HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 545AI / GeneralarXiv:2607.01224

    AutoMem: Automated Learning of Memory as a Cognitive Skill

    • The file system, including read, write, search, append, and create, is promoted to a first-class memory action equivalent to task actions, so every memory decision is traceable.
    • In Outer-loop 1, a meta-LLM reviews trajectories of up to 100,000 steps and automatically improves the agent scaffold, including the code, prompts, and memory schema.
    • In Outer-loop 2, good memory decisions from the agent are extracted and used to LoRA fine-tune a dedicated memory specialist model, while the task model remains frozen.
    • Memory optimization alone yields a 2x to 4x performance gain without changing task-model weights, showing that memory has more leverage than model scale.
    • The optimized 32B agent surpasses 72B and reaches the level of Claude Opus 4.5 and Gemini 3.1 Pro Thinking.
    • This is the first case that turns the cognitive-science concept of metamemory into a concrete optimization target for LLM agents.
    AbstractPDF
  2. 544RoboticsarXiv:2607.00272

    ASPIRE: Agentic /Skills Discovery for Robotics

    • The closed-loop robot execution engine exposes multimodal traces for each primitive, including perception overlays, grasp candidates, trajectories, and collision feedback, enabling autonomous failure diagnosis and repair by the agent.
    • It accumulates validated repair patterns in a skill library and reuses them as in-context guidance on later tasks, improving adaptation speed as the task set grows.
    • Evolutionary search goes beyond single-trajectory repair by exploring multiple candidate programs in parallel, which further improves success rates.
    • It achieves a 31 percent zero-shot success rate on LIBERO-Pro Long, compared with 4 percent for the previous method, demonstrating the generalization effect of the growing skill library.
    • When sim skills are transferred to a real dual-arm robot, inference tokens are reduced by about 10x and success rates also improve.
    • Performance improves step by step from a macro average of 14 percent without the execution engine to 62 percent with the engine and 72 percent with evolutionary search.
    AbstractPDF
  3. 543Computer VisionarXiv:2607.00784

    LeVLJEPA: End-to-End Vision-Language Pretraining Without Negatives

    • This is the first proof that stable large-scale vision-language pretraining is possible with only cross-modal prediction and SIGReg regularization, without negative pairs.
    • When used as a VLM backbone, it outperforms both CLIP and SigLIP, with GQA at 44.6%, VQAv2 at 54.2%, and POPE at 66.9% on the Llama-1N setup.
    • For semantic segmentation, it provides better dense semantic features than contrastive learning, preserving structured information at the patch-token level.
    • In zero-shot classification, contrastive learning still leads, but as a backbone the zero-shot score is actually inversely correlated with performance.
    • The entire training recipe removes the need for negatives, temperature, momentum, and teacher-student mechanisms, simplifying training and reducing cost.
    • It is well suited to object-centric representation learning and is more robust to backgrounds than contrastive methods.
    AbstractPDF
  4. 542Machine LearningarXiv:2607.01232

    Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training

    • Training just a single transformer layer can recover up to 114% of the full RL gain, while the weakest layers contribute less than 30%.
    • High-contribution layers consistently concentrate in the middle 40% to 60% of the network depth, regardless of model scale, family, RL algorithm, or dataset.
    • The layer contribution ranking is stable, with Spearman rho equal to 0.76 across datasets and 0.59 across tasks.
    • Raising the learning rate or selectively training the high-contribution layers adds further gains over full RL, including +43% for Qwen3-1.7B, +27% for 4B, and +32% for 8B.
    • Models trained on different layers show complementary behaviors, and majority voting can improve performance further.
    • Although the weight change magnitude is relatively uniform across layers during full training, the contribution is extremely uneven, which shows that parameter subspace efficiency differs sharply.
    AbstractPDF
  5. 541AI / GeneralarXiv:2606.31651

    FARS: A Fully Automated Research System Deployed at Scale

    • The system uses a fully automated four-stage pipeline: Ideation, Planning, Experiment, and Writing, with a specialized agent at each stage and a shared workspace.
    • It generated 166 papers across 67 fine-grained AI and ML topics and preserved the full output, including successful, failed, and negative results, rather than only selected samples.
    • It collected 282 structured volunteer reviews covering 140 papers, including ratings, soundness, presentation, contribution, integrity checks, and disclosure of LLM use.
    • In a cross-comparison with the Stanford Agentic Reviewer, FARS scores 5.00 out of 10 across 165 papers, which surpasses existing systems such as DeepScientist at 4.13 and ScientistOne at 4.00.
    • Human reviews average 3.23 out of 10, so they are stricter than automated reviewers but still aligned in direction, and 11.4 percent of papers reach the accept threshold under automated review criteria.
    • Repeated failure modes include narrow experimental scope, methodological limitations, and integrity issues such as claim-artifact mismatch, which makes contribution quality, experimental sufficiency, and faithful reporting the true bottlenecks rather than polished presentation.
    AbstractPDF
  6. 540Computer VisionarXiv:2606.31986

    CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts

    • The problem is that text CoT must generate thousands of tokens autoregressively, which causes large inference latency, and the natural-language form can erase alternative reasoning paths.
    • It uses three forms of step-level supervision: forward decoding from latent to text, backward decoding from text to latent, and internal supervision for consistency between latent steps.
    • At inference time it removes the decoder and internal supervision, so reasoning is completed with only 3 latent vectors and no extra parameter overhead.
    • Across 8 benchmarks on Qwen3-VL-8B, including MMStar, MathVista, and SeedBench, it consistently outperforms both text CoT and previous latent reasoning methods.
    • For speed, text CoT takes 7.24 seconds and 142 tokens, whereas CoLT takes 0.32 seconds with 3 latent vectors, which is a 22.6x decoding speedup and a 10.1x end-to-end speedup.
    • It is also robust to noise: under visual and textual noise, Direct Answer drops by 7.2 to 19.0 percent, Text CoT drops by 5.0 to 13.8 percent, but CoLT drops only 2.8 to 9.0 percent.
    AbstractPDF
  7. 539LLMs / NLParXiv:2606.30015

    Parametric Skills

    • The core problem is that in-context skills are limited by the model's ability to understand and follow instructions, and performance drops sharply, especially for open-source models or long contexts.
    • The solution is to feed a text skill into a hypernetwork that generates a LoRA adapter in a single forward pass, applying the skill without consuming context window.
    • It builds a 45.8K-scale high-quality skill library from web crawling and summaries of real agent trajectories, covering 13 domains.
    • On six SWE tasks, ParametricSkills scores 64.09 versus 57.65 for in-context learning and 48.48 for SHINE, averaged over LLM judges. SHINE fails to faithfully compress both the skill content and how to use it.
    • Rank-concatenation merging gives multi-skill composition performance equal to or better than a single LoRA, and it outperforms factor-wise linear merging.
    • In HumanEval-based self-evolving experiments, five rounds of iteration improve accuracy from 75% to 84.76%, and online continual merging further improves it from 29% to 51.6%.
    AbstractPDF
  8. 538Computer VisionarXiv:2607.00678

    ABot-M0.5: Unified Mobility-and-Manipulation World Action Model

    • It identifies three structural bottlenecks: temporal resolution mismatch between coarse video and fine-grained control, action-space mismatch between navigation and manipulation, and context mismatch between training and inference.
    • It introduces an intermediate latent action that bridges video latent representations and embodiment-specific control while capturing local visual state transitions.
    • A dual-level Mixture-of-Transformers separates modality representation from heterogeneous action subspaces, such as base movement and arm manipulation, so that each can be optimized separately.
    • Dream forcing trains dynamics on model-predicted dream video, which aligns autoregressive inference with training conditions and reduces long-horizon rollout error accumulation.
    • On RoboCasa365 mobile manipulation, it reaches 90.8 percent, 75.4 percent, and 58.4 percent across difficulty levels, and on real-robot peg insertion it achieves a 70 percent success rate and a 96 percent process score, compared with 50 percent and 90 percent for pi0.5 and 30 percent and 77 percent for Fast-WAM.
    • On real robot long-horizon tasks such as plate sorting, fruit sorting, and flower arrangement, it achieves 60 to 80 percent success, which is a large improvement over Fast-WAM at 20 to 40 percent.
    AbstractPDF
  9. 537Machine LearningarXiv:2606.32017

    TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning

    • A core problem in agent RL is that GRPO applies trajectory success or failure to every action equally, which punishes useful exploration in failed rollouts and rewards regression in successful rollouts.
    • It introduces a four-role taxonomy, Decisive, Exploration, No-progress, and Regression, and applies differentiated credit rules by role.
    • It proves that role labels alone can support MSE-optimal segment correction and always reduce advantage estimation error within the judge's trusted range.
    • On Qwen3-1.7B, it improves ALFWorld by 18.4 points and WebShop by 7.9 points, while reducing environment turns per completed rollout by 10.4% to 14.8%.
    • Regression suppression is the main driver of the gains, with exploration bonuses adding only 0.6 to 1.7 points as a stable auxiliary benefit.
    • It outperforms scalar process rewards and outcome-supervised value baselines, so the key benefit comes from role typing itself rather than simply adding dense rewards.
    AbstractPDF
  10. 536Information RetrievalarXiv:2606.31984

    GR2 Technical Report

    • Semantic ID mid-training with at least 99% uniqueness lets the LLM directly reason about catalog items.
    • Trace distillation from a strong 32B teacher model installs reasoning priors in the 8B student model.
    • DAPO-based RL with multiple rewards, format, AUC/NDCG, and LLM-as-judge, optimizes re-ranking.
    • It improves industrial traffic by +18.7% in R@1, +7.1% in R@3, and +9.6% in N@3.
    • On-Policy Distillation (OPD) recovers about 82% of the 32B teacher gains with a 1.7B model that is only 5% of the size.
    • Conditional reward design to prevent reward hacking, such as preserving input order and exploiting position bias, is essential.
    AbstractPDF
  11. 535Software EngineeringarXiv:2606.29957

    SWE-Together: Evaluating Coding Agents in Interactive User Sessions

    • It extracts 109 verifiable tasks from 11,260 real sessions, for a conversion rate of 0.97 percent.
    • The state-conditioned LLM user simulator preserves the original intent while adapting feedback to the trajectory.
    • It evaluates along multiple dimensions using final accuracy, user correction, meaning the number of correction turns, and intent coverage.
    • Stronger agents achieve higher success rates with fewer interventions, suggesting an improved user experience.
    • It highlights the limits of static SWE-Bench style evaluation, which misses real interaction by scoring only single-turn commands and final code.
    • It uses four upstream data sources: DataClaw, Pi-staging, Hyperswitch, and SWE-chat.
    AbstractPDF
  12. 534Computer VisionarXiv:2606.29148

    GPC: Large-Scale Generative Pretraining for Transferable Motor Control

    • It learns a discrete motion representation based on Finite Scalar Quantization (FSQ), which avoids the codebook problems of VQ-VAE.
    • It models motion tokens with a GPT-style autoregressive transformer using next-token prediction.
    • It is trained on a dataset with more than 600 hours of motion data and achieves a 99.98% tracking success rate.
    • It shows emergent behaviors such as disturbance recovery and fall recovery strategies.
    • It supports adaptation to new tasks through Parameter-Efficient Fine-Tuning (PEFT).
    • It is a SIGGRAPH 2026 paper from NVIDIA and Simon Fraser University.
    AbstractPDF
  13. 533AI / GeneralarXiv:2606.30639

    Self-Evolving World Models for LLM Agent Planning

    • We present a self-evolving world model that changes only deployment-time context without parameter updates.
    • Episodic memory uses retrieval-based simulation of real action transitions.
    • Semantic memory extracts persistent heuristic rules from prediction-observation mismatches.
    • Selective foresight filters low-confidence predictions before integrating them into the agent's reasoning context.
    • It achieves the best predictive accuracy across all three backbones on ALFWorld and ScienceWorld.
    • Oracle experiments show that noisy foresight harms action accuracy, proving the importance of confidence filtering.
    AbstractPDF
  14. 532Information RetrievalarXiv:2606.31693

    ShopX: A Foundation Model for Intent-to-Item Fulfillment in Agentic Shopping

    • It unifies intent-to-item fulfillment in a single model: intent understanding, execution planning, and SID-native control.
    • Semantic ID (SID) design: semantically recoverable and editable by the LLM.
    • It specializes a general LLM into a shopping agent through continued pretraining and post-training recipes.
    • On single-turn and multi-turn evaluations based on Taobao production logs, it outperforms existing tool-mediated systems.
    • It achieves especially large gains on complex and ambiguous requests, reaching precision of 88% to 97%.
    • The model-native design reduces loss during handoff between orchestration and item execution.
    AbstractPDF
  15. 531AI / GeneralarXiv:2606.29537

    OSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks

    • It contains 108 long-horizon workflows, with a human median of 1.6 hours and an average of 318 tool calls per agent.
    • Claude Opus 4.8 with max thinking completes 20.6 percent of tasks, while GPT-5.5 completes about 13 percent, although its token efficiency is better.
    • It defines 10 challenge phenomena, including dynamic environments, cross-source reasoning, hidden-state reasoning, and conflict resolution.
    • It builds a realistic environment from 31 self-hosted web services, including email, banking, and chat.
    • The main failure causes are forgetting constraints, missing intermediate information, guessing, and skipping verification, not basic GUI manipulation.
    • Agents spend less than 7 percent of their own budget on error detection and correction.
    AbstractPDF
  16. 530Machine LearningarXiv:2606.32026

    AdaJEPA: An Adaptive Latent World Model

    • The first test-time adaptation framework for a JEPA world model, using a Plan-Execute-Adapt-Replan loop.
    • It performs self-supervised adaptation from observed transitions alone, without additional expert demonstrations.
    • It achieves large gains with only one gradient step per MPC replanning step.
    • It produces significant improvements over the fixed model on both in-distribution and out-of-distribution target-reaching tasks.
    • It shows especially large gains when training data is limited.
    • It compares online buffer management strategies using recent-N and hard-N approaches.
    AbstractPDF
  17. 529AI / GeneralarXiv:2606.31410

    Xiaomi-GUI-0 Technical Report

    • A hybrid, real-device-centered infrastructure uses physical devices as the main environment and a sandbox as a secondary one to match the real deployment distribution.
    • An error-driven data flywheel turns failed rollout trajectories into corrective actions, reflection explanations, and recovery demonstrations.
    • A three-stage progressive curriculum: SFT for basic operations, Step RL for long-term planning, and Agentic RL for error recovery.
    • It achieves a 72.0% success rate on RealMobile and 78.9% on AndroidWorld.
    • It strengthens the ability to handle abnormal states in real applications, including account states, permission dialogs, payment authentication, and risk controls.
    • It builds a closed-loop real-device environment to solve emulator-detection mechanisms.
    AbstractPDF
  18. 528LLMs / NLParXiv:2606.30406

    MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training

    • It uses a three-stage pipeline: General SFT, Domain-specialised RL, and Multi-teacher On-Policy Distillation.
    • It aligns teacher and student on student rollouts with per-token reverse KL, which gives dense supervision and remains on-policy.
    • On Qwen3-30B-A3B, it improves by 5.5 points over the next best baseline and closes 91% to 95% of the teacher-student gap.
    • Domain teacher training can run in parallel, which increases development throughput and decouples the domains.
    • Teachers from the same origin are key to stable optimization, while using external teachers risks divergence.
    • It is deployed in the MiMo-V2-Flash industrial model, which validates its utility at frontier scale.
    AbstractPDF
  19. 527AI / GeneralarXiv:2606.28446

    Domain-Informed Multi-View Self-Distillation for Astronomical Light-Curve Representation Learning with JEPA

    • C-RoPE, or Continuous Rotary Positional Embedding, uses rotary position encoding to handle irregular observation intervals.
    • EANE, or Error-Aware Numeric Embedding, integrates measurement uncertainty or noise into the embedding.
    • It performs multi-view self-distillation across three domain views: raw light curves, GLS periodograms, and phase-folded curves.
    • On the StarEmbed benchmark, it beats hand-crafted features on 15 of 16 metrics, with macro-F1 ranging from 42.56 to 63.58.
    • It is validated on downstream applications such as similarity search, zero-point drift detection, and parameter estimation.
    • It reaches state of the art on 5 of the 12 non-astronomy datasets in PYRREGULAR, which highlights the importance of domain-specific design.
    AbstractPDF
  20. 526AI / GeneralarXiv:2606.30626

    DOPD: Dual On-policy Distillation

    • It defines the privilege illusion as a failure pattern in which information asymmetry from privileged information is mistaken for ability gaps.
    • Advantage-aware routing dynamically selects the supervision source, teacher or student, for each token.
    • It improves average performance by +7.5 points over Vanilla OPD on LLMs and +6.0 points on VLMs.
    • It applies strong-teacher distillation over the full vocabulary and weak-student self-supervision by token category.
    • It shows consistent gains in continual learning, OOD generalization, and training stability.
    • Token-level analysis shows that high-teacher, low-student tokens are responsible for transferring core abilities.
    AbstractPDF
  21. 525RoboticsarXiv:2606.28276

    SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation

    • It automatically generates a sim-ready digital twin from a single video, with F1 scores from 0.81 to 0.92 and 0.93 to 0.99 after 3 minutes of tuning.
    • It generates scene variations that preserve affordances while increasing diversity across object, scene, and task cousins.
    • Across 7 tasks and 5 policies, simulation evaluation correlates with real performance at Pearson 0.911, which is 0.59 higher than the previous SOTA.
    • Training policies on cousin data improves zero-shot sim-to-real transfer success by 17 to 40 percent.
    • It shows strong zero-shot transfer on YAM, with 99 percent Pot on Stove success, and on DROID, with 100 percent Stack Dishware success.
    • It handles more complex tasks than prior work, including multi-step manipulation, articulated-object interaction, and bimanual collaboration.
    AbstractPDF
  22. 524Information RetrievalarXiv:2606.29718

    Diagnosing and Mitigating Context Rot in Long-horizon Search

    • Context rot is defined as the phenomenon in which accumulated context causes the model to give up or produce an uncertain answer.
    • It performs a systematic diagnosis using four open-source models, including Qwen3.5-397B and GLM-4.7, across three benchmarks.
    • It evaluates seven management methods, including context compaction, trimming, and isolation, in terms of performance, cost, and their effect on rot.
    • Rot-aware filtering, which removes uncertain or give-up trajectories, improves average performance by 2.6 to 4.9 percent.
    • The combination of compaction and trimming offers the best balance between cost and rot mitigation.
    • Context isolation through subagents is highly model dependent and only works well on stronger models.
    AbstractPDF
  23. 523Machine LearningarXiv:2606.29723

    ScaleAware-JEPA: Latent Representation for Discovery in Multiscale Physical Fields

    • Constrained Diffusion Decomposition, or CDD, decomposes fields into pixel-level scale components.
    • The CDD scale determines the masking geometry, aligning the physical scale hierarchy with the prediction task.
    • It is applied to three heterogeneous domains: MHD turbulence, the NGC 3627 molecular gas field, and Chengdu night lights.
    • It can generate a dense latent atlas without labels, segmentation, or catalogs.
    • It quantitatively verifies the superiority of scale-aware masking over fixed patch masks.
    • It is an independent study performed on consumer hardware, specifically an Apple M3 Pro and an RTX 3090.
    AbstractPDF
  24. 522Machine LearningarXiv:2606.28277

    Towards Automating Scientific Review with Google's Paper Assistant Tool

    • It uses a four-stage agent pipeline of Segmenter, Adaptive Budgeting, Deep Review, and Global Synthesis.
    • It improves recall by 34 percent over zero-shot on math error detection in the SPOT benchmark.
    • It was piloted as a pre-submission tool at two major conferences, STOC and ICML.
    • It proposes a four-level taxonomy for AI roles: Tool, Supporting Reviewer, and Fully Automated.
    • It addresses the precision drop and context limits of Pass@k through segment-level deep review.
    • A NeurIPS 2021 study found that even human reviewers had a 23 percent disagreement rate, which suggests practical value for AI review.
    AbstractPDF
  25. 521LLMs / NLParXiv:2606.30616

    Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent

    • A 35B MoE shows a performance path based on expanding agent horizon rather than parameter count.
    • Knowledge-Action Graph (KAG): a long-horizon trajectory infrastructure that stores evidence, actions, observations, and verification results as connected objects.
    • A three-stage training pipeline: all-domain SFT, domain-specific teacher RL, and domain-routed multi-teacher on-policy distillation.
    • It generates agent trajectories averaging 45K tokens and integrates six heterogeneous domains.
    • It outperforms 1T models on SEAL-0 (56.4), IFBench (80.6), and FrontierScience-Olympiad (79.0).
    • Salient Vocabulary Alignment reduces conflicts between reasoning patterns across domains.
    AbstractPDF
  26. 520Computer VisionarXiv:2606.30534

    Orca: The World is in Your Mind

    • It proposes Next-State Prediction as a unified paradigm instead of Next-Token, Next-Frame, or Next-Action Prediction.
    • It designs two complementary paradigms, unconscious learning through dense state transitions in continuous video and conscious learning through language-conditioned sparse event transitions.
    • It builds a large-scale world-learning dataset with 125,000 hours of video and 160 million event annotations.
    • It freezes the encoder and trains only a lightweight decoder to produce text, image, and action readouts.
    • It validates scalability at the 4B and 0.8B scales, showing that larger world latents translate directly into stronger downstream performance.
    • It also openly states clear limitations, including the current restriction to vision and language modalities, short-horizon transitions, and limited model scale.
    AbstractPDF
  27. 519RoboticsarXiv:2606.27374

    World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays

    • The first continual imitation learning approach to use WAM generative capability, reconstructed future frames, as pseudo-replay.
    • Compared with sequential fine-tuning, it reduces catastrophic forgetting by up to 50% and comes close to experience replay that uses real demonstrations.
    • Recursive generation feeds future observations predicted by the WAM back into the model to synthesize complete trajectories, which is recurrent generative replay.
    • It remains effective on real robots as well: NBT improves from 96.3 to 60.5, reducing forgetting by 40%, and FWT improves from 50 to 80.
    • It identifies two main bottlenecks: visual quality degradation during long-horizon recursive generation, reflected in lower PSNR, and mismatch between imagined and grounded actions, 83% versus 42%.
    • Action representation drift shows that REGEN, at 0.12, preserves representations much better than Seq-FT, at 0.30.
    AbstractPDF
  28. 518RoboticsarXiv:2606.28133

    Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots

    • The paper argues that rotation information from human hands is noisy and that rotation-inclusive 6DoF learning is suboptimal because the human hand and parallel gripper have different contact patterns.
    • It adopts relative wrist translation in the head-camera frame as a bridging action and builds a shared action space between humans and robots.
    • An interleaved action token sequence, 3D-wrist to 6D-eef to gripper, together with attention masking, handles missing components in heterogeneous data.
    • The training strategy has three stages: large-scale human-data pretraining, joint human-robot training, and a small amount of real-robot fine-tuning.
    • Adding bridging-action supervision to robot data is essential, raising overall success from 12.5 percent to 38.3 percent.
    • The upper-bound analysis shows that removing the embodiment gap lets the bridging objective achieve even stronger transfer performance, with overall progress rising from 59.75 percent to 73.54 percent.
    AbstractPDF
  29. 517RoboticsarXiv:2606.27581

    SceneBot: Contact-Prompted General Humanoid Whole Body Tracking with Scene-Interaction

    • Using reference motion plus per-link contact labels as policy conditioning, we unify free-space and contact-rich behaviors in a single policy for the first time.
    • Hindsight scene reconstruction infers the scene-interaction graph from retargeted human motion and reconstructs terrain and object assets in reverse.
    • With 7.5 hours of contact-rich training data, combining AMASS, OMOMO, Bones, and Lafan, it generalizes to unseen motions and environments.
    • Global root tracking is essential: without it, terrain success drops from 100% to 45% and object-task success from 95% to 20%.
    • Removing contact labels makes object-grasp success converge to 0%, proving the decisive role of contact labels.
    • On the real Unitree G1, state-estimation performance is confirmed with an average root-position error of 0.032 m and a rotation error of 0.018 rad.
    AbstractPDF
  30. 516AI / GeneralarXiv:2606.26859

    AgentX: Towards Agent-Driven Self-Iteration of Industrial Recommender Systems

    • The four-stage closed loop is Brainstorm, which proposes evidence-based ideas; Developing, which generates repo-grounded code; Evaluation, which performs guardrail-vetoed A/B decisions; and Harness Evolution, which applies SGPO.
    • SGPO, or Semantic-Gradient-based Prompt Optimization, uses execution trajectories to update agent prompts and doubles weekly throughput.
    • Three AgentX workers turned 374 ideas into 10 launchable rollouts over 3 weeks, which is 8x the concurrency of human engineers and 3.7x the business value.
    • It uses online A/B feedback as a reward signal so that the work stays aligned with real business value, including a 0.561 percent increase in app dwell time and more than 100 million yuan in annual value.
    • It shows that the approach can extend beyond recommendation strategy to model-side research, such as paper replication, module pruning, and architecture combination.
    • It includes a negative-result assetization mechanism that turns failed experiments into structured knowledge assets.
    AbstractPDF
  31. 515Computer VisionarXiv:2606.28128

    PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation

    • We experimentally identify two sources of physical errors: deformation of moving objects and unnatural spatiotemporal correlations between interacting entities.
    • We design complementary pixel-level trajectory alignment loss based on point tracking and semantic-level relation alignment loss based on a frozen video encoder.
    • We propose a physics region focus strategy that concentrates physical supervision on interaction-critical regions, outperforming uniform application by 1.5 points.
    • On R-Bench, we improve Wan2.2-I2V-A14B by 22.3% and Cosmos3-Nano by 9.2%, corresponding to gains of 7.1% and 3.7% over vanilla finetuning, respectively.
    • In the WorldArena action-planner protocol, closed-loop success rises from 16.0% to 24.0%, and average downstream policy success rises from 68.2% to 72.8%.
    • PF-Cosmos achieves the best overall score on R-Bench, PAI-Bench, and EZS-Bench.
    AbstractPDF
  32. 514Computer VisionarXiv:2606.27608

    Qwen-Image-2.0-RL Technical Report

    • It fine-tunes a VLM with pointwise scoring and chain-of-thought reasoning to build task-specific composite reward models for T2I, including alignment, aesthetics, and portrait quality, and for editing, including instruction following and face ID.
    • It introduces a hybrid CFG strategy in the GRPO-based RL framework to preserve pretrained knowledge and training stability at the same time.
    • On-policy distillation, or OPD, integrates the T2I teacher and the editing teacher into a single student model through trajectory-level velocity matching.
    • OPD achieves better performance than mixed RL, which optimizes multiple rewards at once, because it avoids cross-task interference.
    • It reports consistent improvements on Qwen-Image-Bench with 57.84 points, up 2.61, on the T2I arena with an Elo gain of 78, and on the editing arena with an Elo gain of 93.
    AbstractPDF
  33. 513RoboticsarXiv:2606.26025

    In-Context World Modeling for Robotic Control

    • We formulate the problem of restoring missing system configuration psi, such as camera viewpoint and morphology, as context in the VLA setting.
    • We prepend N task-independent random-search clips, typically 3 to 5, as context to implicitly infer system dynamics without parameter updates.
    • Across six OOD viewpoints, the average success rate is 25.0%, a gain of 5.2 percentage points over Multi-View's 19.8%, and wrong context can reduce performance to 18.9%, proving that the model truly depends on context.
    • With 20 to 80 mm spacers attached to the robot gripper, a morphology change is preserved with a 60% margin, with ICWM at 14.4% versus 5.6% for Multi-View.
    • Reusing context hidden states with KV caching restores inference latency to the 0.112 s baseline.
    • This context adaptation ability does not emerge naturally in standard imitation learning and requires an explicit training incentive.
    AbstractPDF
  34. 512Machine LearningarXiv:2606.26294

    The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators

    • Controlled utility evolution, with a fixed criterion within each epoch and replacement only at epoch boundaries, preserves self-improvement guarantees even under nonstationary utility.
    • On coding tasks such as Polyglot, adding agent-as-a-judge code review signals improves pass rate by 1.8 percentage points over the previous SOTA while using 1.35x to 1.72x fewer tokens.
    • In the paper-writing setting, the co-evolved writer's acceptance rate rises from 21.8 percent to 40.5 percent, which is a 1.86x improvement.
    • In IMO-level proof grading, the co-evolved grader reaches higher accuracy with three times less search cost than the fixed baseline.
    • By injecting adversarial goals at epoch boundaries, it corrects reviewer self-preference bias toward AI-generated papers, where AI papers are accepted 1.91 times too often.
    • As the evaluator gets stronger, it also creates a curriculum effect for the task agent, which leads to archive reranking and gradual strengthening.
    AbstractPDF
  35. 511AI / GeneralarXiv:2606.26300

    The Verification Horizon: No Silver Bullet for Coding Agent Rewards

    • When quality judging and trajectory monitoring are added to unit-test rewards, the SWE-Bench hacking solve rate drops sharply from 28.57% to 0.56%, while the normal solve rate rises from 40.22% to 60.53%.
    • For frontend tasks, introducing a browser-based interaction evaluator defends against length attacks that are vulnerable to static code inspection.
    • User-feedback rewards achieve up to a 13.3-point improvement across five internal coding-agent benchmarks.
    • For autonomous-evaluator data filtered for long-horizon tasks, RFT gains +1.91 points over random sampling, from 21.61 to 23.52.
    • There is no single mechanism that satisfies all three dimensions of verification, scalability, fidelity, and robustness, so task-specific design is essential.
    • Verification is not auxiliary but core training infrastructure, and it must co-evolve continuously with policy capability growth.
    AbstractPDF
  36. 510AI / GeneralarXiv:2606.27226

    Ask, Don't Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement

    • It uses a meta-prompt structure that breaks evaluation criteria into atomic binary questions, which greatly improves transparency and diagnostic value.
    • On QAGS for factual consistency, it improves Spearman rho by 0.195 over UniEval, which shows that the decomposition itself contributes most of the information gain.
    • BINEVAL catches subtle factual errors that G-Eval and UniEval give a perfect 5.0 to by asking 7 questions each, reaching 1.57 points and coming close to human judgment at 2.0.
    • It avoids the usual ceiling effect of LLM judges and distinguishes borderline outputs from clearly defective ones more effectively.
    • It improves both summarization and IFBench prompts through cross-model prompt updates that use question-level feedback.
    • Its task-agnostic and training-free design makes it immediately applicable to summarization, dialogue, instruction following, and other tasks.
    AbstractPDF
  37. 509Machine LearningarXiv:2606.25086

    Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models

    • Formulate iterate-average optimization as a continuous-time stochastic quadratic optimal control problem.
    • PACE is a lightweight wrapper on top of AdamW, with extra memory equal to one copy of the model weights.
    • The standard SCO convergence rate is guaranteed in the convex setting, and arbitrarily large bounded errors can be improved in the quadratic setting.
    • It shows consistent gains in three fine-tuning settings on SmolLM2, Qwen3, and Gemma3, and also works for GPT-2 FineWeb pretraining.
    • It is competitive with Schedule-Free and removes the need for LR decay, outperforming WSD across all token budgets with constant LR.
    • Extensive ablations on pullback strength c, EMA power κ, and update frequency show its robustness.
    AbstractPDF
  38. 508RoboticsarXiv:2606.27295

    LA4VLA: Learning to Act without Seeing via Language-Action Pretraining

    • It provides a systematic analysis of vision-language asymmetry in VLA training at both the data level and the input level.
    • It automatically builds 33K LA episodes from existing demonstrations without collecting extra data, in LA4-33K.
    • LA-only pretraining consistently outperforms VLA-only pretraining at the same scale.
    • Mixed LA-VLA pretraining reaches 87.53% on MetaWorld, 96.28% on LIBERO, and 83.3% on real robots.
    • Under visual noise, the LA-pretrained policy achieves a success rate that is 25 percentage points higher than VLA-only.
    • t-SNE analysis of the LA-pretrained policy shows clearly separated representations by command direction.
    AbstractPDF
  39. 507Machine LearningarXiv:2606.27326

    Hallucination in World Models is Predictable and Preventable

    • It defines three types of hallucination: perceptual hallucination from the tokenizer, action hallucination from marginalization in the dynamics, and scene hallucination from rollout divergence.
    • MMBench2 contains 65,600 trajectories spanning 427 hours, 210 tasks, and 10 domains, together with ground-truth actions, rewards, and the simulator.
    • Coverage-aware training improves both the tokenizer and dynamics, which raises rollout PSNR by 0.88 dB.
    • Collecting only 50 curiosity-driven trajectories yields 0.325 adaptation performance on unseen tasks, which is about 90 percent of the expert or human score of 0.362.
    • The hallucination predictors show a strong negative Spearman correlation of about 0.80 with rollout ΔPSNR.
    • The analysis also compares an off-the-shelf tokenizer, Wan 2.1 VAE, with an in-domain tokenizer.
    AbstractPDF
  40. 506Machine LearningarXiv:2606.26217

    Fast LeWorldModel

    • Action-prefix prediction generates multi-horizon latents in parallel, removing the bottleneck of sequential rollouts.
    • The dynamics module time drops from 31.4 seconds to 8.0 seconds, a 3.9x speedup, and overall CEM solve time falls by 48%.
    • Both the initial open-loop latent loss and its growth slope are significantly lower than in LeWM.
    • Average planning success rate rises from 85.8% to 90.5%, including 98% on Two-Room, 96% on Push-T, and 80% on Cube.
    • An ablation confirms that dense prefix supervision is essential for performance, not just terminal-only prediction.
    • In physical-state probing, an MLP probe achieves the best accuracy on agent position, block position, and angle.
    AbstractPDF
  41. 505LLMs / NLParXiv:2606.26790

    OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning

    • Because it directly extracts hindsight skills from completed on-policy trajectories, no external skill library or retrieval is needed.
    • It introduces a two-level hierarchy with episode-level skills for global workflows and failure-avoidance rules, and step-level skills for local decisions.
    • Critical-first routing uses step-level skills at important moments and falls back to episode-level skills otherwise.
    • It achieves 58.9 percent on ALFWorld, compared with 46.1 percent for GRPO, and 74.2 percent on WebShop, compared with 63.3 percent for GRPO.
    • It provides a qualitative analysis showing that skill injection fixes the hallucinated-target error in GRPO-trained agents, where the agent searches for objects that do not exist.
    • It is a training-time-only framework that needs no analyzer, external skill retrieval, or privileged context at inference time.
    AbstractPDF
  42. 504AI / GeneralarXiv:2606.24937

    The Hitchhiker's Guide to Agentic AI: From Foundations to Systems

    • It covers the foundations from LLM architecture, including tokenization, transformers, attention, and GPU systems, all the way to Flash Attention.
    • It covers alignment techniques such as RLHF, PPO, DPO, and GRPO, and reasoning techniques such as Chain-of-Thought and test-time scaling.
    • For agents, it covers trajectory RL, RAG and Agentic RAG, in-context, external, episodic, and semantic memory systems, and agent harness design.
    • It systematically organizes MCP, or Model Context Protocol, A2A, or Agent-to-Agent, protocols, along with tool use and agent skills.
    • It covers centralized, distributed, and hierarchical multi-agent architectures and communication patterns among agents.
    • It covers the full practical stack, from evaluation methods and agent UI design to production deployment, across more than 600 pages and 400 references.
    AbstractPDF
  43. 503Computer VisionarXiv:2606.26907

    Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation

    • It defines the root cause of real-world T2I failures as the Context Gap and proposes an agent paradigm to address it.
    • It structures Context-Aware Planning into three stages, information-level, content-level, and generation-level, for identifying missing context, collecting it, and allocating it.
    • Context Grounding collects context from four channels: reasoning, web and image search, memory, and feedback.
    • It builds IA-Bench with four capabilities, Plan, Reason, Search, and Memory, 17 tasks, 730 test instances, and 1,801 checklist items.
    • The intelligence of the MLLM backbone is decisive for the whole system, and swapping the backbone causes large drops on most metrics.
    • It finds that overusing image search can hurt generation quality, which shows that the search boundary must match the generator's ability.
    AbstractPDF
  44. 502Machine LearningarXiv:2606.25971

    Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors

    • It mathematically analyzes interference between weight magnitude and direction: direction updates are inversely proportional to the current magnitude, and magnitude grows as a byproduct of direction learning.
    • It decomposes weights inside the optimizer into fixed-norm directions plus learnable gains, so it can be applied without changing the model architecture.
    • Weight decay and warmup become unnecessary, greatly simplifying large-scale training recipes.
    • The optimal learning rate transfers across model widths, so no LR retuning is needed, similar to µP but obtained directly.
    • On large MoE models, MuonMD reaches the same loss as AdamW with about half the compute.
    • Fixing the direction on the sphere captures most of the improvement, and learnable gains provide additional accuracy gains.
    AbstractPDF
  45. 501LLMs / NLParXiv:2606.25331

    Improved Large Language Diffusion Models

    • It is the first case of scaling a masked diffusion language model to 12T tokens, trained from scratch with 8B parameters.
    • Grouped-query attention reduces KV-cache memory and tied embeddings slim the parameters down to 7.62B.
    • Confidence-based scoring improves PIQA by +1.3, ARC-C by +0.6, and HellaSwag by +2.3 over likelihood-based scoring on multiple-choice evaluation.
    • Extending SFT to 12 epochs yields continuous gains on GSM8K, MATH, and MMLU-Pro, confirming data reuse effects in diffusion LLMs.
    • The base model reaches MMLU 74.8, BBH 71.3, and GSM8K 81.9, surpassing Qwen2.5 7B across multiple benchmarks.
    • The instruct model is competitive without RL alignment, with MMLU-Redux 76.4 and GSM8K 89.0, but it still trails Qwen2.5 Instruct on code and math.
    AbstractPDF
  46. 500Computer VisionarXiv:2606.27377

    DanceOPD: On-Policy Generative Field Distillation

    • The method formalizes the conflict between text-to-image, local editing, and global editing from a velocity-field perspective and resolves it with on-policy distillation.
    • It removes target-field ambiguity with hard routing that assigns exactly one capability field to each sample.
    • On GEditBench, it improves by 8.1 percent over the existing OPD baseline while preserving GenEval performance comparable to the T2I source.
    • For local-plus-global editing combinations, it improves by 16.1 percent over the best competing baseline and by 7.9 percent over the local-edit source.
    • When absorbing the realism field, it improves by 9.9 percent over off-policy distillation and reaches 85.3 percent of the student-teacher reward gap.
    • It also shows that classifier-free guidance can be absorbed under the same velocity-MSE objective.
    AbstractPDF
  47. 499LLMs / NLParXiv:2606.24775

    Are We Ready For An Agent-Native Memory System?

    • We propose an analytical framework that systematically decomposes agent memory into four modules: representation and storage, extraction, retrieval and routing, and maintenance.
    • We evaluate 12 memory systems plus two baselines on a unified testbed spanning five workloads and 11 datasets.
    • No single architecture is universally best. Hybrid systems are strong for conversational QA and graph-based systems are strong for factual recall, but they are weak at temporal reasoning.
    • Memory extraction should preserve context at write time. Broader retention, rather than overly selective extraction, is better for downstream reasoning.
    • Conservative merging outperforms delayed flushing or excessive summarization in long-term stability.
    • Highly structured systems pay tens to hundreds of times more in index-build time and query latency, but the accuracy gains do not scale proportionally.
    AbstractPDF
  49. 496Machine LearningarXiv:2606.24143

    AsyncOPD: How Stale Can On-Policy Distillation Be?

    • AsyncOPD systematically studies the stale-data problem in asynchronous On-Policy Distillation, or OPD.
    • It shows that the KL direction changes the stale-data problem, where teacher-weighted forward KL is robust and student-weighted reverse KL is fragile, and it shows that recomputing advantages at training time is the most effective fix for reverse-KL OPD.
    AbstractPDF