Large Web platforms operate large fleets of ranking models, each selecting features from a shared pool of tens of thousands. We present an end-to-end agentic system that optimizes features across the fleet under resource constraints shared across models and those set per model, such as feature pipeline cost and serving throughput (QPS), respectively. In this system, LLM agents propose hypotheses, and deterministic tools execute experiments and verify them. For each model, an LLM generates and shortlists cross features from a combinatorial candidate space, and all features are then ranked by several importance methods. Feature blending mixes these rankings into a single selection, using Pareto fronts across tasks for multi-task models. An LLM Researcher agent then tunes the feature selection in 10--25 asynchronous trials under resource caps, and a trimming step keeps the pipeline cost within budget. Across production models, the system improves normalized entropy by 0.31% on average within each model's budget, and the trimming step cuts pipeline cost by 2.3% with no regression beyond measurement noise. The optimized models were launched, and in online A/B tests they improved a primary online metric by 0.20%.
Precise grounding matters. It specifies which object is the target and where that object is, even in clutter and for tiny objects, and it has to be fast enough for closed-loop control. Yet vision-language-action (VLA) and world-action models (WAMs) take perception from general-purpose vision-language and video-generation backbones, which still fail in these settings. We introduce GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary. Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines. Against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, GroundingPI establishes a new state of the art, averaging 73.68%, above the larger GPT-6 Astra (71.54%). As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving. Across RoboTwin 2.0 and RoboCasa-GR1, it outperforms all evaluated backbones in all four out-of-distribution settings, with relative gains of up to 24.8% over the strongest baseline. On RoboCasa-GR1, GroundingPI trained with 50% of the demonstrations outperforms those baselines trained with 75%. Its mean open-loop L2 error on nuScenes is 0.296 m. We systematically analyze GroundingPI's pretraining in scale and data composition. Grounding quality and downstream manipulation improve with pretraining scale, with generalization tracking perceptual gains most closely. Analyzing the data recipe across these 11 perceptual capabilities shows dense grounding's substantial benefits for both, and OCR's potential as a catalyst for perceptual learning. These results support grounding as a perceptual foundation, and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.
World-action models jointly predict future observations and actions, yet reconstruction-oriented visual representations do not explicitly emphasize the spatial relationships and object-state changes needed for manipulation. We introduce Phoenix, which learns representations of scene structure for world-action modeling. Specifically, Phoenix Encoder enriches DINOv3's general visual semantics through dual-teacher distillation of geometric and spatiotemporal semantic knowledge. Next, CS-VAE causally aggregates and compresses Phoenix Encoder features into generative latents while preserving scene information and spatial layout. Finally, Phoenix jointly predicts these latents, complementary appearance latents, and robot actions, linking scene evolution with action generation. To study pretraining data efficiency, we curate approximately 39,000 hours of egocentric and robot manipulation data for representation learning and use the robot demonstration subset for WAM pretraining. Following benchmark-specific post-training, Phoenix achieves state-of-the-art performance across simulation and real-robot benchmarks spanning single-arm, bimanual, and dexterous manipulation. It demonstrates strong robustness to perturbations, achieving success rates of 94.8% on LIBERO-Plus and 72.4% on RoboTwin 2.0 C2R, alongside strong cross-embodiment transfer and real-world manipulation performance on Bench2Dex and RoboDojo-RealWorld, respectively. Pretraining-scale analysis shows larger performance gains from data scaling than a video-VAE-only baseline, indicating stronger scaling benefits and more effective data utilization. Moreover, Phoenix maintains high success rates when adapted with limited downstream demonstrations, demonstrating strong few-shot adaptation. Our code, pretraining data, and models are available on the project page to help the community reproduce our results.
Controllable real-world video generation models can condition on a rendered graybox proxy video, which fixes layout, motion and camera. Coding agents can build such videos by generating a 4D scene in Blender, yet no benchmark checks both these rendered proxy videos and the complex, multi-shot scenes behind them. We introduce VibePrevisBench, a benchmark for 4D scene generation: an agent turns a brief into an executable 4D scene and renders its proxy video. Its 242 tasks span four input kinds, including scripts, photos, continuations, and reference videos. Each delivery is scored along two evaluation dimensions. Rule-based Scene Check re-runs the code and compares the exported 4D scene with hidden ground truth facts in nine dimensions. VLM-based Screen Check asks a VLM judge whether facts that pass the rules are recognizable in the proxy video. We analyze the evaluation results and trajectories of 11 frontier coding agents and identify three key observations. First, multimodal understanding separates the agents, especially in understanding space and camera motion. Second, all agents seem to already master basic camera work and staging, and fail at challenging requests. Third, agents inspect their own renders but fail to correct their errors. Agents thus need better visual understanding of space and camera, explicit checks for challenging requests, and item-by-item verification of renders against the brief.
Real-world harness engineering is driven by explicit and implicit user requirements, which guide development and validation. However, existing harness evolution methods often overlook implicit requirements and interaction needs, while using a single evolution agent to modify the harness and benchmark scores to select changes. This process often produces narrow patches and leads to benchmark-specific local optima. We introduce Agentic Harness Engineering Beta (AHE-Beta), a requirement-centered loop that brings real-world engineering practices into harness evolution. The loop consists of three steps: (1) User-driven evaluation lets simulated users raise their requirements as issues while they work with the harness over multiple turns. (2) Team-based modification implements those requirements as generalizable changes rather than task-specific patches. (3) Requirement-driven selection merges a change only when it resolves a verified requirement and passes review. Empirically, on held-out tasks, AHE-Beta raises pass@1 from 21.9% to 33.3% while using 40.8% fewer tokens than Base Harness, and achieves higher pass@1 than Claude Code, Meta-Harness, and AHE. In a blind human A/B test, real users prefer AHE-Beta in 31 of 48 valid pairs and rate it higher on all five dimensions, matching the directional preference of the user simulator. The human A/B results show that AHE-Beta achieves its goal of aligning harness evolution with real-world user requirements. Without further evolution, the gains transfer to GLM-5.3 and DeepSeek-V4.1-Flash with pass@1 improvements of 15.8% and 5.3% over Base Harness. Together, these results show that requirement-centered harness evolution is a more general method for improving agent capabilities.
Existing AI agents can retrieve information, call tools, run experiments, and produce multi-step outputs. Yet most are built around relatively static workflows and struggle to preserve reliable state as the environment changes. Financial trading makes this limitation especially consequential: prices, news, market conditions, portfolio positions, and available capital evolve continuously. A result is useful only when it remains tied to the correct time, data snapshot, and execution context. We present Vibe-Trading, a comprehensive agent-native trading framework and all-in-one workspace for building, evaluating, and evolving agents for financial trading. It unifies market research, financial data, strategy generation, backtesting, portfolio analysis, paper trading, multi-agent collaboration, and optional broker execution. Its central contribution is an architecture that keeps results, evidence, research history, and trading authority connected throughout the research-to-action process. Each result retains its source, temporal scope, configuration, and producing run. Shared evidence grounding enables verification, versioned research records support reproducible continuation, and an explicit authority gate separates strategy research from live execution. Together, these designs provide a practical foundation for trading agents whose results can be traced, reproduced, and acted on only under explicit authority.
Robot manipulation videos provide a rich visual experience, yet recovering robot action and gripper-state annotations from RGB observations still remains challenging. Using the embodiment-agnostic tool-center-point (TCP) pose as the action space, we present TCP4D, a unified model in which a shared 4D scene representation both tracks observed TCP motion and generates future TCP trajectories. Given a monocular video and initial 2D TCP queries, the tracker directly recovers metric TCP poses in SE(3) and gripper states alongside scene geometry and camera parameters. The generator samples local features from the same 4D representation along the recovered TCP tracks, then combines them with robot observations and language instructions to predict future TCP states. Our new cross-embodiment augmentation pipeline converts existing RGB videos into demonstrations for alternative embodiments by retargeting the recovered motion and synthesizing the corresponding observations, without recorded source-robot states. Experiments show that the tracker achieves absolute trajectory errors of 7.60 and 16.07 mm on DROID and RoboLab and adapts to a new embodiment with only ten demonstrations. The generator achieves 69.19% mean success on RoboTwin and outperforms well-established policy models on two real-world tasks, while joint training on source and augmented demonstrations improves target-robot success by 41.5% over augmented-only training.
Photographic color grading shapes the tone and color of a photograph to achieve a desired look. Text-driven photographic look transfer lets users specify these adjustments in natural language, with the goal of reproducing the requested look across photographs while preserving their content. However, prior research on text-driven color grading uses different datasets, text instructions, target definitions, and metrics. These differences make it difficult to compare how faithfully the methods reproduce the same look. To compare methods under a shared protocol, we introduce LUTBench. We collect 91 three-dimensional lookup tables (3D LUTs) and define one photographic look per LUT with a specified input conversion. Applying these looks to 5,000 MIT-Adobe FiveK photographs produces 455,000 input-target image pairs with reproducible target images. For each look, we use templates to turn measured tone and color changes on fixed color samples into candidate clauses. We use training image measurements to verify applicable clauses before composing the text description. The evaluation measures target fidelity, tests sensitivity to predefined input conversions, and separates generalization to unseen looks from reproduction of seen looks on new photographs. Across 16 methods, stronger text alignment does not guarantee target fidelity. However, most supervised variants improve fidelity on both seen and unseen looks.
When shown matches between two photographs of the same cathedral, people can tell at a glance which ones are wrong: a window matched to its look-alike jumps against its neighbors, while correct matches shift together. Two-view correspondence learning, also called correspondence pruning, asks a network to make this judgment for thousands of candidate matches, most of which are wrong. Mamba, a fast sequence model, now drives strong pruners, but it reads the matches as a plain sequence, and whether its scan can capture such agreement is unclear. We propose SFMambaNet, which scans the matches cluster by cluster, so that correct matches vary slowly along the sequence while wrong ones stay erratic, and filters the scanned features with learnable frequency-domain gates between its forward and backward passes. On camera pose estimation, outlier removal and homography estimation, we find that SFMambaNet outperforms spatial-domain Mamba baselines with about two million parameters. Contrary to the usual expectation that such gains come from extra capacity, we find that a parameter-free low-pass filter beats point-wise layers as large as the gates. The learned gates turn low-pass on their own and make the scan far less sensitive to how the matches are shuffled. Together, these results show that frequency is a natural language for describing agreement between correspondences, with implications for other sequence models that must read unordered sets, such as point clouds and graphs.
Robot-use agents (RUAs) have emerged as a promising approach to general-purpose robot control. However, strong performance has so far been achieved only by frontier proprietary models such as GPT-6 Astra and Claude Fable 5.2, whereas open-weight models remain substantially less capable. We show that an open-weight 27B reasoning model can be trained to use real robots natively through reinforcement learning, entirely in simulation and without demonstrations. Like a computer-use agent (CUA), the model observes the environment, reasons in a chain of thought, and generates the next end-effector pose, without any harness. We build a Post-training Engine for Agentic Robotics (PEAR) and use it to train the model on simulated robot tasks at scale with robot-use reinforcement learning (RURL), mirroring how frontier models acquire computer use. We find that the skills acquired in simulation generalize well beyond the training tasks: they transfer to an unseen robot, to the real world, and to novel instructions. After training on LIBERO and RoboDojo in simulation, the model succeeds in 60% of our trials on a real dual-arm robot, whereas the base model succeeds in none. On LIBERO, training raises the success rate from 1.6% to 74.4%, approaching GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort. Compared with $π_{0.5}$, a vision-language-action model fine-tuned on demonstrations, our model is substantially more robust to changes in the task itself. On LIBERO-Pro, $π_{0.5}$ succeeds in only 0.75% of episodes when assigned a new task in a familiar scene whereas our model succeeds in 57.4%. Finally, the model retains its general capabilities, with no measurable degradation on nine reasoning, coding and vision benchmarks. These results highlight robot-use reinforcement learning as a promising component of foundation-model post-training.
Model architecture shapes the scaling behavior, training efficiency, and inference cost of LLMs, making it a central focus of both industry and academia. In industry, teams such as those behind DeepSeek, Qwen, and Kimi routinely introduce architectural innovations, yet public reports rarely document which ideas succeeded, which were abandoned, and why. In academia, limited compute and inconsistent baselines often restrict evaluations to small-scale, single-point comparisons, making it difficult to establish whether improvements hold at larger scales or translate into broader impact. ArchSpace addresses both gaps as a public platform for Recording, Reviewing, and Referencing Architecture Research: each proposal is recorded with its experiments and reasoning, reviewed by the community, and referenced by later work. To evaluate ideas before costly large-scale training, ArchSpace leverages scaling laws to predict model loss at larger scales from smaller experiments, a practice widely adopted in industry. We provide initial references that make the platform usable: Qwen3.5 and OLMo3 baselines, fitted scaling laws, and a reusable training recipe. Ideas validated through this review can in turn become new baselines, and ArchSpace thereby aims to broaden the range of ideas that shape the next generation of models.
While spatiotemporal deepfake detectors achieve high AUC, our experiments reveal their susceptibility to evasion attacks. These models tend to overfit on fragile temporal spectrum cues, rather than learning robust semantic causality. To mitigate this vulnerability, we propose SpInShield, a temporal spectral-invariant defense framework explicitly designed to decouple semantic motion from manipulatable spectral artifacts. We propose a learnable spectral adversary that dynamically synthesizes severe spectral deformations, simulating extreme attack scenarios. By employing a shortcut suppression optimization strategy, SpInShield compels the encoder to extract reliable forensic cues while purging unstable spectral statistics from the latent space. Experiments show that SpInShield obtains competitive performance on widely used datasets and outperforms the strongest baseline by 21.30 percentage points in AUC under simulated amplitude spectral attacks.
For robots to assist, serve, and collaborate with humans, they must not only interpret human requests conveyed through language and multimodal intentional cues, but also execute long-horizon tasks that interleave navigation and manipulation. In this paper, we introduce HEIR, a resource for studying how to Harness Egocentric Intent for Human-Robot Interaction. HEIR pairs human speech, measured gaze, and egocentric video with robot observations and teleoperated navigation-manipulation trajectories, while each session preserves successive episodes, with each request fulfilled through multiple execution subtasks. The captured data, together with human-authored annotations, support the study of intent grounding, execution planning, and task completion within a unified interaction setting. To demonstrate the value of HEIR, we conduct extensive offline experiments to evaluate how models benefit from additional human-centric context across these tasks, together with real-robot experiments that assess execution task completion. HEIR thus provides a valuable resource for studying user-centric robot interaction in naturalistic household environments. Our demo and dataset are available online.
Modern coding agents rely on establishing executable software environments before they can perform software engineering tasks such as testing, debugging, and patch validation. Existing setup agents typically control progress at either single-command or whole-script granularity, making recovery either too local to remove polluted state or too coarse to preserve useful progress. We present HerAgent (Harness EnviRonment Setup Agent), a block-level harness for coding-agent environment setup. HerAgent's main contribution is semantic setup blocks: a single progress-control abstraction used for execution, validation, checkpointing, repair, and rollback. Evaluated against six prior agents on four environment-setup datasets, HerAgent improves setup success in every comparison, achieving the largest gains on Installamatic-Bench (52.0% to 85.0%) and SetupBench (46.3% to 70.4%). It also reduces setup costs by up to 94.5%, showing that block-level progress control improves reliability and efficiency.
Decoder-only Large View Synthesis Models (LVSMs) with KV-cache have recently achieved state-of-the-art quality by regressing novel views from neural networks without reconstructing 3D geometry, while Test-Time Training (TTT) layers save computation but sacrifice quality. We propose Test-Time Baking (TTB), a strategy that learns to bake global scene information from cross-view attention into a lightweight MLP, enabling fast and high-fidelity novel-view decoding. The first key is to repurpose the fast-weight MLP of TTT to learn the cross-view attention mapping explicitly from its input-output pairs, rather than replacing attention entirely, achieving the same rendering efficiency as TTT while delivering substantially higher quality. The second key addresses a modality misalignment between context prefilling and novel-view rendering: we disentangle context tokens into camera-only context rendering tokens to serve as TTB input, and context ground-truth tokens with image content to serve as the TTB mapping target, ensuring the baked MLPs are relieved of the modality generalization challenge between context and novel-view processing. Our method achieves up to **23x** rendering speedup with quality on par with the state-of-the-art across several benchmarks. Project page: https://test-time-baking.github.io
Applying machine learning systems to new domains introduces bottlenecks in both labeled data acquisition and manual pipeline engineering. While recent MLE agents show promise in automating pipeline design, they have been studied almost exclusively in fully supervised settings, leaving them untested in realistic regimes where labeled data is scarce but unlabeled data is abundant. In this work, we introduce Agentic Semi-Supervised Learning (AgentSSL), framing SSL as a program search problem and studying how agents navigate this space under limited supervision. We find that even under weak feedback and matched compute, agents can synthesize code that combines and adapts established SSL techniques, exceeding the strongest human-engineered baselines in 8 of 12 settings by up to +21.8%. However, our results identify program selection as a critical bottleneck in AgentSSL, where performance drops by as much as 47% when selecting based on the agent's feedback metric. To systematically study this bottleneck, we introduce ProgramSelectBench, a large-scale benchmark with a combined total of 16,977 agent-generated and human-engineered SSL programs, and evaluate both unsupervised and active program selection methods. We further evaluate AgentSSL on real-world ecology detection tasks, highlighting the opportunities and challenges of transferring agent-discovered strategies beyond controlled benchmarks.
Large Language Models (LLMs) exhibit strong informal mathematical reasoning but struggle to generate mechanically verifiable proofs in formal languages like Lean. We present LEAP, an agentic framework that enables general-purpose foundation models to achieve state-of-the-art performance on automated formal theorem proving. LEAP leverages foundation model capabilities, such as informal reasoning, instruction following, and iterative self-refinement. By decomposing complex problems into smaller units, the system bridges formal proof construction with informal blueprints through continuous interaction with the Lean compiler. To provide a rigorous evaluation beyond increasingly saturated benchmarks, we introduce Lean-IMO-Bench, a benchmark of IMO-style problems formalized in Lean, with short statements yet highly non-routine and multi-step proofs across a wide range of difficulty levels. Empirically, on the latest 2025 Putnam Competition, an annual mathematics competition for undergraduate students in North America, LEAP solves all 12 problems, matching recent breakthroughs by frontier formal mathematical models. On Lean-IMO-Bench, LEAP boosts the one-shot formal solve rate of general-purpose LLMs from below 10% to 70%, notably surpassing the 48% benchmark set by a specialized, gold-medal-caliber IMO system. Furthermore, we demonstrate LEAP's research-level utility by autonomously formalizing complex proofs for open combinatorial challenges, including a verified proof for a key subproblem in Knuth's Hamiltonian decomposition of even-order Cayley graphs.
Multi-agent deep-research systems use delegation to parallelize search and keep long tool traces out of the main agent's context. Yet delegation helps only when task allocation and subtask execution are coordinated: the main agent must issue useful subtasks, while subagents must return reliable evidence. Existing training either optimizes only the main agent or broadcasts the same team advantage to all agents, which can reward misleading subagent reports and penalize useful ones. We introduce LEAD-RL (Learning Effective Agent Delegation with Reinforcement Learning), a joint main-agent-subagent training framework with contribution-aware credit assignment. A training-time rubric evaluates the research utility of each subagent report and gates its inherited team advantage, while final-answer correctness remains the only team reward; no token, step, or speedup reward is used. Across five benchmarks and two model families, LEAD-RL improves answer quality. On BrowseComp, it improves accuracy by 2.49 percentage points and reduces critical-path cost by 24.5% over joint training with shared team credit. During validation, accuracy increases by 6.78 points at approximately constant critical-path cost. These results show that contribution-aware credit assignment improves the quality-efficiency trade-off of multi-agent deep research.
This report summarizes the Mobile Unified Multimodal Understanding (MUMU) track of the 8th Large-scale Video Object Segmentation (LSVOS) Challenge, held in conjunction with ECCV 2026. MUMU asks for a single unified model that jointly performs multi-concept image tagging over quality, scene, and event vocabularies, open-vocabulary object detection, and concise image captioning, within a budget of 0.5B parameters and 8 GB of peak inference memory. Submissions are ranked by a weighted task score scaled by an efficiency factor. We describe the tasks, rules, and evaluation protocol, report the final results, and review the methods of the top three teams. All three leading solutions build on the compact Florence-2-base model, but extend it in different ways: lightweight tagging heads with cross-task inference refinement, a shared low-rank adapter trained on coverage-aware multi-source data with task-dependent decoding, and a training-free pipeline that maps captions to tags and applies detection test-time augmentation. The results show that a compact prompt-based vision-language model can serve all three outputs when supervision, decoding, and cross-task cues are designed for each task, while open-vocabulary detection on novel categories remains the main bottleneck.
AI-RAN distributes network control across multiple autonomous agents, namely xApps and rApps operating over shared radio resources. When an alert or degradation appears, operators currently lack a structured way to trace which decision caused it, what context triggered that decision, and how other agents interacted. Existing explainability tools focus on single-model feature attribution or linear event logging, neither of which captures cross-agent decision chains. We propose the Context Graph, a heterogeneous directed graph that organizes policies, measurements, decisions, and alerts into a queryable explanation layer. Its foundational primitive is the Decision-Event node acting as a causal junction, where incoming edges capture triggering context and outgoing edges track downstream consequences that are progressively upgraded through counterfactual verification. Instantiated over existing E2, A1, and O1 data flows, our proof-of-concept demonstrates that compliance and diagnosis queries are resolved at sub-millisecond latency over graphs of tens of thousands of nodes, while reliably distinguishing true causal impacts from spurious correlations.
Recommender systems, which retrieve items of interest to users based on their interaction histories, have recently witnessed several key technological advancements, benefiting from the semantic representation space and model-data scaling. Nevertheless, practitioners and researchers have frequently encountered critical inconsistencies between simple offline evaluations and complex online results, and this gap still exists in the era of large models. In this work, we address several key factors that contribute to this inconsistency: 1) the \textit{large-scale continual learning challenge} that assumes a practical demand to maintain long-term model effectiveness under dynamic item pools and distributional shifts; 2) the \textit{combinatorial diversity challenge} that requires the generation strategy to solve the accuracy-diversity trade-off in both the finer token level and the holistic item level; and 3) the \textit{multi-task reward balancing challenge}, which is particularly critical for modern generative models that can accurately model the sequence distribution but are less controllable for the multi-task demands. To address these challenges and further pave the way for the development of more realistic recommendation solutions, we introduce KuaiRecV2, a deliberately constructed benchmark with industrial-level datasets from Kuaishou. Specifically, we provide a large-scale interaction dataset with million-scale multi-modal item information and long-term interaction records across 30 days, with a dynamic video candidate pool. Additionally, we provide a rigorous benchmark that evaluates retrieval accuracy, diversity, and multi-task performance in one rubric system, where an effective advancement occurs if and only if the new method improves all metrics. The code is available at https://github.com/cxli2333/KuaiRecV2
Reinforcement learning with verifiable rewards (RLVR) produces rich training experience, yet outcome rewards reduce each attempt to a single number and leave most of this information unused. Training-time scaffolds, such as strategies or hints added to the prompt during training, offer a natural way to turn this information into guidance, but their design and adaptation are largely left to humans and tailored to specific domains. We introduce AutoScaffold, a unified AI-for-AI framework that turns a student's training experience into adaptive scaffolds, with minimal domain-specific information and no human intervention during training. A frozen teacher periodically examines this experience, diagnoses bottlenecks, and designs scaffolds in two forms: (i) strategies shared across a class of tasks and (ii) hints for individual hard tasks, adjusting the content and amount of support as the student learns. The scaffolds are removed at evaluation. We apply the same framework to ALFWorld, search-based QA, mathematical reasoning, and GPU kernel generation. When evaluated without scaffolds, students trained with AutoScaffold outperform those trained with the corresponding RL baselines in all four domains, with gains of 7.2 points in ALFWorld success and 4.0 points in average QA exact match. On GPU kernel generation, AutoScaffold increases the fraction of tasks with a correct kernel that runs faster than the PyTorch reference by up to 5.0 points. On math, our student surpasses QuestA's reported accuracy using approximately one-third as many student training steps. Analyses show that the teacher identifies a different failure in each task class and shifts its scaffolds from step-by-step workflows to targeted fixes as the student improves, and that its revised scaffolds help the student learn more than keeping the previous scaffold or using no scaffold.
LLM agents are increasingly used for open-ended, data-driven scientific discovery, yet most of them evaluate hypotheses pointwise, testing an empirical claim $h$ only against its complement $\neg h$. When evidence rejects $h$, the agent learns that one explanation is unlikely but not which alternative the evidence supports. To recover that alternative, we introduce Contrastive Discovery, which replaces the single hypothesis with a contrast set, a scientific question paired with mutually exclusive and collectively exhaustive competing explanations. Each experiment is scored by how its evidence redistributes belief across the set, and the question is resolved when belief concentrates on one explanation. Because contrast sets change only the unit an agent proposes and scores, an existing agent can use them to re-evaluate evidence it has already collected or to guide its search. Across six domains, $89\%$ of the pointwise candidates with the largest belief shifts are rejections. On the same evidence, contrast sets resolve $73\%$ of these candidates to a specific alternative. When contrast sets guide search, the agent resolves $67\%$ more questions than pointwise surprisal under the same experiment budget. In a blinded study, $38$ human experts chose the same explanation on $78\%$ of resolved questions and on $82\%$ of questions where the pointwise agent rejected its hypothesis and stopped. Together, these results support explicit competing explanations as a way for discovery agents to resolve more scientific questions from the same evidence and validation budget.
Recent advances in one-step text-to-image generation have enabled real-time synthesis with remarkable efficiency and quality. Previous reinforcement learning methods for one-step generators combine image-space reward optimization with diffusion noisy-space distribution matching. This paradigm brings challenges due to a mismatch between terminal reward optimization and the underlying generative dynamics. As a result, optimization tends to exploit stochastic degrees of freedom, often improving reward at the expense of image fidelity. To address this issue, we propose Diff-Instruct with Diffused Reward (DIDR), a data-free trajectory-level alignment framework derived from Integral KL minimization. DIDR propagates the RLHF-optimal reward-tilted clean-image distribution across all noise levels along the diffusion trajectory. We show that this objective admits the same minimizer as clean-image RLHF, while naturally inducing the Diffused Reward Score (DRS), which acts as a reward-driven correction to the reference score function. To make this practical, we further introduce the Diffused Reward Proxy (DRP), an efficient estimator of DRS based on differentiable short-step denoising. Extensive experiments demonstrate that DIDR consistently Pareto-dominates existing one-step SDXL baselines. Moreover, when transferred to a 6B DiT backbone (Z-Image), DIDR surpasses its 50-step teacher in preference alignment while requiring only a single generation step. Code: https://github.com/Junyi-W/DIDR
Fetal brain biometry provides a quantitative basis for assessing prenatal growth and identifying congenital brain abnormalities. However, comprehensive assessment remains limited by labor-intensive manual measurements and automated models that lose accuracy across imaging sequences and institutions. We introduce BioTTA, an anatomically guided test-time adaptation framework that measures 11 fetal brain biometrics from reconstructed three-dimensional magnetic resonance images. BioTTA adapts pretrained models to individual unlabeled volumes using geometric consistency and atlas priors, without access to source images. In a retrospective study across five centers in China and Switzerland, BioTTA improved measurement accuracy across three out-of-distribution datasets and two network architectures. It reduced mean absolute measurement errors by 46-82% relative to the corresponding models without adaptation, achieving errors similar to those of supervised fine-tuning. Its measurements were mapped onto normative growth trajectories derived from expert annotations, enabling gestational-age-specific assessment through a web-based reporting platform. Across four disease cohorts, the platform achieved 82.1-95.0% accuracy in binary detection of ventriculomegaly or cerebellar abnormalities, with performance similar to that of a senior radiologist. BioTTA provides a practical foundation for standardized fetal brain biometry across heterogeneous imaging settings, supporting scalable prenatal growth assessment and quantitative evaluation of brain abnormalities.
Pretrained video diffusion models (VDMs) offer a powerful foundation for World Action Models (WAMs), which jointly predict robot actions and future observations. Building such models requires injecting low-dimensional robot signals (e.g. proprioception, actions, and values) into a latent space shaped by natural videos, and a key design choice is the interface through which these signals enter: existing schemes treat it as a formatting detail, spending a separate latent frame on each signal or bolting on dedicated encoders and decoders. We argue that signal injection is instead a distribution-alignment problem: how well a representation matches the prior of the pretrained VDMs during denoising. Using the relative reconstruction error of a frozen VDM denoiser as a probe, we find that conventional spatial encodings sit far off this prior, whereas frequency-domain packing yields latents that the denoiser corrects far less. Motivated by this diagnostic result, we propose MetaWAM, which folds all robot signals into a single unified meta-frame: each signal is assigned to a disjoint frequency-domain channel band and lifted to the spatial latent by an inverse DCT, yielding a decodable interface aligned with the pretrained VDM prior without any added module or architectural change. A fixed band layout together with slot-aware routing lets the same meta-frame support policy, world-model, and value prediction over a shared set of robot variables. On RoboCasa, MetaWAM reaches 68.9% success while collapsing the per-signal frames into one and cutting inference latency by 39%, and attains 85.0% success on three real-world Franka tasks. Code and models are available at https://github.com/showlab/MetaWAM
Research agents achieve increasingly strong benchmark scores, yet these scores do not necessarily translate into reliable delivery in ongoing research projects. Many research-agent benchmarks emphasize standalone tasks and relatively self-contained workflows in clean or curated workspaces. However, researchers ask agents to complete one step within an ongoing project, where requests depend on earlier session decisions and workspaces contain unfinished files, outdated versions, and failed attempts. We introduce ARE-bench to evaluate whether research agents advance ongoing projects while respecting prior decisions and existing workspace state, grounded in real session context and in-progress workspaces. ARE-bench includes 75 tasks built on agent sessions and full workspaces from recruited researchers' own projects, spanning seven research stages, 33 requirements, and 150 sub-requirements, with 1,173 binary checklist items in total. Each task is anchored at a real user request, preserves its requirements, and summarizes the earlier requests still in effect as background. The complete workspace is replayed to that boundary without cleanup, preserving artifacts from earlier turns. Each checklist is compiled from an intent-aligned rubric and credits any valid solution. Across seven frontier models, the strongest passes every checklist item on only 66.2% of tasks, and Design and Execution rank among the hardest stages for every model. Three models pass about 69% of tasks when only delivery contracts and general norms are checked, but just 28.4% to 40.0% once historical and material or workspace requirements are also checked. Evaluating delivery without these contextual requirements can therefore overestimate research agents and mask differences between them.
Masked Diffusion Models (MDMs) are well known for fast text generation via parallel decoding. However, independent parallel sampling can commit to inconsistent predictions, and less is known about how to selectively revise these early token commitments. In this work, we introduce $\underline{\mathrm{P}}$lanned $\underline{\mathrm{U}}$nmasking-and-$\underline{\mathrm{R}}$evision Joint $\underline{\mathrm{L}}$earning (PURL) for MDMs, a training-based framework that seamlessly unifies parallel generation with selective token revision. We derive a normalized joint transition whose likelihood factorizes into edit selection, unmask selection, and token prediction. These components are trained jointly on partially masked sequences constructed according to the model's own unmasking policy. During inference, all three decisions are supported by a single shared forward pass, enabling the model to decode masked tokens and revise noisy tokens. Across logic puzzles, mathematical reasoning, and code generation, PURL improves accuracy at matched unmask budgets, with relative gains up to 47.4% accuracy improvement and up to an $8\times$ reduction in scheduled decoding rounds. Beyond these gains, our analysis further reveals general principles of revisable decoding: token edits become more effective as context accumulates; early edits can steer unfinished trajectories, whereas later edits tend to remain localized; and effective edits therefore depend on both calibrating when to intervene and learning where and how to revise. Together, these suggest that effective parallel generation depends not on revising more tokens, but on making the right commitments revisable at the right time.
The escalating complexity of deep neural networks hinders the deployment of federated learning (FL) on resource-constrained devices, especially for multimedia applications such as on-device image classification and speech recognition. Parallel split learning (PSL) has emerged as a promising solution by offloading a substantial portion of the computational workload to the server through layer-wise model partitioning, making it particularly suitable for resource-limited visual and speech intelligence scenarios. However, the absence of client-side aggregation renders PSL sus ceptible to severe training divergence and potential convergence failure under heterogeneous data distributions. To address this challenge, we propose PSL-RGHA, a gradient-aligned PSL framework that employs reference gradient half-space alignment mechanism to mitigate training divergence among client devices. PSL-RGHA comprises two novel components: reference gradient progressive construction (RGPC) and gradient half-space alignment (GHA). Specifically, RGPC evaluates the directional consistency of client gradients and progressively synthesizes a robust reference gradient via Weiszfeld updates, effectively approximating the global descent direction. Then, GHA constrains client gradients within the reference half-space defined by RGPC to mitigate gradient conflicts and rectify local updates. Extensive experiments demonstrate that our PSL-RGHA framework outperforms existing baselines in various settings.
Streaming multi-reference video generation establishes a scene from specified people and objects, then continues it under changing instructions. We introduce TaoMate-R1, a streaming multi-reference generator that organizes adaptation parameters through temporal role specialization. It assigns separate low-rank adapters to scene establishment and continuation on a frozen backbone. Each adapter performs both denoising and context writing for its role; Self Gradient Forcing jointly trains the roles through the history they construct. This allocation contrasts with operation-based specialization, which separates denoising from context writing across all chunks. A fixed-replay gradient analysis characterizes the update directions admitted by the two sharing schemes. On 32 multi-reference cases, we compare the schemes at the same two-adapter budget. Across ten approximately two-minute streams with changing instructions and shared teacher-audio conditioning, TaoMate-R1 reduces the mean boundary luminance jump from 1.96 to 0.613, with higher boundary feature similarity. It also achieves higher mean CLIP and DINO temporal similarity across short interactions, long streams, and public reference compositions. These results support grouping adaptation parameters by scene establishment and continuation for streaming multi-reference generation. Our code and model are released at https://github.com/TaoLiveAIGC/TaoMate-H3.