In briefThe Default Trap: rethinking how tool-using AI agents are evaluated against their plans
SummaryThe paper shows that an agent can respond strongly when a plan's priority is switched, yet barely change its choices when a default-aligned whole plan is removed, a pattern it calls the "default trap." Experiments on Retail, Airline, and AgentDojo tasks compare decision windows and full-task episodes against a no-plan reference. The authors argue that priority responsiveness, presentation-dependent defaults, and task success and cost should be reported together.
In briefCalibrating Decisions That Shape Future Outputs: On-Policy Quantization for Multimodal LLMs
SummaryThe paper notes that post-training quantization for multimodal large language models usually calibrates on fixed sequences with local objectives, overlooking how quantization-induced token changes redirect later generation. The authors propose OnPTQ, which calibrates on trajectories visited by the current quantized policy, uses short counterfactual rollouts to compare competing tokens, and prioritizes critical states via a Decision-Consequence risk. Across vision-language and omni-modal Qwen models at several low-bit settings, it improves downstream performance and reduces correctness flips relative to Dense/FP16 references without changing the deployed inference graph.
In briefSelf-consistency learning for diffusion-based video reasoning
SummaryThis paper studies whether self-consistency, which aggregates multiple reasoning samples, can improve visual reasoning with diffusion-based video generation. It samples several video rollouts, extracts paths, locations, or masks, and combines them into a consensus prediction, reading out results early to cut denoising steps by more than half. It also introduces Rejection Fine-Tuning to distill this consensus into the model so a single generation performs better. Tests on maze solving, visual search, and referring segmentation show gains, such as visual search accuracy rising from 48.4% to 99.0%.
In briefRED: A Method for Detecting AI-Generated Images by Tracking How They Are Rebuilt
SummaryThis paper proposes RED, a framework that detects AI-generated images by examining how they are reconstructed step by step, from coarse to fine, rather than only at the final output. It uses a frozen multiscale VQ-VAE and a frozen CLIP encoder to represent this reconstruction trajectory, and a frozen VAR model to weight each stage. The authors report an average accuracy of 92.5% and average precision of 97.5% across six benchmarks, with robustness to common image degradations.
In briefCorrGRPO: Correlation-Normalized GRPO for Learning from Multiple Rewards
SummaryThis paper addresses Group Relative Policy Optimization (GRPO), a method for training reasoning language models, which can let large-scale correlated rewards dominate normalization and suppress smaller-scale reward signals. The authors propose CorrGRPO, which normalizes pairwise covariances into Pearson correlation coefficients while keeping the centered total reward unchanged. The method is compared with GRPO and other variants on code generation, tool calling, and agent security using models from 0.5B to 8B parameters. Results reportedly show improvements across all three domains.
In briefACPO: A More Stable Training Signal for Reinforcement Learning of LLMs
SummaryThis paper proposes Advantage Clipped Policy Optimization (ACPO), a reinforcement learning method for post-training large language models. Instead of clipping only the importance sampling ratio, ACPO clips the product of the ratio and the advantage to obtain more stable gradient estimates. The authors connect this to gradient clipping in policy mirror descent and prove convergence of the clipped variant. On mathematical reasoning benchmarks, ACPO outperforms PPO and GRPO in accuracy and efficiency, reporting 4-6 point gains with Qwen3-8B+PPO.
In briefCurvSpec: Adaptive Multi-Curvature Learning for Partial Relevant Video Retrieval
SummaryThis paper addresses Partially Relevant Video Retrieval, which finds long untrimmed videos containing a moment matching a text query, without temporal annotations. It proposes CurvSpec, which learns content-adaptive curvature for video representations instead of a fixed geometry, using parallel Euclidean and hyperbolic attention with content-aware fusion. Videos are represented by semantic centroids matched to queries by geodesic distance. The authors report state-of-the-art retrieval results on ActivityNet Captions, TVR, and Charades-STA.
In briefRESCUE: Fixing Language Model Errors by Repairing Sparse Circuits with Reinforcement Learning
SummaryThis paper introduces RESCUE, a framework that locates sparse internal circuits linked to language model errors and fine-tunes only those parameters to repair the failures. It uses reinforcement learning with multiple masked-model rollouts to refine the circuit masks so they better match failures that occur during generation. On math reasoning and medical QA repair sets, the identified circuits have densities of 1.40% and 1.44%, and repair-set accuracy rises from 6.0% to 75.5% and from 0% to 81%.
WhyExtends circuit analysis to error repair with RL-refined masks, a notable idea; validated on math and medical QA, though with limited scale and comparisons.
In briefProposal-Conditioned Refinement Flows for Improving Diffusion Policies
SummaryThis paper proposes PReFlow, a policy extraction method for offline reinforcement learning that pairs critic-based proposal selection with a conditional refinement flow. The refinement flow can represent multiple high-value action modes for the same proposal, while a Gaussian reference centered on the proposal limits large action changes. Evaluated on 50 OGBench tasks, the authors report competitive offline performance and the highest aggregate score among compared methods after online fine-tuning, reaching 91% after 500K environment steps.
In briefMeteoVerse: A Unified Weather-Controllable Video World Model
SummaryMeteoVerse is a video world model that generates future videos from a single sunny or adverse-weather image, guided by a weather-free scene description, a target-weather instruction, and a camera trajectory. It explicitly estimates observed and target weather states and uses a transition-aware mixture of weather experts to preserve, introduce, or remove weather while allowing fine-grained control of intensity. The authors also release a dataset of over 50K real-world weather video clips with annotations. Experiments report much better weather controllability while keeping competitive scene consistency and camera control.
In briefPredicting FDA approval of drug-disease pairs using only public pre-trial data
SummaryThis paper asks whether FDA approval of a drug for a specific disease can be forecast using only information available before clinical trials. The authors built a cohort of 14,134 drug-disease pairs from ClinicalTrials.gov, enriched with molecular, disease, ADMET, target and pathway features, and trained a classifier on pre-2019 data. On 2,637 post-2019 programs, the model prioritized an eventually approved program about three times out of four, and choosing the top-ranked disease for multi-disease drugs identified an approved disease 61.3% of the time versus 40.7% at random.
In briefGeometry-conditioned fixed-scaffold encoder for time-warp robust sequence retrieval
SummaryThis paper studies a source of instability in patch-based sequence encoders, where patch boundaries chosen from signal geometry shift under temporal distortion. The authors propose GeoPatch, which keeps token positions fixed and uses geometric descriptors only as continuous conditioning variables. Experiments on ECG, speech, and multivariate time-series retrieval show improved early-rank retrieval under timing variation. The paper also notes a trade-off between local surface matching and strict non-overlap retrieval.
In briefSpotter: Embodied Model Leads, Vision-Language Model Reflects and Corrects Errors
SummaryThe paper finds that current embodied robot models rarely repair their own failures on retry. The authors propose Spotter, in which the embodied model executes continuously while a vision-language model (VLM) monitors in parallel and intervenes only when an error is confirmed, reflecting on and correcting it before returning control. Experiments on RoboCasa, RoboTwin 2.0 and a real robot show improved success rates, with successful episodes taking only modestly longer than the embodied model alone.
WhyProposes reversing VLM-led control so the embodied model leads and the VLM intervenes only on detected errors, with gains on RoboCasa, RoboTwin, and a real robot.
In briefCANTO: A Transformer Model That Predicts Physical Fields Directly from CAD Geometry
SummaryThe paper introduces CANTO, a transformer neural operator that takes continuous CAD geometry, represented as NURBS patches, as input and predicts surface and volume physical fields without meshing. It is evaluated on four automotive and aircraft aerodynamics benchmarks, achieving state-of-the-art accuracy on most tasks. Because the model is differentiable with respect to CAD parameters, it also supports gradient-based inverse design.
WhyDirect NURBS-to-field neural operator avoiding meshing, with strong results on four aerodynamics benchmarks and CFD-verified inverse design.
Daniel Leibovici, Nikola Borislavov Kovachki, Dawon Ahn et al.
In briefUpliftMem: Learning Set-Level Gains for Agent Memory Retrieval
SummaryLLM agents reuse external memory for new tasks, but must learn which memory sets actually improve execution. UpliftMem learns retrieval from the execution uplift of memory sets relative to the same executor without memory, allocating limited training rollouts by an expected value of sample information criterion. Across ALFWorld, WebShop, and BigCodeBench, it achieves the best success rates among the evaluated baselines on the main evaluation sets.
In briefVAA-CSEC: Vote-guided advantage allocation for Chinese semantic error correction
SummaryThis paper proposes VAA-CSEC, a multi-stage framework for correcting semantic errors in Chinese text. It combines chain-of-thought distillation, supervised fine-tuning, reinforcement learning, and self-consistency decoding. The authors design a reward function aligned with minimal editing and introduce GLPO, which reallocates advantages to match the inference-time self-consistency objective. Reported results include an F0.5 of 47.72% on CSED-C and 41.55% on NaSGEC-Exam.
In briefEGSD: Event-Grounded Self-Distillation for Streaming Video Understanding
SummaryThis paper addresses how models keep an up-to-date memory of streaming video for real-time understanding. It proposes Event-Grounded Self-Distillation (EGSD), which treats memory as incremental updates over verifiable events such as key entities, actions, and details. Reported results on mainstream benchmarks are 79.8% on StreamingBench and 73.4% on the OVO-Bench Real-Time track.
In briefEasyPPO: Stabilizing the Critic Is Key for Training Large Language Models
SummaryThis paper finds that the learned critic in Proximal Policy Optimization (PPO) is a major source of instability when reinforcement learning is applied to large language models, and identifies two critic failure modes. The proposed EasyPPO method filters overlong truncated rollouts only on the actor side, weights each prompt's critic loss by the inverse standard deviation of its returns, and uses moderately smaller critic mini-batches. Across FrontierCS, AIME24, and Search-R1, the authors report that EasyPPO stays stable over training and improves best validation scores over PPO by 14.89%, 2.28%, and 9.47%, respectively.
In briefScene Retargeting: transferring object layouts between rooms by analogy
SummaryThis paper introduces Scene Retargeting, the task of transferring semantically coherent spatial organization of objects from one 3D indoor layout to another. Instead of relying on text descriptions or pairwise object relations, the method transfers object clusters and uses the spatial distribution of foundation features to preserve each cluster's semantic context, then applies physical constraints such as wall contact, pairwise alignment, and clear passageways. The authors report that it outperforms state-of-the-art methods for layout generation on the 3D-FRONT dataset and demonstrate applications including real-to-sim transfer, analogical trajectory transfer, and multi-reference composition.
In briefAI as a Compiler: Generating GPU PTX Code from Triton Kernels with an LLM
SummaryThis paper studies "AI lowering," where a large language model agent translates Triton GPU kernels directly into NVIDIA PTX code instead of using the conventional compiler pipeline. Across twelve common kernels and ten kernels from recent ML papers, the approach reaches 0.83x to 3.34x the performance of autotuned Triton. The authors also extend the Volta PTX verifier to model Blackwell features such as tensor memory and asynchronous execution.
Fran\c{c}ois Costa, Charly Castes, Thomas Bourgeat et al.
In briefDiagnosing and fixing a shortcut where omni-modal AI leans on images for audio questions
SummaryThe paper finds that omni-modal LLMs, when answering audio-related questions, rely on the image nearly as much as the audio, and sometimes more. The authors introduce the Factorized Modality Diagnostic, which swaps audio and images between samples to measure each modality's causal contribution. They propose DMC-Repair, a training method that trains on cross-modal swapped samples and assigns supervision according to the modality the question specifies, reducing image-induced answer effects by 59.9% without hurting audio question answering.
In briefWhere randomness matters in neural cellular automata: whether random cell updates help learning and whether they are still needed at run time.
SummaryThis paper studies how random cell updates in neural cellular automata (NCA) affect both learning a useful rule and execution. Using controlled Growing NCA experiments, the authors separate the update modes used in training and evaluation. Asynchronous training passed the short-horizon quality test more often than synchronous training, and most asynchronous models kept the target under deterministic evaluation.
In briefRAE-PPG: Duration-Based Retain-and-Extend Pretraining for PPG Foundation Models
SummaryThis paper introduces RAE-PPG, a pretraining method for photoplethysmography (PPG) foundation models. A single Transformer encoder is trained successively on 10 s, 30 s, and 240 s signals, adding supervision for features each longer duration supports while reusing earlier learned parameter groups. Across 18 tasks from eight datasets, the final frozen encoder achieves the best observed score on 12 tasks compared with five existing PPG foundation models.
In briefGitHarness: Git-style versioned working memory for changing user requirements
SummaryThe paper addresses long-horizon LLM agent tasks where user requirements are supplemented, added, or revised mid-task. It proposes GitHarness, a pluggable framework that stores requirement states and their work states in a branchable, Git-style version history. A trainable Git Agent identifies requirement changes and selects a compatible historical state, which a unified version interface restores and branches from so execution focuses on affected parts. The authors also introduce MTAgentBench, a benchmark spanning several task domains.
In briefPerfect Reconstruction Cannot Reveal If Two VAE Codebooks Agree
SummaryMany systems read a variational autoencoder's latent space as discrete codes and assume each latent is decoded as the code it was written as. The paper proves that no item in a standard training report can show whether written and read codes agree on encoded data. It builds two VAEs with perfect reconstruction, exact inference, and identical ELBO and code histograms, yet with agreement of 1 and 0. The ELBO's terms bound how far agreement can shift, and the authors test this on probabilistic PCA and 240 trained VAEs, proposing a protocol called Certify-or-Measure.
WhyProves standard training reports cannot reveal code agreement, with counterexamples matching all ELBO metrics, plus ELBO-derived bounds validated on pPCA and 240 trained VAEs.
In briefUsing large language models to compile task context into Bayesian optimisation
SummaryThe paper studies whether large language models can generate and run Bayesian optimisation programs that turn domain knowledge and external observations into executable code for standard BO. The authors propose HarBO, a BO-specialised harness that compiles context through a validated multi-stage workflow, and analyse regret under imperfect compilation and the effect of adding new context. Across synthetic and real-world benchmarks, LLM harnesses achieve performance competitive with specialised embedding-based and direct LLM-in-the-loop BO methods. General coding harnesses work well in familiar domains such as hyperparameter optimisation but fall short in unfamiliar, context-rich ones.}
In briefPolicy gradient with learned Gaussian mixtures for games with continuous actions
SummaryThe paper proposes a scalable policy-gradient algorithm for large sequential games with continuous or mixed discrete-continuous actions. It combines magnetic mirror descent with a mixture-of-Gaussians parametrization, trained through self-play. The authors report that it approximates equilibrium where gradient descent fails, outperforms neural fictitious self-play, and matches or beats policy space response oracles with 3.5 to 5.5 times fewer samples. In heads-up no-limit Texas hold'em, it performs on par with Slumbot.
Ond\v{r}ej Kub\'i\v{c}ek, Viliam Lis\'y, Tuomas Sandholm
In briefDecoding Affective Nuances: Improving Multimodal AI's Ability to Tell Similar Emotions Apart
SummaryThis paper examines why multimodal large language models struggle with emotion reasoning, especially distinguishing closely related emotions from fine-grained visual cues. It proposes DAN (Decoding Affective Nuances), a training-free inference-time framework with two components: a Hierarchical Emotional Reasoning Chain and Contrastive Discriminative Visual Pruning. The authors report that DAN improves emotion discrimination on several benchmarks, with a +10.47% gain using Qwen3-VL-8B-Instruct on the WebEmo25 dataset of 25 fine-grained emotion categories.
In briefAperture: Merge-Consistent Rotary States for Compressed Tokens
SummaryThe paper addresses how rotary position embeddings lose positional information when several tokens are merged into one during compression. It proposes Aperture, which stores Fourier moments of each token's weighted support at the model's rotary frequencies, and proves these have minimal real dimension among sufficient continuous states. In a prespecified native video question-answering comparison, stored support reached 65.63% accuracy versus 67.12% for the deployed merging rule.
In briefDRHeC: RGB-based differentiable rendering for robot hand-eye calibration
SummaryThis paper proposes DRHeC, an RGB-based differentiable rendering framework for calibrating the hand-eye transformation of robotic arms. By incorporating color and mask geometric features, it gives richer cues that improve calibration accuracy and optimization stability. The method is validated in simulation and real-world experiments, achieving grasping and insertion success rates of 88.9% and 57.4% on a UR5e, outperforming the prior method EasyHeC by 46.3 and 48.1 percentage points.
In briefCode4Scene: A benchmark testing coding agents on building and editing 3D scenes
SummaryCode4Scene is a benchmark of 190 Unreal Engine cases that tests whether coding agents can build 3D scenes from language descriptions and edit existing scenes to match reference images. It scores the generated engine-native scenes rather than code or renders, checking task fulfillment, artifact integrity, and physical validity. Results show construction and editing performance are correlated but distinct, and editing remains imprecise, with many edits introducing unintended changes.
Xiaokang Ye, Siddhant Hitesh Mantri, Zimeng Chen et al.
In briefCausal-EVC: Removing Spurious Emotion-Cause Links via Spatiotemporal Grounding and Counterfactual Intervention
SummaryThis paper addresses emotional video captioning, where existing methods rely on simple attention matching and can be misled by background co-occurrence, such as labeling sadness as joy on a sunny beach. The authors build EVC-CauseGround, a benchmark with dense spatiotemporal causal annotations and a Causal-Faithfulness Subset for measuring genuine emotion-cause attribution. They also propose Causal-EVC, which uses motion-guided causal localization, interpretable sparse emotion routing, and a counterfactual contrastive objective. Experiments reportedly show the best semantic metrics and clear gains on the causal-faithfulness subset.
In briefDSPO: Diversity-Aware Subjective Policy Optimization for Robust Emotional Reasoning
SummaryThis paper proposes DSPO, a reinforcement learning framework that improves emotion reasoning in multimodal large language models. It builds a context-grounded emotional distribution prior in VAD space and uses a diversity reward to keep plausible subjective emotion interpretations, along with a visual intervention gate that down-weights interpretations lacking visual evidence. The authors report state-of-the-art results on multiple public benchmarks, with a 10.8% average cross-domain accuracy gain over EMO-R3.
In briefBeyond Conditional Independence: Root-Cause Analysis Using Deep Causal Models
SummaryThis paper introduces RCA-DCM, a method for finding the root cause of faults by comparing anomalous system observations with normal ones. It models the normal system as a causal model and treats anomalies as changes in its structural functions, using neural networks and handling unobserved confounders. The authors report higher top-1 accuracy than the strongest baseline on Sock Shop and Online Boutique, and better exact root-cause recovery when the true cause is latent.
Md Musfiqur Rahman, Kenneth Lee, Ziwei Jiang et al.
In briefSelf-trained vision networks spontaneously specialize without a shared controller or cross-network gradients
SummaryThe paper studies whether independently trained neural networks can develop a division of labor when each has its own weights and asks others for help through forward passes only. The agents are finetuned on an unlabeled mix of six visual domains, and specialization is measured by whether the best agent for an input matches its latent domain. Semantically routed populations outperform a single generalist, and local routers pick the emergent expert for 98% of inputs.
Aram Davtyan, Pablo Acuaviva, Sebastian Stapf et al.
In briefCan Fixed Prefixes Replace LoRA Adapters in Frozen Attention Heads?
SummaryThis paper asks whether a fixed continuous prefix can replace a given low-rank adapter (LoRA) while the attention head stays frozen. It identifies three conditions (observability, realizability, and implementation) that determine when this is possible, and tests them with exact resource-capped analysis. On three first-layer GPT-2 attention heads, the capped optimum leaves 18.4% to 74.2% of the projected adapter effect uncompiled. All claims concern local approximation at a single head, not whole-network equivalence.
In briefGraph-spectral flow matching for detecting anomalies in multivariate time series (GRASP)
SummaryThis paper proposes GRASP, a flow matching framework for anomaly detection in multivariate time series. It builds graph structure into the probability path by minimizing a fixed-endpoint action combining kinetic and graph Dirichlet energy, yielding a closed-form path. A velocity predictor trained on normal data flags anomalies through weighted velocity discrepancies, and experiments on four benchmarks report superior detection performance.
Zepeng Zhang, Jhony H. Giraldo, Wenbin Wang et al.
In briefFederated clustering when neither local nor global cluster counts are known
SummaryThis paper addresses federated clustering where neither the server nor the clients know the number of clusters, a situation arising in fault diagnosis across independently operated industrial sites. The authors propose a two-phase framework in which each client first estimates its local cluster count, and these estimates are then used by aggregators such as FedGEM. The first phase introduces Adaptive Split–Merge (ASM), which uses BIC-driven splitting of a spherical Gaussian mixture followed by merging, without labels. Across eight datasets, ASM with FedGEM reaches a mean ARI of 0.333.
Mitushi Goyal, Tarun S., Riddhanya Senapathi et al.
In briefQuantMLA: Function-aligned dual-path quantization to shrink MLA attention caches to low-bit precision.
SummaryMulti-Head Latent Attention (MLA) keeps separate caches for content and decoupled RoPE paths, but cache memory still grows with context length and batch size. The paper analyzes how quantization errors in each path distort attention outputs, then proposes QuantMLA, which learns path-specific transformations with function-aligned objectives. The authors report joint INT4 quantization of both caches across four MLA model families with minimal accuracy loss, plus a 3.59x compression at 128K context. They also describe a native low-bit attention kernel and a 5.168x throughput gain over BF16 in a cache-pressure serving workload.
In briefDuLBE: A Dual-Mode Low-Rank Learner with Bridge-Prototype Ensemble for Vision-Language Class-Incremental Learning
SummaryThis paper proposes DuLBE, a method for exemplar-free class-incremental learning built on CLIP. It uses gradient routing to split visual low-rank updates into a reusable shared mode and residual modes for task-specific changes, reducing knowledge overwriting. It also builds geodesic bridges between visual prototypes and text embeddings on the unit hypersphere, and ensembles reliable bridge prototypes to address the modality gap. The authors report state-of-the-art incremental learning performance across multiple settings while keeping high parameter efficiency.
In briefFastVR: Fast, Smooth Video Restoration in One Diffusion Step
SummaryThis paper presents FastVR, a streaming video restoration framework built on a one-step diffusion model. It uses a lightweight VAE and chunk-wise causal attention to cut computation, and adds velocity consistency regularization and continuous trajectory learning during training to improve quality. The authors report processing 1080p video at 11 FPS on a single H20 GPU and state-of-the-art results on synthetic and real-world benchmarks.
In briefNesTok: A Nested Self-Aligned 1D Tokenizer for Image Generation
SummaryThis paper presents NesTok, a nested self-alignment framework for dynamic 1D visual tokenizers that can use variable token lengths. It introduces cross-length training, which optimizes reconstruction across token lengths and uses the full-length sequence to guide shorter ones. On ImageNet, NesTok achieves an rFID of 0.98, and for autoregressive image generation at 256×256 it reports a gFID of 1.46, the best among existing variable-length methods.
In briefDrag as Evidence: Motion-grounded latent recomposition for precise drag-based image editing
SummaryThis paper proposes MoRe-Drag, a drag-based image editing method that treats coarse pixel-space warping as motion evidence and injects it into the generative sampling process. It performs region-aware latent recomposition over refinement, inpainting, and preservation regions, with stage-adaptive conditioning that shifts from motion-grounded structure to semantic refinement. Experiments on DragBench-SR and DragBench-DR report improved drag precision over base editors and strong semantic consistency and visual realism.
In briefDoes a prosody-trained representation beat trainable fusion? A parameter-matched study with frozen HuBERT
SummaryThis paper tests whether explicit prosodic cues improve speech recognition. It uses a frozen HuBERT backbone and a 64-dimensional representation trained to predict pitch, voicing, energy, and spectral tilt. On Buckeye, Switchboard, and AMI IHM, the learned representation shows no significant word error rate gain over a parameter-matched control, though removing it at inference increases errors.
In briefcktFormer: A Transformer-based method for automating analog circuit design
SummaryThis paper presents cktFormer, a dual transformer architecture for automated analog circuit design. It converts netlist data into graph-based representations, with a node prediction model proposing components and an edge prediction model inferring valid connections. The authors report that it outperforms recent models such as AnalogGenie and cktGNN in the validity of generated circuits.
Pasindu Dodampegama, Praveen Wijesinghe, Naveen Basnayake et al.
In briefGroup-Marginalized Self-Rewarding RL for Label-Free Self-Improvement of LLMs
SummaryThis paper studies self-rewarding reinforcement learning, which lets large language models improve themselves without human labels. The authors note that existing group-based methods compute rewards using one randomly sampled group of other responses, which may miss useful reward signals. They propose Group-Marginalized Advantage Estimation (GMAE), which aggregates rewards across possible group contexts into a response-level distribution to estimate expected advantages. Experiments on eight benchmarks and four base models reportedly show strong performance and cross-domain generalization.
In briefLifelong Editing of LLM Knowledge That Generalizes Better
SummaryThis paper addresses updating specific facts in large language models through continual, lifelong editing rather than one-off changes. It notes that existing methods overfit to target prompts, weakening generalization of edited knowledge and general model abilities. The authors propose GLIME, which combines knowledge editing with preference optimization over generation behavior, plus replay-based editing and a gradient constraint to retain earlier edits. Experiments reportedly show improved knowledge generalization while maintaining editing performance and general capabilities.
In briefEASE: A Method That Adapts Skill Curation to Executor Behavior for Self-Evolving Agents
SummaryThis paper observes that existing learned skill curators for self-evolving agents do not explicitly model how downstream executors behave, so curators trained with one executor degrade when paired with others. The authors propose EASE, which trains one shared curator that uses an online behavioral profile, the current trajectory, and retrieved skills to add, modify, or remove skills, using reinforcement learning across multiple frozen executors. Reported results on ALFWorld, ScienceWorld, and WebShop show EASE outperforming skill- and memory-based baselines while maintaining 34.5–41.0% fewer skills.
In briefSIPO: A Training Method Unifying Reinforcement Learning and On-Policy Self-Distillation for LLMs
SummaryThe paper proposes self-instructing policy optimization (SIPO), a method for training large language models that addresses the lack of token-level credit assignment in reinforcement learning with verifiable rewards. SIPO uses a contrastive self-teacher that pairs reference answers with mistakes from the same group to give dense, token-level feedback on each rollout. According to the authors, SIPO outperforms RLVR and on-policy self-distillation baselines across multiple reasoning and code-generation benchmarks without requiring an external teacher or extra generation.
In briefDistinguish or Homogenize: Last-Chance Policy Identification for Irreversible Resource Depletion
SummaryThis paper studies how an agent facing irreversible resource depletion can either spend resources to tell apart latent fault models or change the system state so the remaining models share an acceptable continuation. It introduces Last-Chance Policy Identification (LCPI), an exact recursion for deterministic diagnostic graphs, and Risk-Budgeted Compatibility Planning (RBCP) for noisy finite-horizon recovery. In experiments on microservice incident recovery and latent-damage navigation in MiniGrid, RBCP improves risk-feasible recovery while respecting a failure budget.
In briefOffline learning of ranking policies by splitting top-k selection from the rest
SummaryThis paper addresses off-policy learning of ranking policies, which learns new rankings from historical logged data, a task that is hard because the action space of permutations is extremely large. Existing policy-based methods can have high variance, while regression-based methods can be biased. The proposed R-POD method splits a ranking policy into a first-stage policy selecting the top-k items, learned with a new policy gradient estimator, and a second-stage policy choosing the remaining items, learned by regression. The authors state the first-stage estimator is unbiased under a conditional pairwise correctness condition.
Ren Kishimoto, Koichi Tanaka, Haruka Kiyohara et al.
In briefFrontier Autolab: a simulated AI firm testing memory, dissent and hindsight bias across fifty years of tech change
SummaryThe paper builds a long-horizon simulation in which a multi-agent LLM firm, with sixteen role personas and a Red Team, repeatedly re-founds itself across nine technology eras from 1990 to 2040. Each decision is made from a dated briefing, then a historian-judge reveals what happened and scores it on five dimensions, with lessons stored in a persistent Playbook. The authors report a consistent gap where the firm recognized coming shifts better than it chose where to build, and that scores rise over eras while the judge's hindsight subscore falls. They conclude that apparent learning may be confounded with recall of history.]
In briefBackpropagated Output Momentum: moving optimizer history from parameters to task outputs
SummaryThe paper introduces Backpropagated Output Momentum (BOM), which stores a compact moving average of prediction errors at the model output instead of a parameter-sized gradient average, and reprojects it through the current network each step. A batch-level analysis describes what information this relocation keeps and omits, and BOM can serve as a plug-in for momentum-based optimizers. The authors report 49.7–99.8% reductions in parameter-shaped optimizer state, a 4.0% average step-time increase across three language backbones, and a 1.42-point mean validation gain in a five-task language and vision fine-tuning comparison.
In briefRebuilding the Vocal Tract Shape from Sound Using Differentiable Acoustic Simulation
SummaryThis paper presents a differentiable, GPU-accelerated acoustic simulator that synthesizes speech by propagating sound through a tube model of the vocal tract. Its gradients allow the vocal tract shape to be reconstructed from the sound alone. The authors report that three contributions, a frequency-domain fluid formulation, a differentiable turbulence model, and a neural-network geometry parameterization, make gradient descent succeed. Applications shown include self-supervised autoencoding of vocal tract shapes across 11 languages and reconstructing moving vocal tract MRI from speech without paired data.
In briefFrom Neurons to Conversation: A Systems Review of Speech Brain-Computer Interfaces
SummaryThis review examines speech brain-computer interfaces, which convert neural activity linked to speech, language, or communicative intent into text, synthesized voice, or avatar control to help restore communication. It covers the neural substrates of speech, recording and decoding choices, closed-loop adaptation, evaluation, clinical translation, and ethics. The authors argue that future systems should be judged not only by offline accuracy but also by cross-session robustness, calibration burden, latency, uncertainty, and usability.
Moein Khajehnejad, Forough Habibollahi, Tommaso Boccato et al.
In briefBiFE: LLM-driven two-fidelity search for CPU-only branching rules in MILP solvers
SummaryThis paper studies how to choose branching variables in branch-and-bound for mixed-integer linear programming, a key factor in solver speed. It proposes BiFE, a bi-fidelity evolutionary framework in which large language models generate lightweight CPU-friendly branching rules. Low-fidelity imitation scores pre-screen candidates, and full on-instance evaluation is applied only to elite ones. The authors report that the discovered rules outperform the SCIP solver and other baselines on CPUs, and some surpass GPU-based neural policies.
In briefExplainer: A confidence-aware selective distillation method for large language models (CaRE-KD)
SummaryKnowledge distillation trains a smaller student model to imitate a larger teacher, but in large language models teacher predictions can be high-entropy and hallucinated, which can degrade the student. The paper proposes CaRE-KD, which adaptively switches between forward and reverse KL divergence based on teacher-student confidence at the token level, and suppresses batch updates when the teacher is more uncertain than the student. The authors report consistent gains over strong baselines across eight teacher-student pairs and eleven benchmarks.
In briefCan AI Agents Design Code Libraries for Other AI Agents?
SummaryThis paper introduces LibraryDesignBench, a two-phase benchmark in which an agent implements a full-featured library from a specification listing required capabilities and use cases, without prescribing the design. The library is judged by the correctness and simplicity of programs that three user agents from different model families write with it. The benchmark covers 242 expert-validated programming problems across 15 library-design tasks in four languages.
In briefCan AI scientists revise beliefs when evidence conflicts with prior knowledge?
SummaryThis paper introduces Synthetic Universes, a benchmark that pairs famous physical worlds with matched twisted versions governed by nearby noncanonical mechanisms. Each law an agent reports is tested both by its predictive accuracy on held-out data and by whether it recovers the true generating mechanism. The completed runs show these two properties can dissociate, so predictive adequacy and mechanism recovery require separate tests. The authors do not claim a confirmed causal prior-conflict effect, since key controls remain incomplete.
In briefRouting in Gradient Space: Balanced Usage Is Not Expert Specialization. A study of how sparse expert models can route training signals by gradient alignment.
SummaryThe paper treats routing in sparse expert models as a gradient-partitioning problem and proposes gradient-aligned routing (GAR), which groups observations with aligned gradients. On five multi-task text-classification mixtures with a fully trainable RoBERTa backbone, GAR achieves the highest aggregate validation accuracy, 1.07 points above task-loss-only routing. It also ranks first with frozen DeBERTa and Qwen3-1.7B backbones and adapter experts, with better-balanced load and higher gradient-mass purity.
In briefATTUNER: Reusing Agent Context KV States Without Recomputation
SummaryLLM agents repeatedly load reusable content such as skills, documents, and memory into their context, and re-encoding it each time wastes computation. The paper proposes ATTUNER, a query-side method that inserts low-rank adapters into a frozen artifact cache and is trained by distilling full-prefill distributions. It trains fewer than 0.05% of parameters and needs no cache recomputation at inference. On seven benchmarks with Qwen3-4B and Qwen3-8B, it outperforms prior position-independent caching baselines and matches full-context prefill quality with up to 3.73x speedup.
WhyLocalizes PIC quality loss to attention rather than KV recomputation, and validates a query-side adapter across seven benchmarks with speedups.
In briefCALIBUDGET: Calibration-Guided Data Source Allocation for Fixed-Budget Adaptation
SummaryThis paper introduces CALIBUDGET, a method for deciding how many training examples each of several heterogeneous data sources receives under a fixed budget. It uses small calibration splits taken from the training data, combining model need, availability after a floor is applied, and bootstrap stability to produce integer quotas that respect capacity limits, without changing the model, objective, or total budget. In a setting combining mathematical and commonsense data, it improved commonsense and overall averages in all three paired LLaMA-2-7B LoRA+ runs, while the mathematics average decreased slightly.
In briefSIRNET: Adapting Pretrained Embeddings for Causal Treatment-Effect Estimation with Surrogate Signals
SummaryThe paper proposes SIRNET, a framework for estimating treatment effects when experiments produce unstructured post-treatment surrogates such as text, images, or video. It starts from pretrained embeddings and learns a shared adapter that serves outcome prediction, treatment propensity, and arm-specific surrogate marginalization. The final estimate and confidence interval come from cross-fitted orthogonal scores, which the authors state attain the semiparametric efficiency bound under stated nuisance-rate conditions. The method is validated on synthetic, semisynthetic, and A4 Alzheimer's trial data, using frozen brain-MRI embeddings as surrogates for a long-term outcome.
In briefInformation-based receding-horizon active learning for identifying nonlinear dynamical systems
SummaryThis paper studies how to collect informative data efficiently from a single trajectory when learning a stochastic controlled nonlinear dynamical system. It builds a parametric estimator with fixed nonlinear features and recursive Bayesian updates, quantifying predictive uncertainty online. The adaptive experiment design is posed as a receding-horizon problem solved with the cross-entropy method. Numerical tests on a noisy multistable system show it reduces uncertainty and reconstruction error more efficiently than common excitation baselines.
Juncal Arbelaiz, Anushri Arora, Jonathan W. Pillow
In briefStable extrapolation in high-dimensional function and operator learning
SummaryThe paper studies regression when test and training distributions differ, asking when stable extrapolation is possible for the target function or operator. It identifies holomorphic function and operator classes whose out-of-distribution error converges at algebraic rates even under large distribution shifts, attributed to increasing smoothness in higher-index coordinates. The authors derive explicit rates for polynomials, deep neural networks and deep neural operators, and support the results with numerical experiments.
In briefLAURA: Knowledge Distillation for Explainable Identification of Ambiguous Clauses in Legal Contracts
SummaryThis paper proposes LAURA, a framework that identifies ambiguous clauses in legal contracts and explains its reasoning. It uses knowledge distillation with an IRAC-Unlearning prompting technique to transfer knowledge from a teacher LLM to an open-weight student model with at most 1B parameters, trained on classification and rationale generation losses. According to the authors, LAURA with Flan-T5 (250M) achieves the best interpretability among interpretable baselines while matching the identification performance of the strongest opaque baseline.
In briefJudgeProfile: Understanding and Steering Subjectivity in LLM Judges
SummaryThis paper introduces JudgeProfile, a framework that splits LLM pairwise evaluation into perception (how judges compare responses on attributes such as clarity and correctness) and prioritization (how much each attribute affects the final choice). The authors built SubjectiveSet, with 50,013 response pairs from 17 public sources, evaluated by 21 LLM judges across 87 attributes. Judges often agree on attribute judgments even when their overall choices differ. Reweighting attributes using reference labels raised average held-out agreement from 66.48% to 71.97%.
In briefWhen Coarse Labels Pay Off: Cost-Aware Learning with Unknown Aggregation
SummaryThis paper studies learning with two label resolutions: expensive fine labels reveal a vector response, while cheaper coarse labels reveal a scalar aggregate with unknown weights. The authors characterize the information geometry of this setting and propose an estimate-and-track policy that learns the aggregation rule and tracks the optimal resolution mix. They derive a closed-form break-even condition for coarse supervision and show the online policy is asymptotically optimal, and synthetic experiments show a finite-budget gain over all-fine acquisition when coarse supervision is favorable.
WhyDerives a closed-form break-even condition and matching minimax lower bound for cost-aware two-resolution learning under unknown aggregation, a notably novel theoretical contribution.
In briefLost in Conversation or Translation? Multi-Turn RAG Degradation
SummaryThis paper examines how retrieval-augmented generation (RAG) and GraphRAG perform in multi-turn conversations, which current evaluations largely overlook. The authors turn multi-hop question answering questions into underspecified dialogues and test ten LLM assistants with eight retrieval systems across 1.5 million simulated conversations. Multi-turn interaction causes widespread degradation, with relative performance drops of up to 21% and a 47% increase in unreliability. The authors identify two failure modes: conversational rephrasing distorting retrieval queries, and the LLM failing to synthesize evidence spread across turns.
In briefDual-headed transformers that replace the value projection with structured transforms
SummaryThe paper proposes dual-headed transformers, which replace the third head of standard attention with a structured transform to cut parameters and cache memory. The authors examine the Walsh-Hadamard, discrete cosine, discrete Fourier, Shearlet, and multiplication-avoiding operators as the second head. They report better ImageNet results than triple-headed transformers.
In briefKnow Thyself, Teach Thyself: Internal Information Flow for Selective Self-Distillation
SummaryThis paper introduces InFlow, a self-distillation framework that needs no external teacher. It retrieves potentially useful sources using certainty-calibrated hidden-state trajectories, then measures their realized effect via the Jensen-Shannon divergence between the teacher's initial and retrieval-conditioned answer beliefs. Examples with larger belief shifts are selected for on-policy distillation. The authors report the strongest cross-model average among compared selection methods across four open-weight language models and three knowledge domains.
In briefAESplat: Better pose-free 3D scene rendering by separating view-independent and view-dependent appearance
SummaryAESplat is a framework for pose-free feed-forward 3D Gaussian Splatting, which generates novel views of a scene without requiring camera poses. It separates appearance into a base view-independent component, taken directly from the input images, and view-dependent components, predicted by a shallow MLP with two 3D-aware inductive biases. The paper reports PSNR gains of 0.8 dB over the pose-free NAS3R and 1.1 dB over the pose-required DepthSplat on RealEstate10K.
In briefNormalize-then-precondition: a hierarchical optimizer for training large language models
SummaryThe paper revisits the Muon matrix optimizer through its full-Gram representation and proposes the Normalize-Then-Precondition (NormPre) framework, which first applies marginal normalization and then spectral preconditioning. It introduces two variants, NormPre-G and NormPre-L, and provides O(T^-1/2) convergence guarantees for simplified versions. In pretraining experiments on GPT-2 Small, LLaMA and Qwen3, both variants outperform AdamW, Muon and MANO under matched training budgets. The code is open-sourced on GitHub.
In briefVideo2Skill: Turning Streaming Video Experience into Reusable Robot Skills
SummaryThis paper introduces Streaming Embodied Skill Discovery, where a model watches videos in sequence and maintains a persistent library of skills. The authors release Video2Skill, a benchmark on robot tabletop manipulation and human kitchen activity that tests event localization, grouping of same-transformation events, and deciding whether to reuse or create skills. Across 19 open-source vision-language models, many group events near chance, and fine-tuning improves grouping but models rarely expand their libraries. The authors identify recognizing when existing skills are insufficient as the central challenge.
In briefCHAIN: Calibrated LLM Forecasting via Causal-Temporal Hypergraph Inference
SummaryThis paper addresses systematic calibration bias in large language models' probability outputs for event forecasting, which varies across domains and question types. The authors split probabilistic prediction over causal-temporal hypergraphs into three stages (evidence weighting, evidence aggregation, and source fusion) and propose CHAIN, which applies stage-specific mechanisms to reduce bias. Experiments on cross-domain forecasting benchmarks report that CHAIN outperforms existing methods in expected calibration error, Brier score, and accuracy.
In briefMyoCodec: A Streaming Neural Codec for Muscle Electrical Signals
SummaryThis paper presents MyoCodec, a streaming neural codec for electromyography (EMG), the recording of muscle electrical activity. It combines causal Transformers with residual vector quantization to encode EMG signals into representations from continuous features to discrete tokens at 50 Hz. Trained on twelve public EMG datasets, it performs well on codec quality and downstream tasks such as typing, hand pose, and speech decoding. It needs only 0.482 ms per 20 ms frame, enabling real-time streaming, and code and model weights are released.
WhyAdapts causal Transformers with residual vector quantization from audio codecs to EMG, validated across four downstream tasks against prior models.
Jihwan Lee, Kleanthis Avramidis, Junhyeok Lee et al.
In briefWhy root cause analysis fails: splitting retrieval from reranking
SummaryThis paper argues that the common top@k accuracy metric hides two different failures: the true cause is never retrieved, or it is retrieved but ranked too low. The authors split root cause analysis into a retrieval stage and a reranking stage and audit four benchmarks. Their two-stage pipeline, which combines a multi-signal retriever with an LLM reranker, matches or exceeds the best baseline's top@1 accuracy on all six benchmark suites without needing a causal graph or labeled data.
WhyDecomposes top@k into retrieval and reranking failures, validated across six benchmarks with a retrieval-plus-LLM-reranker pipeline, useful for root cause analysis.
Hada Melino Muhammad, Luan Pham, Laure Barri\`ere et al.
In briefBalancing Meaning and Rhythm: Reliability-Aware Control for Speech-Synchronized Gesture Generation
SummaryThis paper studies co-speech gesture generation, which aims to produce gestures that match speech timing and its semantic content. The authors propose a reliability-aware framework that first learns a discrete motion prior, then estimates how much text semantics changes predicted motion using conditional information gain between multimodal and audio-only branches. This estimate drives a controllable objective that strengthens semantic guidance in content-relevant segments. Background noise is treated as an acoustic reliability condition to improve robustness, with experiments on benchmark datasets reported to balance semantics, rhythm, diversity and robustness.
In briefProgressCompass: Embodied progress reward models need the right context
SummaryThe paper addresses progress estimation for embodied agents on long tasks, where the current frame alone often cannot show how far a task has advanced. The authors build ContextProgress-Bench, with 24 manipulation tasks and 120 episodes, covering state recall, sequence tracking, and recurrence disambiguation. Supplying the right context to the same five PRMs cut their progress error by 77-82%. They also propose ProgressCompass, an agentic loop that reduces the error of a frozen PRM by 63%.
In briefMARCO: Multi-Round Agentic Reinforcement Learning for Conditional Molecular Optimization
SummaryMolecular optimization is an iterative process in which candidate molecules are proposed, evaluated on several objectives, and revised while staying related to the source molecule. MARCO trains molecular editing models with evaluator-grounded reinforcement learning over bounded proposal, feedback, and revision trajectories. Tested on the three-objective MuMOInstruct benchmark with three Qwen backbones, the SFT-initialized MARCO achieved the highest product of property success rate and similarity in every reported primary setting. Allowing up to five responses further improved the observed score.
In briefReprogramming Vision-Language Models via Structured Prompt Reparameterization: a method that improves few-shot classification with a frozen backbone
SummaryThis paper proposes RVP (Reparameterized Inter-Class Visual Reprogramming), which aggregates multiple text prompts within each class and applies residual correction across classes to improve visual reprogramming. The authors show that CLIP-based linear output aggregation can be written as a linear mapping from frozen image embeddings to downstream logits, and use this to design a structured reparameterization that models shared semantic components and class-specific differences. Across 11 few-shot classification benchmarks and four CLIP backbones, RVP outperforms prior visual reprogramming methods with comparable or better inference efficiency.
Zizhao Li, Chengyi Cai, Mohammed Yaqoob Ansari et al.
In briefMLToolBench: Training AI Agents to Use Diagnostic Tools for Machine Learning Development
SummaryThe paper introduces ToolMLBench, a set of executable tools for data inspection, code verification, and experiment diagnosis, along with an SFT and reinforcement learning pipeline that teaches agents when and how to use them. It proposes SPICE, which uses the change in a sampled tool action's likelihood under privileged context as a turn-level reward alongside the final outcome. Trained on 80 synthetic tasks, in-domain success rises from 24.8% to 52.4% for Qwen3-8B and from 35.6% to 69.2% for Qwen3.5-35B-A3B.
In briefReWorld-Track: A recursive event world model for tracking people across cameras using language cues
SummaryThis paper proposes ReWorld-Track, a model that keeps a target's identity consistent across camera handoffs, even when the target is temporarily unobserved. It treats candidate matches and continued waiting as alternative target states and updates a recurrent belief using their posterior probabilities. This belief predicts the next camera, arrival time, and entry region. The model reports HOTA scores of 65.19 on CityFlowV2 and 45.36 on MTMMC.
In briefGödel Forest: Balancing search depth and breadth for data-centric recursive self-improvement
SummaryThis paper proposes Gödel Forest, a multi-agent framework that organizes recursive self-improvement as an ensemble of co-evolving search trees. Each agent grows its own tree by deepening, branching, or pruning data strategies based on model feedback, while parallel trees explore different regions of the data space, and a shared memory distills successes and failures into compact lessons. Evaluated on RSIBench-Data across six domains, the framework outperforms the single-agent baseline by an average of 10.70% and reduces wall-clock time on five tasks. Ablations show co-evolving shared memory yields a +7.00% gain over independent parallel search.
In briefCAST: Faster language model decoding by checking more draft candidates at once
SummarySpeculative decoding speeds up large language models by cheaply drafting future tokens and verifying them in parallel. CAST packs the candidates a one-pass block drafter already scored into a tree and verifies it in a single target-model pass, without changing the model, drafter weights, or decoding rule. Tree width adapts automatically using a latency measurement rather than a width sweep. The authors report speedups over the standard chain in all eight tested settings, up to 43%, and prove the target output distribution is unchanged.
WhyPacks already-scored block-drafter candidates into a verification tree with cost-adaptive width; eight-setting GPU experiments plus a distribution-preservation proof support it.
In briefHuman-Inspired Task-Dimension-Guided Exploration for Efficient High-Dimensional Learning
SummaryThis paper proposes TDGE, a model-agnostic exploration algorithm inspired by how humans narrow large decision spaces by probing feature dimensions. It uses an automatically built task-dimension, feature, and item hierarchy and explores top-down. Experiments on MovieLens-20M, LastFM, and Amazon recommendation datasets report better exploration efficiency and cold-start adaptation than baselines.
In briefFairDiff: Reducing Self-Reinforcing Popularity Bias in Diffusion Recommenders
SummaryThis paper finds that Diffusion Recommender Models amplify popularity bias through their generative process, creating a self-reinforcing Matthew Effect. It proposes FairDiff, a plug-and-play framework with Popularity Condition Guidance, applied at inference time, and a Semantic Calibration module that aligns forward and reverse distributions via one-step optimal transport. The authors report that FairDiff achieves state-of-the-art performance while mitigating this effect.
In briefFineSID: Scalable, Efficient Semantic ID Learning for Generative Recommendation
SummaryThis paper proposes FineSID, a quantization framework for learning semantic identifiers used in generative recommendation. Existing methods rely on Top-1 hard assignment, which concentrates gradients on a few codewords, leaving many under-trained and causing SID collisions. FineSID distributes learning signals softly and differentiably across the entire codebook. The authors report that it is robust to initialization and improves codebook utilization and recommendation accuracy on public benchmarks.
In briefStochastic heavy ball with Polyak step size and Armijo line search: a convergence analysis
SummaryThis paper analyzes the convergence of stochastic heavy ball (SHB) methods using Polyak step size and Armijo line search. It introduces a modified Armijo rule closely paralleling Polyak step size and a decoupling analysis isolating momentum-induced history dependence. The authors establish expected convergence for strongly convex, convex, and non-convex objectives, and stronger almost sure and last-iterate results under interpolation or strong growth.
In briefEgoRefine: A Method for Asynchronous Collaborative Perception Using Ego-Referenced Alignment and Reliability-Aware Fusion
SummaryThis paper addresses asynchronous collaborative perception, where shared features from connected vehicles arrive with delay and remain misaligned with the ego agent's current view. The proposed EgoRefine framework uses the ego feature to guide trajectory-field prediction and refine sampling offsets, then fuses the aligned streams with adaptive reweighting based on trajectory discrepancy and refinement magnitude. On the V2V4Real and DAIR-V2X-Seq datasets, it improves average AP@0.5 and AP@0.7 over TraF-Align by 1.6 and 2.9 points, respectively.
In briefGenLimitLib: A Formal Library for Language Generation in the Limit and AI-Assisted Math Research
SummaryThis paper presents GenLimitLib, a Lean 4 library aligned with source papers, for studying language generation in the limit. This theoretical field, introduced by Kleinberg and Mullainathan at NeurIPS 2024, asks how to generate valid new strings from observed examples. The library formalizes 30 papers, extracting shared definitions and reusable proof components while recording relationships among them. Through mathematical case studies and LLM experiments, the authors show it can support both human and AI-assisted mathematical research.
In briefAutoLoCo: Communication-efficient distributed LLM training through adaptive synchronization
SummaryThis paper introduces AutoLoCo, a framework that reduces communication in distributed large language model training. It adapts the number of local optimizer steps between synchronizations using scalar training statistics and corrects the outer update to address the mismatch this creates. Under communication constraints, the authors report a 27% reduction in communication frequency relative to DiLoCo while maintaining training performance.
In briefYou Only Reprogram Once: A Faster Way to Adapt Pretrained Vision Models
SummaryThis paper addresses the high training cost of visual reprogramming, which adapts frozen pretrained models by optimizing visual prompts over many epochs. It proposes YORO, which builds a downstream predictor from the frozen model's response space in a single forward-only pass using a covariance-aware Bayesian Discriminant Mapping, with no backpropagation. Across three full-data settings, YORO improves average accuracy by 18.4 to 24.4 percent over the strongest prior gradient-free mapping method, and raises the 16-shot CLIP four-backbone average from 71.4 to 77.2 percent.
Zizhao Li, Mohammed Yaqoob Ansari, Xinyu Su et al.
In briefByzantine-Robust Federated Representation Learning: a method where each client learns its own linear head while jointly estimating a shared representation through attack-resistant aggregation.
SummaryThis paper studies federated learning with adversarial clients, aiming to minimize the average loss of honest clients without knowing which clients are honest. Each client learns a personalized linear head while a shared nonlinear representation is estimated through Byzantine-robust aggregation, avoiding the bias of a single shared model under heterogeneous data. The authors give non-asymptotic parameter recovery error bounds, extend the analysis to multiclass classification, and validate it empirically on CIFAR-10, FEMNIST, and School Exam Score datasets.
Leonardo F. Toso, James Anderson, Rafael Pinot et al.
In briefLensBridge: Frequency-Guided Adaptation for Correcting Lens Aberrations and Removing Veiling Glare
SummaryThis paper presents LensBridge, a two-stage framework for correcting residual lens aberrations and veiling glare in simplified optical systems. The first stage builds a reusable aberration-correction foundation from point spread functions in the large-scale LensLib library. The second stage adapts this foundation to compound degradation using frequency-domain guidance and only a few unpaired target observations. Experiments across multiple optical systems show joint aberration correction and veiling glare removal without paired supervision from the target system.
In briefPaper on how on-policy training's parameter update direction drives better generalization in LLM post-training
SummaryThis paper studies the cumulative parameter update direction in on-policy post-training of large language models. It finds that supervised fine-tuning moves parameters along consistent directions, while on-policy training keeps adjusting direction during training. The authors propose OPSFT, which constrains SFT updates to the direction identified by on-policy methods, and report that this transfers the generalization advantage to SFT.
In briefConstitutional adapters: inference-time tools to curb AI misalignment and misuse
SummaryThis paper shows that behavior following an explicit set of principles, or a "constitution," can be distilled from synthetic data into lightweight low-rank adapters and steering vectors. Although the training data contained no harmful requests or jailbreaks, these objects improved jailbreak defense and measured alignment, especially at long context lengths and against multi-turn attacks, outperforming prompted and steered baselines. The authors call the resulting defenses "constitutional adapters" and report they can be transferred zero-shot to post-trained checkpoints and scaled at inference time to trade off defense against benign compliance.
In briefNot Every Correction Helps: Gain-Guided Continual Test-Time Adaptation
SummaryThis paper addresses continual test-time adaptation (CTTA), where a model must keep adapting to an unlabeled test stream whose distribution shifts over time. The proposed GAIN framework is backpropagation-free: it builds a correction proposal from accumulated target statistics and uses a posterior-predictive evaluator to estimate its gain over the source prediction, setting how strongly each sample is corrected. The authors report strong accuracy, calibration, and efficiency across five benchmarks, including 61.9% accuracy on ImageNet-C and a 15.9x speedup over a representative optimization-based CTTA baseline.
In briefReplay the Curvature: Accurate, Scalable NVFP4 Quantization for LLM Inference
SummaryThe paper introduces Schur Replay, a scale-selection algorithm for NVFP4 quantization that replays the GPTQ updates caused by each block scale and scores block error after compensation from unquantized columns. It also presents an execution infrastructure that keeps only the active layer resident, tiers activations across device, host, and disk, and distributes output rows across tensor-parallel ranks. Together they reach 99.35% and 100.84% question-weighted recovery from BF16 across seven benchmarks on Qwen3.5-397B-A17B and Llama-3.3-70B-Instruct.
In briefAdaptive Update-Step Scheduling for Recursive Reasoning in Looped Transformers
SummaryThis paper studies how the step size of each iterative update should vary in recurrent reasoning models that refine latent states with shared parameters. The authors decompose the sensitivity of terminal loss to update scale into persistent-progress and centered-fluctuation terms, and use this to propose TAPS, a scheduler that adapts step size online. According to the paper, TAPS improves terminal accuracy on structured reasoning tasks without retraining, and incorporating the principle into training yields further gains and up to 1.56x wall-clock speedup at matched baseline accuracy.
