWorld-Verified. They also report that training in a mocked train-ticketing app improves success on a physical phone from 28.8% to 48.1%.
WhyUnified GUI agent evaluation and RL platform with synthesized tasks; benchmarks and real-phone transfer results support solid evidence and moderate impact.
In briefRAC: A contract that lets agents specify relations while tools handle exact engineering details.
SummaryThis paper introduces RAC, a contract for the action and feedback layer of engineering agents, where the agent chooses which relations an artifact should satisfy and specialized tools derive the concrete results. On PCB placement, routing, and repair tasks, RAC reached an average clean-pass rate of 90.9% across nine models, versus 74.0% for the strongest baseline. In job-shop scheduling, direct relational actions reached 96.3% quality pass compared with 59.6% for code-mediated delivery.
In briefPrune the Summary, Not the History: Source-Judged Compaction for Long-Horizon Agents
SummaryLong-horizon LLM agents compress their interaction history into natural-language summaries when the context window fills up. The authors argue these summaries mix environment-verified facts with the compressor's own reasoning, and that many failed runs trace to summary entries unsupported by environment output. They propose Source-Judged Compaction (SJC), a training-free step that checks each summary entry for a source and removes unsupported ones. On AppWorld and OfficeBench it improves cross-run reliability by 2.1%-8.9% without raising the agent's token usage.
In briefSkillNEO: Non-elitist evolutionary optimization for improving reusable AI agent skills
SummaryThis paper introduces SkillNEO, a method for improving the reusable "skills" that LLM agents use to carry out procedures. It keeps lower-scoring skill versions in development so that partial repairs can pay off later, and compares how related versions perform to decide what to keep and what to fix. On 86 SkillsBench tasks, GPT-5.4 Mini's average reward rose from 49.1% to 66.4%. The evolved skills also improved stronger models without further editing.
In briefPlans That Learn as They Fly: Real-Time Robust Adaptive Robot Control on GPUs
SummaryThis paper presents A-SLS, a robust adaptive model predictive control framework that accounts for future learning when planning, so robots can stay safe while reducing uncertainty. It forecasts future parameter uncertainty to build tighter reachable sets, and includes a GPU-accelerated solver for real-time use. The authors report less conservative reachable tubes on nonlinear robotic systems, scaling to whole-body control of a quadruped with 61 states, 8 uncertain parameters, and 12 inputs. A physical Crazyflie drone is also shown learning an uncertain parameter online to stabilize.
WhyIncorporates guaranteed future parameter learning into robust adaptive MPC reachable sets, with GPU solver validated on quadruped and hardware drone.
In briefFormalizing how parts combine into compositional structure, and generating such structure
SummaryThe paper formalizes compositional structure using algorithmic information theory, defining it as shared structure that appears when parts are considered together and is not captured by modeling them separately. Using tractable approximations, the authors invert the framework to optimize strings and images with a specified compositional form without prescribing their content. The generated objects show recognizable hierarchical reuse and part-whole organization, and the authors discuss implications for machine learning, including depth in neural architectures and data curricula.
Eric Elmoznino, Tejas Kasetty, Blaise Ag\"uera y Arcas et al.
In briefMulti-Tiered AI Accessibility Model (MAIAM): Tiered Access for Frontier AI Safety
SummaryThe paper proposes a Multi-Tiered AI Accessibility Model (MAIAM) that sorts who may use AI capabilities by purpose, safety, and capability rather than treating release as a binary choice between API access and open weights. It defines three tiers: broadly accessible, capability-bounded AI for learning; registered and controlled AI for research and development; and full-scale frontier AI available only through recorded, accountable access. The authors pair MAIAM with the SAFE AI Operating Point to form a Capability Boundary and an Accessibility Boundary. They argue that capability limits alone are insufficient and that tiered access administered by regulators is also needed.
In briefAI weather models forecast storm power outages to time crew deployment
SummaryThe study uses two AI weather models, GraphCast and FourCastNet, to predict county-level power outages across the contiguous US 24 to 120 hours ahead. It measures how forecast lead time affects the outage burden captured where crews would be staged. Captured burden declines as lead time grows, and early deployment with all requested crews is compared against later deployment with fewer crews.
Mohammad Kian Golkar, Luciano Alves de Oliveira, Mohamad Alipour
In briefDecision-FLIP: Budget-free pruning for Gaussian occupancy models
SummaryOccupancy prediction models rely on large banks of Gaussians or learned queries whose size is set by hand. The paper introduces Decision-FLIP, which scores each row by how many voxel decisions change when it is removed, using only unlabeled calibration inputs. A Otsu-based split selects the number of survivors automatically, and the kept rows are compiled into a smaller checkpoint. Across several settings, it removes 12-79% of eligible rows and reduces bank-dependent latency, with semantic mIoU changing by at most 0.18 point.
In briefRecursive Offline Distillation via Entropy Gain Sampling and Answer Mixing: a method for training small models on large models' reasoning
SummaryThis paper studies distilling reasoning chains from large language models into smaller models, for cases where full access to the teacher's internal representations is unavailable. The method samples reasoning chains weighted by entropy difference against a medium-sized auxiliary model, then adds single-token ground-truth answers to the thinking traces. The authors report an average of 57.15 ± 1.01 versus 51.61 for naive distillation at comparable data budgets, with a steeper learning curve.
Andrey Goncharov, Daniil Vyazhev, Simon Epanov et al.
In briefReplicable Dueling Bandits: Choosing the Same Pair of Options Across Runs Despite Noisy Comparisons
SummaryThis paper studies replicable dueling bandits, where two runs sharing internal randomness but with independent noisy comparison outcomes must choose the same ordered pair in every round with probability at least 1-ρ. Building on Beat-the-Mean, the authors design the first such algorithms with sublinear regret under relaxed and strong stochastic transitivity, and for linear dueling bandits. Experiments on synthetic and language-model preference data show replicability and sublinear regret.
In briefMultimodal Screening Using Video, Audio and Text From Alexa Conversations to Detect Cognitive Impairment
SummaryThis paper proposes a contactless screening framework that uses video, audio, and text collected during structured voice conversations with Amazon Alexa to detect mild cognitive impairment. Three conversational tasks were compared across unimodal and multimodal models. The multimodal model performed more consistently across tasks, reaching an accuracy of 0.76.
Sankalp Vaish, Rishank Singh, Sumukh Veeramalla et al.
In briefEgoJEPA: A Retargeting-Free Method for Learning Dexterous Manipulation from First-Person Videos
SummaryThis paper presents EgoJEPA, a three-stage framework that learns embodiment-agnostic representations from egocentric human videos without hand-crafted human-to-robot retargeting. It was tested on five real-world manipulation tasks across three robot embodiments. Under out-of-domain evaluation it reached a 74% average success rate, versus 65% for the retargeting baseline, while cutting cumulative cross-embodiment wall-clock cost by 47.2%, from 743.2 to 392.7 hours.
In briefSign-betting regression: a robust estimator that adapts to noise and approximates optimally under misspecification
SummaryThis paper studies bounded regression when the model is misspecified, measuring error by the L1 distance to the conditional mean. It analyzes the min-max estimator of Baraud and Maillard (2025), called sign-betting regression. The authors show high-probability L1 oracle inequalities under misspecification, with bounds that adapt to average conditional variance under realizability. They also show the optimal universal approximation factor for proper learners is 3, and support the theory with synthetic and real-world experiments.
In briefTGrasp6D: Zero-shot task-oriented 6DoF grasping using vision-language priors and pose refinement
SummaryThis paper presents TGrasp6D, a method that lets robots generate task-oriented 6DoF grasps without task-specific training data. Given a task and one RGB-D image of an object, it uses vision-language reasoning to produce a task-informed grasp prior, then refines the pose to avoid collisions while keeping the grasp consistent with the task. It reports a 94% success rate on the TaskGrasp dataset, compared with 74% for the best baseline. The authors also deploy it on a Franka Research 3 robot to show zero-shot generalization to real-world objects.
Manav Kulshrestha, Apoorva Vashisth, Damon Conover et al.
In briefWrong Votes, Right Gradients: Majority-Vote Rewards Versus Verifiable RL
SummaryThe paper examines why reinforcement learning using majority-vote labels from a model's own samples can recover much of the gain of training with verifiable rewards. The authors argue that label errors matter only when they point in the same direction across prompts, and that incoherent errors cancel out as batch size grows. Experiments on five models show the majority-vote update closely matches the gold update. However, the equivalence is local, and gold-reward training still achieves higher held-out majority accuracy.
WhyOffers a new batch-level explanation for majority-vote rewards, identifying error coherence as key, validated by gradient measurements across five models.
In briefStochastic gradient descent analyzed as a nonlinear time series
SummaryThis paper studies constant-step-size stochastic gradient descent (SGD) using iterated random functions. It proves Geometric Moment Contraction, which yields convergence in Euclidean distance rather than only in probability measure. The authors establish quenched central limit theorems, invariance principles and α-stable convergence for averaged SGD regardless of initialization. They also propose an enhanced Richardson-Romberg extrapolation, an online inference method, and empirical confidence intervals whose coverage approaches the nominal level as n grows.
In briefFlexible doubly robust proximal causal inference for censored time-to-event outcomes
SummaryThis paper studies estimating causal effects on right-censored survival outcomes when unmeasured confounding is present. The authors propose Proximal Survival Bridge learning (PSB), which uses proxy variables and yields outcome-bridge, treatment-bridge, and doubly robust estimators with kernel (PSB-K) and neural (PSB-N) bridge learners. The methods are compared on a SEM benchmark, a semi-synthetic ACTG study, and the Framingham study, with additional tests under weakened or invalid proxies and heavier censoring.
Dhanajit Brahma, Ruolin Meng, Lawrence Carin et al.
In briefLearning simplex warm starts from measured solver cost
SummaryThis paper studies using machine learning to predict starting bases for the simplex method so that fewer iterations and less solve time are needed. The authors find that a good starting basis removes many iterations but saves only a fraction of the time, because the removed iterations are cheap and closeness to an optimal basis does not rank starts by cost. They propose CostStart, which uses a small policy trained on measured instruction counts to choose per LP the number of predicted columns, the algorithm and pricing rule, and whether to warm-start. On unseen LPs of the training classes, it roughly halves solve time relative to the solver's default start.
Felix Parker, Danielle A. Ripsman, Emmett Springer et al.
In briefDo Emergent Misalignment Models Share a Circuit?
SummaryThe authors fine-tuned three Qwen2.5-3B-Instruct models with LoRA on harmful medical, financial, and extreme-sports advice to study whether emergent misalignment relies on shared components. They identified three attention-head locations common to the top five heads of all three models. Replacing those heads' contributions with base-model contributions lowered free-generation misalignment substantially without reducing judged coherence.
In briefFrom Pixels to Tokens: Token-Space Active Learning for Very Low-Label Semantic Segmentation
SummaryThe paper introduces TEAL (Token-space Efficient Active Learning), the first pixel-level active learning framework built on vision transformer (ViT) representations. It performs diversity-based selection on the frozen backbone's native token lattice, then refines choices with margin-based uncertainty on a finer decoder grid. On CamVid, Cityscapes, ADE20K, and Pascal Context, TEAL outperforms prior baselines under extreme label scarcity, improving by up to +24.36 mIoU with one labeled pixel per image and reaching up to 97% of fully supervised performance with 10 pixels.
In briefDefoliationCAM: Edge AI for Real-Time Grapevine Insect Defoliation Assessment
SummaryThis paper presents DefoliationCAM (D-CAM), a solar-powered edge AI system that estimates insect-caused leaf loss in grapevine canopies from 4K images and sends results via LoRa to a web dashboard. It uses YOLOX-Nano to locate cordons and a lightweight MobileNet-based classifier to grade image patches, aggregating them into a canopy-level NGDS score validated against an entomologist's ratings. The authors report a five-week deployment of two field units demonstrating autonomous, solar-powered operation and remote monitoring.
In briefCogDrift-4D: Hierarchical Multimodal Modeling of Behavioral Decay for MCI Detection
SummaryThe paper proposes CogDrift-4D, a hierarchical 4D Transformer that detects mild cognitive impairment from behavioral data. Data were collected over 18 months across seven unscripted multimodal interactions and modeled as a 4D tensor to capture within-session fading and cross-session drift. On structured verbal and linguistic interactions, the method reports 0.733 ± 0.13 accuracy and 0.716 ± 0.14 F1-score, which the authors say outperform advanced baselines.
Sankalp Vaish, Sumukh Veeramalla, Rishank Singh et al.
In briefMachine-learning model for neutrino events, calibrated to experimental measurements
SummaryThis paper presents a generative surrogate model for charged-current ν_μ-nucleus scattering that can be calibrated to measured differential cross sections, without needing event-level experimental data. The authors fine-tune the model to a MicroBooNE measurement and report that it also improves the description of observables not used in the fit. The calibrated model can be applied to measurements with different neutrino fluxes without retuning, while still predicting outgoing-muon kinematics event by event.
Jose Luis Bonilla, Krzysztof M. Graczyk, Rwik Dharmapal Banerjee et al.
In briefA machine-learning study that estimates biological age from health survey data and explains which factors drive the estimate.
SummaryThis study uses data from 24,983 adults aged 18 and older in the U.S. NHANES survey (2011 to March 2020). A CatBoost model estimates biological age from 44 demographic, laboratory, physical examination, and lifestyle predictors, with the model's estimated age defined as biological age. The authors apply SHAP explanations and add a prespecified automated procedure that quantifies feature-value ranges, effect sizes, and inflection points on SHAP dependence plots. In held-out testing, estimated biological age correlated strongly with chronological age (Pearson r = 0.798).
In briefPredictive Commonsense Reasoning About Moving Objects: An ASP-Based Explainable Framework for Interacting Trajectories
SummaryThis paper proposes a general declarative framework for anticipatory reasoning about interactive motion in complex dynamic environments, aiming to form explainable space-time mental models at a semantic level. The method combines an ASP model of interactive trajectory dynamics, semantically guided prediction that produces branching trajectory futures, and an energy-based optimisation model that ranks and filters anticipations. Evaluated on the Argoverse 2 benchmark, each predicted trajectory carries an explicit, inspectable event-level hypothesis, unlike end-to-end stochastic prediction.
In briefSHRED: Demo-free self-play training for mixed road users
SummaryThis paper presents SHRED, a framework that trains vehicles, pedestrians, and cyclists jointly through large-scale heterogeneous self-play from random initialization, without expert demonstrations or logged trajectories. It combines type-specific dynamics and objectives with procedural scenario generation and per-agent randomization. On the Waymo Open Sim Agents Challenge it outperforms a demonstration-free self-play baseline and self-play methods using demonstration-trained reference policies, and on the InterPlan benchmark it outperforms all compared planners.
WhyDemonstration-free heterogeneous self-play beats baselines on WOSAC and InterPlan with ablations, extending existing self-play paradigms rather than creating a new one.
Wei-Jer Chang, Zhouchonghao Wu, Akshay Rangesh et al.
In briefRunning frozen short-context drafters on very long inputs for speculative decoding
SummaryThe paper addresses how speculative decoding loses its speedup at ultra-long context, since drafters that attend over the full prefix become costly. The authors propose keeping released short-context drafters frozen and running them past their training length using two inference-time mechanisms: bounding the attention cache and re-basing rotary position indices. They report that the speedup stays flat out to two million tokens across three target-drafter pairs and five corpora. The work also includes proofs and bounds on acceptance.
In briefSNOAR: Self-training for speech denoising without clean reference audio
SummaryThis paper proposes SNOAR, a self-training framework for speech denoising that does not require clean reference recordings. It removes residual noise from a teacher model's speech estimates using differential noise filtering (DNF), then adds back part of the noisy input to restore over-suppressed speech and reduce artifacts. Compared with RemixIT, SNOAR trains more stably and performs better on signal-based, perceptual, and word error rate metrics. It reaches 97% of a fully supervised topline's SI-SDR improvement on WHAM! and improves all metrics for three teachers on real CHiME-3 recordings.
In briefTime-partitioning Gaussian Process Mixture Model for Regime Detection in Time Series
SummaryThe paper proposes the Time-partitioning Gaussian Process Mixture Model (TPGPMM), a Bayesian nonparametric method that identifies regimes in nonstationary time series and predicts within them without Markovian assumptions. It combines an adaptive Pitman-Yor prior that partitions time into contiguous blocks with regime-specific Gaussian processes. On simulated data, the CBOE Volatility Index, and the 10-year US Treasury yield, the identified regimes align with documented market and monetary-policy events.
In briefDoes Morphological Correctness Matter for Tokenization? A Tagalog Scrambled Control Study
SummaryThis paper compares MorphAustro-Tok, a tokenizer that splits words at morpheme boundaries, with a corpus-matched BPE baseline. Tagalog marks voice with infixes that break the root into non-contiguous pieces, which BPE merges cannot recover. A scrambled-morpheme control randomizes the pseudo-infix while keeping morph count and invertibility. MorphAustro-Tok improves Tagalog machine translation over BPE, but scrambling the analysis changes results by less than seed variance, so no contribution from morphological correctness is detected at this scale.
In briefProbing how vision-language models understand negation via object attention
SummaryThis paper examines how vision-language models process negation by analysing their attention patterns and comparing them with human eye-fixations on affirmative and negated statements. Evaluations use both synthetic and realistic datasets. The authors find that models generally do not follow human-like attention when resolving negation, and that all models show a performance gap between affirmative and negated questions.
In briefComplete or Sparse: Two Identifiability Cases for Interpretable Neural Representations
SummaryThis paper studies when neural networks learn interpretable representations by asking whether supervised learning can recover the true latent variables behind the data. It introduces two data generating processes differing in causal direction and proves that with enough representation dimensions, optimal encoders are linear in the latents. Adding constraints such as non-negativity and energy efficiency, or sparsity with the spark condition, lets individual latent variables be recovered. The authors validate these predictions on synthetic benchmarks and on activations from vision models, large language models, and macaque inferotemporal cortex.
In briefRank-constrained score networks in diffusion models and how they affect generalization
SummaryThis paper studies how architectural rank constraints in diffusion score networks affect generalization. Experiments show that higher rank lowers training error, but its effect on test error depends on sample size and data dimension. Using a random-feature model with Gaussian data, the authors derive asymptotic training and test errors and identify the dimension of the orthogonal complement of the data subspace as a natural rank threshold.
In briefCo-evolving skills and policies can make self-improving agents fragile when skills drift
SummaryThis paper studies self-improving agents whose policy model and skill bank improve each other during training. The authors argue that current evaluations test the policy only on the final skill bank used in training, missing the drift that occurs when skills are rewritten or changed after deployment. Theoretical and empirical analysis links this fragility to the shared cooperative acceptance rule, with sensitivity to drift growing as co-evolution proceeds. They introduce CODA, a deployment-oriented audit protocol, and report that worst-case success falls 5-13 points below each delivered pair's peak across four methods and five benchmarks.
In briefMARS: A foundation model that auto-regressively generates map road networks and buildings
SummaryThis paper introduces MARS (Map Auto-Regressor), described as the first map foundation model, which generates multi-polyline road networks and polygon buildings in one unified, auto-regressive process. The authors trained it on MAP-3M, a multi-class map extraction dataset of 3.4M examples. Across four road and building datasets, MARS matches or outperforms multistage baselines, and it supports interactive human-in-the-loop map generation and correction.
In briefAnalog radio frequency computing that combines wireless data and AI services on one network
SummaryThe paper proposes sending neural network weights as a multicast waveform from a base station, letting edge devices perform AI matrix-vector multiplications with passive mixers at very low power and memory. It studies a multi-user MIMO downlink where this AI flow shares resources with unicast data flows, and formulates a joint beamforming problem solved with a WMMSE-based algorithm. Numerical results using 3GPP channels show the relative aggregate data rate reduction is below 5% in representative settings.
In briefNeural geometry of task concepts inside vision-language-action robot policies
SummaryThe paper studies the geometric shape of task-relevant concepts inside vision-language-action (VLA) robot policies, going beyond whether they can be linearly read out. Using controlled sweeps, natural rollouts, linear probes, and activation interventions, the authors build a taxonomy of seven concepts. They find the realized geometry is mostly multi-dimensional and shaped by physical constraints, such as joint limits producing open arcs for wrist orientation. The same geometries appear in two other VLA models trained on real-world robot data.
Mozes Jacobs, Aiden Swann, Mathilde Papillon et al.
In briefConfidence-calibrated controller deciding when to spend more compute during evaluator-guided reasoning
SummaryThe paper treats evaluator-guided reasoning as a per-step choice among trusting the current answer, returning a fallback, or spending more inference. It proposes a compute-aware three-action controller that uses calibrated confidence estimates for each action. On held-out width expansion and revision tasks, the controller improves observed utility over calibration-selected fixed policies while reducing normalized online compute by about 30% and 16%, respectively.
In briefTraining LLMs to Gather Information Efficiently Through Reasoning
SummaryThis paper studies how large language models can learn to gather information adaptively in experimental design tasks. The authors train the model's reasoning with information-gain rewards, amortizing sequential Bayesian experimental design into the LLM. On Twenty Questions, the resulting policies improve information acquisition and task performance at lower computational cost than inference- and training-time baselines. The skill also generalizes to unseen categories and transfers positively to the Medi-Q clinical dialogue benchmark.
Anita Kriz, Antonin Berthon, Nicol\'as Astorga et al.
In briefPEDBench: Can Speech Models Detect Phonetic Errors?
SummaryThis paper introduces PEDBench, the first benchmark for phonetic error detection, covering natural to controlled speech. It includes real child recordings of spoken errors (334), scripted errors by human actors (1,188), and synthetic errors from phone-conditioned text-to-speech (1,083). Evaluating 13 state-of-the-art speech models, the authors find they fail to reliably detect phonetic errors, even under controlled conditions, and often misclassify correct phonemes. Speech-language models fine-tuned on synthetic data outperform baselines at word-level detection, though a sim-to-real domain gap remains.
Zhonghao Shi, Zackary Rackauckas, Jasmine Yip et al.
In briefCoTranslator: A Tool That Converts Answer-Conditioned Reasoning into Verified Forward Reasoning for Training Smaller Models
SummaryThis paper proposes CoTranslator, a method for producing chain-of-thought training data more efficiently for reasoning distillation. It learns to translate reverse chain-of-thought responses, which are conditioned on the known answer, toward the distribution of verified forward responses. The authors report up to 5.82 percentage points higher average accuracy than semantic suppression across six mathematics benchmarks, and transfer to a new dataset and STEM problems without retraining.
In briefRouting Statistics Are Not a Reliable Guide to MoE Layer Quantization Sensitivity
SummaryThis paper tests whether per-layer, per-task-phase expert-routing statistics in a Mixture-of-Experts language model can guide mixed-precision quantization of expert FFN tensors. The authors build a profiling pipeline for Qwen-3.6-35B-A3B and tune layer bit allocations under a fixed average bits-per-weight budget. Routing-guided reallocation did not outperform the uniform baseline or a random control, and under the aggressive budget it showed a significant loss. The authors conclude that routing-behavior statistics are an unvalidated proxy for true per-layer quantization cost.
In briefLinear independence of polynomial compositions with bounded input degree
SummaryThe paper proves the weak composition-independence conjectures posed by Kohn, Marchetti, Massarenti and Mella. The proof combines deformations of rational curves with a single kernel of a linearized equation, and uses Riemann-Roch to derive a contradiction from tangent fields with controlled poles. As a consequence, neural networks with a general polynomial activation function of sufficiently large degree are shown to be identifiable.
WhyProves a conjectured weak composition-independence result, resolving an open problem via deformation and Riemann-Roch arguments; purely theoretical.
In briefOpen-weight language models adapted for Wolof and ready for industry use
SummaryWolof, spoken by about 23 million people mainly in Senegal, has weak support in open-weight language models. This paper presents open-weight models from 0.6B to 14B parameters, adapted to Wolof through instruction tuning, preference alignment, and model merging. Experiments on 63 public and in-house tasks show good Wolof understanding and generation, with effectiveness retained on other languages, and the 8B and 14B models are competitive with much larger commercial systems. The models are being evaluated in internal customer-care pilots and are publicly available on Hugging Face.
Gw\'enol\'e Lecorv\'e, Delphine Charlet, Pierre Adam et al.
In briefPre-learning feasibility limits for drone-mesh interception sizing and timing
SummaryThe paper studies a mesh of drones spread over a predicted arrival point to intercept an approaching object, deriving closed-form bounds on coverage before any policy is chosen, validated by Monte Carlo simulation. Under a minimum 2r separation constraint, it reports coverage ceilings of 90.69% in the plane and 74.05% in a volume. The authors argue such ceilings should be computed and reported before learning experiments.
In briefTemporal Specification Guidance for Text-to-Video Generation
SummaryText-to-video models often generate requested visual concepts but miss required ordering, timing, and persistence. The paper expresses these requirements as formal Temporal Logic specifications and proposes TSG, a training-free inference-time method that guides sampling with them. On a 220-prompt benchmark covering 11 temporal constraint categories, it improves specification satisfaction over classifier-free guidance by 9.5 to 20 percentage points across five models.
Avishree Khare, Hideki Okamoto, Bardh Hoxha et al.
In briefIntern Lumina U2: a diffusion language model using multiple visual codebooks for understanding and generating images, video, and 3D.
SummaryThis paper presents Intern Lumina U2, a diffusion language model that represents visual content as fully discrete tokens drawn from multiple complementary codebooks, with each visual position encoded by a group of codes. It builds on the AToken unified 4D representation and handles image, video, and 3D inputs in a single backbone. The authors report leading performance among unified models in visual understanding, image generation, and image editing, and validate the system on Ascend and NVIDIA platforms. A stability-aware sampling strategy is said to speed up image generation by 4.45x.
In briefO-RAN-native jamming detection across heterogeneous networks
SummaryThis paper addresses radio-frequency jamming that can disrupt cellular networks, detecting it from per-UE KPI telemetry exposed via the O-RAN E2 interface, without proprietary hardware. The authors propose SEM, a statistical ensemble combining adaptive MAD-based baseline detection, count-based anomaly detection, and abrupt change detection through weighted voting, with a cold-start mode needing no labels and a supervised mode tuned offline by grid search. Evaluated on about 25 hours of data from eight heterogeneous 5G environments against ML and DL baselines, the learning-based models tended to lose recall under distribution shift, while SEM kept detecting jamming at the cost of more false alarms.
Sneihil Gopal, Fernando J. Cintr\'{o}n, and Richard A. Rouil
In briefGibbs-Gap Policy Optimization: off-policy RL for verifiable rewards
SummaryThis paper addresses reinforcement learning with verifiable rewards when training responses come from a different or outdated behavior policy. It proposes Gibbs-Gap Policy Optimization (GGPO), which minimizes a log-mean-exp contrast of KL-adjusted rewards across responses to the same prompt, without learner-to-behavior importance weights. Experiments on mathematical reasoning show GGPO outperforming GRPO and AGRO under training-inference mismatch and on fixed offline data.
In briefRIGen: Reciprocal Interaction for Generative Gameplay Beyond Simple Action Response
SummaryThis paper introduces RIGen, a framework for reciprocal interaction in generative gameplay, combining persistent game state, configurable rules, contextual interaction reasoning, and action-conditioned audiovisual generation. The authors train an audiovisual generator on a Xonotic gameplay dataset and post-train it into RIGenArena, an interactive generative first-person shooter prototype. Experiments show more reliable combat outcomes and reciprocal interaction between the player and the generated world.
In briefA self-improving agent that expands its own domain language for exact question answering without external judges
SummaryThis paper studies whether an agent can expand what it can compute exactly without rewards, judge models, or answer keys. A single frozen model runs two nested loops over a small domain-specific language: the inner loop accepts answers only when two independently written expressions agree on every record, while the outer loop identifies missing capabilities and writes new operators that are admitted only if an independent implementation agrees and a conservation law holds. The authors report that the mechanism also expands answerable question sets in two domains with different schemas, and that its admission rule rejects all 138 buggy operator variants tested.
WhyNovel oracle-free operator induction with independent-implementation and conservation-law gating, validated across three domains and several model families; impact is confined to exact question answer
In briefPRISM: How activation interventions transfer across languages and model sizes to curb unsafe tool calls
SummaryThe paper observes that language models can make unsafe tool calls even when they textually refuse a request. The authors introduce PRISM, a framework that evaluates how activation interventions transfer across precision, languages, and model scales. Using labels from a quantized NF4 model and cross-lingual transfer offers a practical route to improve alignment. Five-language NF4-supervised directions reduced unsafe violations in Llama-3.1-8B on held-out Arabic by 21.95 percentage points, and the interventions were also transferred to models up to 70B parameters.
Giovana Kerche Bon\'as, Thales Sales Almeida, Marcelo da Silva Reis et al.
In briefWhen a reasoning trace conflicts with new evidence, the answer setup determines which one wins
SummaryThe paper studies whether a model's retained reasoning trace helps accuracy or competes with current evidence when the two disagree. It tests three model families on HotpotQA and MuSiQue, comparing continuing the model's own thinking with a separate answer call. The separate call follows current evidence far more often, while continuation often repeats stale conclusions.
In briefOptimal explicit and implicit regularization in high-dimensional continual linear regression
SummaryThis paper studies continual linear regression in high dimensions, where a learner sees tasks in sequence and updates its parameters each time. The authors derive a closed-form expression for the generalization loss of gradient descent run for a fixed number of steps per task. For single or i.i.d. teachers, the optimal step size decays roughly as ln T / T, and the resulting loss matches that of optimally tuned isotropic L2 regularization. The analysis also covers generalized ridge regression and identifies a regime where the optimal step size stays finite as tasks increase.
In briefPredicting music enjoyment from brain signals with explainable deep learning
SummaryThis paper treats music enjoyment prediction as a regression task using EEG recordings from 20 listeners in the public MUSIN-G dataset. Six deep learning models were compared with listeners split into training and held-out test sets, and the temporal convolutional network reached the highest correlation with ratings (r = 0.21). Gradient-based and occlusion explanations identify which time segments within each two-second window most influence those predictions, offering a reproducible cross-listener benchmark.
Parshav Singla, Simranjit Kaur, Maanya Walia et al.
In briefBlame Wrong Words, Not the Canvas: Region-Level Feedback for Accurate Text in Images
SummaryThis paper addresses errors in AI-generated images, where text may be misspelled or its letters malformed. It introduces TextDART, a reward model that locates each text region and separately judges its meaning and letter quality. It also proposes SCANNER, which uses these regional judgments to give region-specific feedback when training the image generator. The authors report state-of-the-art results in text accuracy and glyph fidelity on existing benchmarks and on their new PolyText-Bench.
In briefCompact distribution sketches with computable error bounds
SummaryThe paper studies compact, weighted-atom approximations of one-dimensional probability distributions that are optimal under quadratic Wasserstein distance, extended to higher dimensions through slicing. It derives an exact decomposition of squared distance error that yields computable error certificates for any centroid sketch. Experiments on document, point-cloud, and flow-cytometry data show optimal sketches greatly reduce storage while keeping classification performance close to uncompressed sliced distances.
In briefEvolvingBrain: An adaptive skill system that helps multimodal agents improve themselves in open worlds
SummaryThis paper proposes EvolvingBrain, an adaptive skill system that serves as the "brain" for self-evolving multimodal agents. It has two parts: Stratified Skill Self-Evolution, which extracts procedural and problematic patterns from trajectories to build Domain, Plan, and Action/Trap skill layers, and State-Aware Adaptive Execution, which summarizes long interactions into structured memory to guide skill use by current state. The authors report that experiments on VAB-Minecraft and DiscoveryWorld show improved success rates on open-world tasks and some generalization ability.
In briefAPRiL: A benchmark testing whether language models understand and respond to social situations in Chinese
SummaryThis paper introduces APRiL, an open-ended benchmark for Chinese social awareness in large language models, with 2,196 evaluation instances across six categories. It includes two tasks: Social Scenario Analysis, which tests understanding of social cues, mental states, and conflicts, and Socially Aware Response, which tests whether models can turn that understanding into appropriate replies. Experiments on 12 models from six families show a consistent gap between understanding and response, with better scenario analyses improving responses and weak analyses possibly lowering their quality.
In briefOpen-world AI: Three Requirements for Versatile Machine Learning
SummaryThis paper contrasts closed-world AI, which assumes fixed, i.i.d. data, with open-world AI, which must handle unfamiliar tasks from few examples. The authors derive three requirements: inference-time learning, a rich representation, and predictive disentanglement. They test language models up to 8B parameters and find that models meeting these requirements improve as data from related tasks accumulates, while others degrade, and neither scale nor layer-type interleaving fixes this.
In briefReasonUni: Reasoning-Augmented Reconstruction Improves Unified Multimodal Models
SummaryThis paper proposes ReasonUni, a framework that couples image understanding and generation in unified multimodal models through reasoning. Given an image and an editing instruction, the model first reasons about what and how to edit, then generates the target image from that understanding. The authors also introduce TreeGRPO, a tree-structured reinforcement learning method that jointly optimizes both modules. Experiments across multiple benchmarks report gains in both understanding and generation.
In briefARIA: A Framework for Inspecting and Fixing Multi-Agent AI Systems While They Run
SummaryThe paper addresses how a single mistake by one AI agent can spread through a multi-agent system and cause the whole task to fail. The authors introduce ARIA, a framework that learns when and how to repair ongoing runs, built on a dataset of 3,953 execution-verified repair pairs drawn from over 24,000 trajectories. A compact trained model is taught to recognize when intervention is needed and how to continue the run. The authors report strong repair performance in both post-hoc and runtime settings, with generalization to unseen frameworks and host models.
In briefVerdict-MQ: an open encoder that reads a case once and answers many typed questions
SummaryThe paper presents Verdict-MQ, an open encoder that reads a case once and answers all of its typed questions at once, with structured attention so questions do not influence each other. On the typed-decisions benchmark, a 395M-parameter model matches soft gold labels better than the released 421M Laya specialist (KL 0.080 vs 0.117, Brier 0.045 vs 0.061). Results on RACE, BoolQ and MultiRC are mixed: it wins on MultiRC but loses on RACE and BoolQ.
In briefSelection-safe conformal early exit with coverage under adaptive depth
SummaryThe paper addresses multi-exit early stopping in deep networks, showing that calibrating a conformal prediction set at each exit fails to preserve coverage once a data-dependent exit is chosen. The authors build one terminal path score from prefix maxima, calibrate a single split-conformal quantile, and return nested prefix sets that contain the valid terminal set. Across eight encoder task-backbone cells, pathwise selection reaches macro coverage 0.901, compared with 0.814 for naive layerwise selection.
In briefEEG-text semantic alignment: a contrastive learning method for decoding imagined speech from brain signals
SummaryThis paper studies decoding imagined speech from EEG by aligning EEG trials with embeddings from a frozen pretrained text encoder using contrastive learning. On the Chisco dataset across five subjects, it reports fixed-intent decoding performance comparable to a task-optimized closed-set classifier over 39 categories. The authors also report that maximum EEG-text cosine similarity indicates subsets of trials with higher accuracy, supporting similarity-based retrieval.
Shahryar G. Namdari, Francesco Iacomi, Riccardo Barbieri et al.
In briefA new way to measure how fast reality can change before AI fails
SummaryThe paper asks how quickly reality can change before an intelligent system can no longer reliably perceive, reason, and act. It defines an effective Speed of Coherent Cognition (S_C^eff) as the highest tested speed at which a system keeps meeting a predefined reliability criterion. In reinforcement-learning tests, the Co^4 architecture reached substantially higher S_C^eff than Transformer and gated-MLP baselines across the scenarios tested.
In briefReference-guided transfer of turn-taking style for full-duplex spoken language models
SummaryFull-duplex spoken language models decide at each frame whether to take the floor, which shapes their conversational personality. The paper identifies a single late-layer residual-stream direction, the initiative vector, extracted from the model's own conversations without training or annotation. This direction anticipates speech onsets up to 800 ms ahead and, when scaled, shifts speech ratio substantially. A short reference recording transfers a speaker's turn-taking style, and the transfer is reference-specific.
WhyIdentifies a single residual-stream direction controlling turn-taking in full-duplex speech models and transfers speaker style from reference audio; novel, with consistent multi-corpus experiments, bu
In briefPurify Before You Align: Better Representation Alignment for Diffusion Models
SummaryThe paper introduces local appearance purification (LAP), which removes locally predictable content from the alignment target so diffusion models align only to the residual information. LAP changes only the alignment target and requires no modification to diffusion training or inference. On ImageNet-256 with SiT-B/2, it reduces FID from 53.1 to 42.3 for MAE, with gains across five pretrained encoders.
In briefOptimal Hydra: Data-Informed Hydra-Nets for Heterogeneous Federated Fine-Tuning of LLMs
SummaryThis paper addresses fine-tuning large language models with federated learning when clients have limited computing resources. It proposes Optimal Hydra, which combines global layer selection with client-specific attention head selection aligned to local data, producing client sub-networks called Hydra-nets. The authors state this is the first model heterogeneous federated learning approach tailored to LLM fine-tuning. Experiments on three datasets reportedly show gains over baselines while reducing local resource demands.
In briefInterpreting Language Model Parameters: Adversarial Parameter Decomposition (VPD)
SummaryThis paper seeks to understand how neural network parameters implement computations, an area where prior work has made little progress. The authors propose adVersarial Parameter Decomposition (VPD), which splits a language model's parameters into subcomponents that each implement only a small part of the learned algorithm. The decomposition is trained so that only a small number of subcomponents are needed to explain behavior on any input, and the input-output behavior is preserved even under adversarially chosen ablations. The authors also study how sequences of interactions between subcomponents produce outputs for particular inputs, enabling a new kind of circuit analysis.
Lucius Bushnaq, Dan Braun, Oliver Clive-Griffin et al.
In briefErgodic topology sampling toward the continuum using stochastic normalizing flows
SummaryThe paper provides numerical evidence that a flow-based method can mitigate topological freezing in SU(3) Yang-Mills theory at lattice spacings as fine as 0.03 fm. It analyzes how the computational cost scales for a strategy combining open boundary conditions with Stochastic Normalizing Flows. Across the explored parameter range, the autocorrelation time of the squared topological charge and the statistical error on the topological susceptibility stay under control with current resources.
Claudio Bonanno, Andrea Bulgarelli, Elia Cellini et al.
In briefVertical Sliced Wasserstein Distances: A Scalable Way to Compare Distributions of Distributions
SummaryMany learning problems compare distributions whose individual observations are themselves probability measures, called meta-measures. Optimal transport distances on this space are often too expensive to compute. The paper introduces the vertical sliced Wasserstein (VSW) distance, which projects each inner measure with continuous test functions and compares the resulting one-dimensional distributions. The authors prove VSW is a valid metric that metrizes weak convergence and has dimension-independent statistical guarantees. Experiments show it is accurate and efficient, and a practical particle gradient enables gradient-based transport of meta-measures.
Elaheh Akbari, Ahmadreza Moradipari, Kyungtae Han et al.
In briefQuantile-Guided Diffusion for Probabilistic Time-Series Forecasting
SummaryThis paper studies using predicted quantiles as distributional controls for diffusion-based time-series forecasting. The authors propose Control-TS, which learns a quantile-control pathway while keeping a pretrained backbone frozen, and report consistent CRPS improvements across eight benchmarks. They also introduce the Conditional Linear Predictive Score (CLPS) to evaluate the downstream forecasting utility of generated futures.
In briefWorldCast: Exploring Visual Generation for Predicting Future World States
SummaryThe paper introduces WorldCast, a benchmark of 320 manually designed cases across eight categories that tests whether models can forecast successive future world states from multimodal observations and references. Twelve representative models were evaluated, with GPT-Image-2.5 System scoring 72.8 while the best open-source model scored 20.1. The results show substantial gaps in handling complex context, entangled references, and reasoning-heavy tasks.
In briefCan Learning to Answer Spatial Questions Improve Novel-View Generation?
SummaryThis paper asks whether 3D language understanding can provide complementary supervision for generating novel views of a scene. The authors introduce Ask4NV, built on a pretrained unified multimodal model with a Mixture-of-Transformers architecture, which incorporates camera geometry and uses understanding features during generation. Experiments report strong image quality and geometric consistency in novel-view generation, along with competitive 3D language understanding. Joint training is also reported to improve cross-view correspondence in VAE representations.
In briefOrbia: Do Generated Videos Form a Consistent 3D World?
SummaryThis paper introduces Orbia, the first benchmark for systematically evaluating 3D consistency in interactive video world models. It tests input preservation, persistence of generated content, and 3D self-consistency through controlled exploration and revisits. The benchmark includes 720 indoor and outdoor instances across six protocols and evaluates 27 text-, action-, and camera-conditioned models.
In briefRepresentation Conditioning (RECON) for generation: feeding vision-encoder semantics into later diffusion layers to speed up training
SummaryThis paper examines REPA, a representation-alignment method that trains diffusion transformers by supervising an intermediate layer with a pretrained vision encoder through a discarded MLP head. The authors argue that later layers, supervised only by the velocity prediction loss, may not access these semantics well. They propose RECON, which feeds pretrained vision-encoder tokens into later layers during training and reuses the MLP head as a proxy at inference. RECON adds no parameters, and the authors report it converges 63 times faster than SiT and 25 times faster than REPA.
In briefHierarchical residual policy learning for real-world mobile manipulation guided by sparse human input
SummaryThe paper presents HiRe-MoMa, a framework for stable, sample-efficient fine-tuning of mobile manipulation robots directly in the real world. It learns additive corrections on top of a frozen base policy, splitting the residual into mobile-base and manipulator-arm heads. Sparse expert interventions, covering under 20% of training, are used to shape the critic's value function. Experiments on the TIAGo++ platform report consistent performance gains after about 30 minutes of online interaction.
In briefGenerating and Interpreting Neural Weights of Humanoid Control Policies via Diffusion
SummaryThis paper examines whether high-performing humanoid control networks can be generatively modeled directly in weight space. It uses a latent diffusion framework in which a linear autoencoder compresses actor parameters and a Diffusion Transformer models the latent distribution. Trained on Unitree G1 reinforcement learning networks, the approach generates deployable networks that perform walking, dancing, and spin kick on the real robot. The authors also report that the first-layer weights reflect a kinematic prior that emerges under both reinforcement learning and imitation learning.
In briefVITAL: Turning Visual Motion into Precise Driving Actions (a method that derives driving action targets from visual motion)
SummaryThe paper notes that world-action models usually train action experts on trajectory representations defined independently of visual dynamics. It introduces VITAL, an action tokenizer that pairs a motion scaffold derived from visual motion with a compact residual encoding the remaining trajectory error, and uses flow matching to jointly predict both from the current image, ego state, and navigation command. On NAVSIM v1 navtest, a single denoising step reaches 90.8 PDMS, within 0.13 points of the ten-step score, at 3.29x the end-to-end inference speed.
In briefCommerceGen-Studio: An end-to-end system for generating commercial product main images
SummaryThis paper presents CommerceGen-Studio, a framework that generates commercial-grade main images for e-commerce products (SKUs) through three stages: information extraction, semantic rewriting, and multi-condition generation. The authors state it runs in production at millions of images per day, with a latency of a few seconds per image. They also release CommerceGen-Bench, a benchmark of hard cases from live traffic, where the framework achieves the highest overall scores among evaluated open-source baselines.
In briefDynThinker: A Unified Model for Static 3D and Dynamic 4D Spatial Reasoning
SummaryThe paper observes that existing models struggle to handle both static 3D and dynamic 4D spatial understanding, since 3D post-training can weaken dynamic reasoning and 4D-specialized models often lack strong static 3D reasoning. It proposes DynThinker, combining Decoupled Exponential Moving Average (D-EMA) and Spatial-Dynamic Fusion (SDF) to jointly strengthen both capabilities. Across ten metrics on nine static 3D and dynamic 4D benchmarks, DynThinker achieves the highest average score among all evaluated models.
In briefCo-Evolving Acting and Forecasting for Language Agents: a training method that uses intermediate steps to learn faster
SummaryThis paper addresses language agent reinforcement learning, which uses only a terminal reward and leaves intermediate-step supervision largely unused. It proposes CoPE, which trains the policy on successful trajectories to generate multi-step action sequences following each history, so acting and forecasting co-evolve. The policy still acts one step at a time at inference with no added overhead. The authors report gains over GRPO and RL baselines across five benchmarks, with up to 1.9x and 2.0x higher sample efficiency than ECHO and GRPO on ScienceWorld.
WhyProposes multi-step action targets co-evolving with acting, validated across five benchmarks against GRPO and auxiliary-target baselines with reported sample-efficiency gains.
In briefGeoThinker-F: Predicting 3D-aware world representations inside an MLLM for faster spatial reasoning
SummaryGeometry-aware multimodal language models usually call an external 3D encoder at inference, which adds latency and compute cost. GeoThinker-F instead turns the model's own hidden states into predictive latent representations of current geometry and future semantics. Training is two-stage: first, a frozen backbone with lightweight prediction heads learns from an offline 3D geometry expert; then the expert is removed and the model is jointly trained to integrate its self-predicted priors. The authors report up to a 3.6× speedup, with latency reduced from 1.45 s to 0.40 s while keeping competitive spatial reasoning performance.
In briefFeed-forward multi-view feature coding for efficient 3D scene representation
SummaryThis paper proposes MFGS, an end-to-end feed-forward framework that codes 3D scenes for Gaussian splatting. Instead of compressing an explicit Gaussian representation, it extracts multi-view features from input images, encodes them into a compact bitstream, and lets the receiver predict Gaussian attributes from the decoded features. Reported results on RE10K, ACID, and DL3DV show 99.76-99.86% BD-rate savings over a generate-then-compress pipeline, with 52-55% fewer GMACs and about a 12x end-to-end speedup.
In brief"Your Transformer Is Secretly Elastic": retraining-free method that makes pre-trained transformers resizable to fit different compute budgets
SummaryThe paper proposes a retraining-free method that converts a pre-trained transformer into a single model that can adapt its size on the fly to different computational budgets. It uses a differentiable relaxation to jointly rank every MLP neuron and attention head and soft-prunes across many sparsity targets at once. Across 27 vision and text encoders from 22M to 8B parameters, the authors report smooth degradation up to 60% sparsity and better results than one-shot baselines under deep pruning.
Walter Simoncini, Dheeraj Varghese, Gertjan J. Burghouts et al.
In briefEfficiency loss in AI agent prompt auctions
SummaryThe paper studies a decentralized market where AI agents bid for user prompts using first-price auctions, in which the highest bidder pays its bid and wins the prompt. Because bidders may shade bids below their true value, allocations can fail to maximize welfare, which the authors measure as the price of anarchy. They show that a competitive floor and near-independence of values given the prompt each limit this welfare loss, with the loss approaching zero in the first case and an efficiency bound approaching 1 - 1/e² of optimal welfare in the second. These results are validated through simulations with strategic bidders on the EmbedLLM prompt routing dataset.
Shubhendu Trivedi, Jacob Eisenstein, Alekh Agarwal et al.
In briefContinuum Attention: a new attention method that integrates over a continuous memory trajectory instead of discrete positions
SummaryThis paper introduces Continuum Attention (CA), which links neighboring key and value vectors into a continuous, piecewise linear memory trajectory, and each query retrieves information by normalized integration over its visible prefix. The authors state that the segment integrals have closed forms, making the operator analytically differentiable, and that it preserves causal masking and the standard Transformer decoder structure. In controlled byte-level experiments, exact continuous integration outperformed matched softmax attention, and in multilingual evaluation CA beat all baselines in every language, with an average 5.8% reduction in bits per byte relative to the strongest baseline.
In briefA learnable Hodge-Laplacian mixture of experts for steady PDE surrogate models on coarse meshes
SummaryThis paper proposes an architecture for steady partial differential equation surrogate models that splits the solution among several vanilla transformer experts, each handling a different spectral scale. The split is guided by a discrete Hodge Laplacian that measures high-frequency content, and a gating mechanism computed from the mesh recombines the expert outputs. On a reduced, remeshed version of the AirFRANS airfoil benchmark, the method is reported to cut relative L2 error by 31% and interior mean squared error by 47% compared with Transolver.
In briefPlayer-1: Learning Game Control from Synthetic Experience
SummaryThis paper presents Player-1, a post-training framework that learns game control from synthetic experience rather than relying only on demonstrations. It extends the computer-use action space with relative pointer motion, sustained inputs, and composite actions, and builds sixteen synthetic 2D and 3D environments for multi-environment reinforcement learning with PPO. Player-1-9B reaches 38.33% success on MineDojo-300, compared with 12.63% for the strongest evaluated baseline, with further gains on ViZDoom-10 and GameWorld.
In briefLatent Compaction: Learning a Smaller Per-Document KV Cache via Test-Time Training
SummaryThis paper addresses the memory and attention cost of long-context key-value caches during generation. It proposes Latent Compaction, which learns a compact, document-specific cache through test-time training while keeping model weights frozen and using only generic document prompts. On RULER at 32× compression, the method reaches 96.7% accuracy versus 40.0% for KVZip+. The authors report retrieval gains across Qwen3 and MiniCPM5 models.
WhyApplies test-time training to optimize compressed KV caches via distillation, markedly beating baselines at 32x compression across multiple benchmarks and models.
In briefOne Task, Many Interfaces: Making LLM Agents Robust to Interface Changes
SummaryThe paper shows that post-trained LLM agents often perform worse when a meaning-preserving interface change, such as renaming an action or changing notation, is applied. The authors propose Cross-Interface Policy Regularization (CIPR), a self-supervised objective that replaces action names with uninformative identifiers and penalizes KL divergence between action distributions under two interfaces. Evaluated in supervised and reinforcement-learning post-training, CIPR improves robustness on agentic function calling, recovering 10.5% of accuracy lost to held-out interface perturbations that otherwise reduced accuracy by nearly 25%, while leaving accuracy on the unchanged interface intact.
In briefMechanical Verification Beyond LLM Judges for Agentic Autoformalization
SummaryThis paper presents a method for verifying Lean formalizations that combines LLM-based semantic review with mechanical numerical and vacuity checks, providing checkable feedback and catching errors LLM judges miss. The authors apply it, together with their Extract-Verify-Proof pipeline, to build FLMF, a partial formalization of the NIST Digital Library of Mathematical Functions. On three labeled datasets, the method achieves the lowest accepted-error rate among the evaluated methods. Expert review also identified translation and source defects in existing datasets.
In briefCritique-guided on-policy distillation to restore reliable teacher supervision
SummaryThis paper studies a weakness of on-policy distillation (OPD), where a teacher model fails to guide a student out of its own incorrect rollouts even though the teacher can recognize the errors. The authors propose CriticOPD, in which the teacher critiques a wrong solution, the student rewrites it following the critique, and the teacher scores the revision. Tested on two Qwen3 student models and six mathematical reasoning benchmarks, CriticOPD outperforms OPD, with the largest gains on competition problems.
In briefAny Time, Any Space: A World-State Framework for Consistent Long Video Generation
SummaryThe paper presents ATAS, an agentic framework that turns a screenplay into shot requests with explicit location, camera, and event-time references. Before generation it retrieves evidence from scene and event memories, and after generation it reconstructs and registers the clip, updating memory only when checks accept it. Across four video generators, ATAS improves cross-shot consistency measures and outperforms compared agents and 3D-guided methods on spatial revisit and event recapture.
In briefManacá-1B-Instruct-2: A reproducible safety-aligned open model for Brazilian Portuguese, with negative DPO results and LLM-judge failures
SummaryThis paper presents Manacá-1B-Instruct-2, a 1.72B-parameter model trained from scratch for Brazilian Portuguese and aligned for safety. The authors report three failed DPO attempts and attribute the safety gains to a dedicated safety SFT, with an over-refusal knee point at 40-50% general data. Against four external PT-BR models, its safety lower confidence bound exceeds all baselines' upper bounds. The authors also found that an LLM judge silently refuses to score two safety items in MT-Bench-PT. Code, data, and logs are publicly released."}}
Bruno Leonardo Santos Menezes, Carlos Leonardo Souza Cardoso, Fabio Andr\'e Machado Porto
In briefScaling AI research needs forgetting: periodically resetting agent search state improves automated discovery
SummaryThis paper studies automated research using LLM agents and finds that running agents longer, in parallel, or with communication all eventually plateau as agents accumulate history. The authors propose periodically discarding the search state, including every agent's session and candidate history, and restarting from the best candidate found so far. Evaluated on four tasks spanning heuristic programming, kernel optimization and model training, the approach yields up to 15% improvement over parallel agents and up to 39% over a single agent.
