Ask questions across all of research

17 Aug 2026

Weiliang ChenHaowen Sun

Jun Gao

HarnessEval-W introduces an agentic evaluation framework for interactive visual world models, generating transparent "evidence trees" to explain model performance beyond scalar scores. The system aligns strongly with human preferences and provides fine-grained diagnoses for eight distinct evaluation axes, outperforming existing benchmarks in discriminative power.

17 Aug 2026

Reza Bayat

Ali Behrouz

Vahab Mirrokni

Proteus introduces incremental memory activation for long-context sequence modeling, which adaptively expands effective memory capacity during processing. This mechanism consistently improved average downstream accuracy and lowered perplexity across four memory-based model architectures, exhibiting enhanced robustness and length extrapolation, particularly at extended context lengths up to 16K.

A "latent-to-pixel" adaptation strategy from Alibaba Token Hub enables the training of pixel-space text-to-image diffusion models that achieve quality comparable to or surpassing latent-space models. This approach demonstrated 3.18x to 4.75x end-to-end inference speedups over latent-space counterparts and a 100.6x speedup over the original 100 NFE latent-space pipeline.

The rapid evolution of text-to-image (T2I) generation models has effectively solved the foundational challenge of raw pixel synthesis, shifting the community's focus toward fulfilling increasingly intricate user requests. While recent agentic image generation workflows enhance static inference with advanced capabilities like external knowledge retrieval and iterative reasoning, they mostly operate in isolated silos with fixed ``one-size-fits-all" topologies. This inevitably leads to severe compute-mismatch, where simple queries are forced through computationally heavy pipelines. To bridge this gap, we present GenRouter, the first unified workflow routing framework for agentic image generation. We first formulate GenCanvas, standardizing diverse agentic pipelines into a universal set of foundational primitives and executable templates. Operating over this unified space, GenRouter adaptively routes heterogeneous prompts to their optimal workflows via (i) demand profiling, (ii) experience matching, and (iii) Pareto filtering. Extensive experiments across diverse benchmarks demonstrate that GenRouter achieves superior visual alignment while reducing execution costs by over 95% and latency by 65% compared to heavyweight static pipelines. Furthermore, the system continuously self-evolves via accumulated experience, enabling robust zero-shot generalization that boosts performance and halves computational overhead.

Boundary representation (B-Rep) generation is a fundamental task in computer-aided design, yet the direct synthesis of high-fidelity and structurally valid B-Reps remains a major challenge. Existing deep generative methods suffer from two forms of brittleness: representation brittleness, caused by padding noise and feature contamination in the latent space, and generation brittleness, stemming from sequential error propagation and a train-inference mismatch due to non-differentiable validity enforcement. We propose HiFi-BRep, a novel framework that addresses these limitations through two synergistic contributions. First, a topology-aware encoder constructs a high-fidelity latent representation by eliminating padding via learnable queries and preventing feature contamination with topology-guided attention. Second, a single-stage decoder jointly predicts geometry and topology in parallel, embedding core manifold constraints as a differentiable learning objective. This design ensures mutual guidance between geometry and topology while avoiding cascaded errors. Extensive experiments show that HiFi-BRep significantly outperforms state-of-the-art methods in both structural validity and geometric fidelity, providing a robust solution for high-quality B-Rep synthesis. Code and models are publicly available at this https URL.

HiPHI, a large-scale benchmark, provides 617.5 hours of high-precision optical motion capture data, including 245.7 hours of human-object interaction with synchronized object trajectories and meshes. This dataset, developed by researchers primarily from Noitom Robotics, offers superior physical consistency and broader motion-space coverage than previous datasets, leading to enhanced performance in humanoid robot learning tasks and successful real-world deployment.

τ0-VLA is a hierarchical robot foundation model that employs world-model-guided test-time computation for high-level subtask generation. This approach dynamically allocates computational resources to evaluate hypothetical subtask outcomes, leading to improved task success rates on long-horizon manipulation tasks and enhanced robustness in out-of-domain scenarios.

ClawGym II introduces a unified black-box reinforcement learning framework for optimizing general autonomous agents through complex and opaque agent harnesses. The framework consistently improved agent performance, achieving Pass@1 gains of 9.98 to 17.28 points across different harnesses and benchmarks, while maintaining stable training dynamics and supporting joint optimization across heterogeneous harness environments.

GLM-5.3 is built on the same base model as GLM-5.2, with every reported gain coming from scaled post-training on long-horizon task environments. It posts open-source SOTA on Terminal Bench 3.0 (28.3 vs 4.6) and Agents' Last Exam (28.5), and a 50% improvement on Z.ai's in-house Code Bench while spending fewer output tokens. Cyber capability grew fastest of all: 84.5% on CyberGym is the best result on that benchmark, and exploitation scores more than doubled. Weights are slated for release two weeks after launch, once safety hardening completes.

13 Aug 2026

Damon Falck

Samer Sabri

Anja Surina

Researchers at Inherent developed Faraday, an AI Scientist agent designed to replicate research papers by reproducing experimental figures using a "Coding Agent as a Tool" (CAT) paradigm. Trained on the Replica task space with a novel rubric-based reward system, Faraday demonstrated superior scientific rigor and replication accuracy compared to leading frontier models, including Claude Opus 4.8 and GPT-5.5.

Amazon researchers developed an approach to quantify information gain in multi-turn conversational recommender systems by measuring the reduction in the Large Language Model's recommendation uncertainty. Fine-tuning LLMs with this uncertainty-reduction reward improved recommendation accuracy and conversational efficiency, with models achieving up to 3.32% Hit@1 and 5.21% Hit@5 on the INSPIRED dataset and requiring fewer turns to identify ground-truth recommendations.

Zetta introduces a closed-loop embodied harness that enables online self-evolution of code-based runtime critics and recovery skills for robotic agents, significantly enhancing task success rates without modifying base policy models. It integrates a dedicated rollout infrastructure (Z-Infra) to accelerate the learning process, demonstrating robust and transferable physical intelligence.

Latent On-Policy Self-Distillation (LOPD) introduces a framework where the privileged context for self-distillation is a learnable latent representation, moving beyond human-engineered artifacts. This approach consistently enhanced agent performance on tool-use and code generation benchmarks, yielding superior aggregate results and improved sample efficiency across multiple LLM backbones.

A neurosymbolic framework for embodied agents combines task-directed visual grounding with constrained symbolic planning to ensure plan executability by design. This approach achieves 94.5–99.5% task success in VirtualHome and 90.3–97.8% in ALFWorld, substantially outperforming direct VLM policies while reducing token generation and visual queries.

16 Aug 2026

Junfeng Chen

A new class of Flux-Form Spatiotemporal Neural Operators enables stable and accurate coarse-grained predictions for multiscale PDE systems by explicitly embedding local conservation laws and causal history dependence. The operators consistently reproduce long-horizon dynamics and time-averaged statistics, surpassing both traditional physics-based models and purely data-driven approaches across Burgers', Kuramoto-Sivashinsky, and Navier-Stokes equations.

Reinforcement Learning (RL) has been promising in single-turn LLM fine-tuning. However, long-horizon agentic reasoning introduces increasingly branching interactions and sparse rewards, exposing several limitations of RL: its heavyweight backpropagation makes it impractical to fine-tune larger LLMs, and longer-horizon trajectories make the credit assignment substantially harder. This paper argues that evolution strategies (ES) can be a better choice for fine-tuning long-horizon LLM agents. Compared with agentic RL, ES offers three key advantages: 1) Model Scalability: ES enables full-parameter optimization requiring only minimal, inference-level GPU memory, making it possible to fine-tune large LLMs. 2) Flexibility: its lightweight, black-box feedback interface makes ES fine-tuning easy to compose with prompt-space evolution (e.g., skill optimization & test-time compute); and 3) Long-Horizon Scalability: ES performs trajectory-level parameter attribution without decomposing rewards across horizons, yielding better scalability than Agentic RL as the horizon length grows. Based on this insight, we propose Agentic ESOpt, a full-parameter agentic fine-tuning framework tailored to flexible parameter–context co-evolution. At each step, Agentic ESOpt samples perturbations around the current LLM parameters, evaluates the resulting agents with rewards, and applies an online reward-weighted update. To improve the exploration–adaptation trade-off, Agentic ESOpt further introduces a cosine decay schedule of the perturbation scale σ. We evaluate Agentic ESOpt across both train-time fine-tuning and agentic test-time compute settings. On long-horizon Sudoku, Agentic ESOpt outperforms RL methods by 12.50% with Qwen3.5-4B. On WebArena-Lite, full-parameter optimization of Qwen3.5-27B improves the No Skill baseline by 6.69%, and combining Agentic ESOpt with Trace2Skill further improves the Trace2Skill baseline by 2.42%. In test-time automatic heuristic design, Agentic ESOpt performs online prompt–parameter co-evolution, improving its matched baseline in 28 of 36 settings.

Researchers from Google DeepMind, Carnegie Mellon, MIT, and Columbia University reduced the upper bound on the matrix multiplication exponent $\omega$ from 2.371339 to 2.371177 by employing modern gradient-based optimization and AI-driven algorithmic refinement with AlphaEvolve, enabling computation at an unprecedented scale.

A method combining Momentum Contrastive Learning and variational regularization is developed to improve decoding of heard continuous speech from multi-session electroencephalography (EEG) signals. This approach enhances linguistic fidelity (Character Error Rate) while preserving acoustic reconstruction fidelity (Mel-spectrogram Pearson Correlation Coefficient) by mitigating inter-session variability and preventing latent representation collapse.

Despite recent advances in unified multimodal models for multi-reference image generation, existing benchmarks remain organized around predefined task types (e.g., "subject composition"), which are ill-suited to this combinatorial setting and lead to fragmented coverage, uncontrolled complexity, and little diagnostic value. Recognizing that diverse multi-reference tasks share a common set of atomic operations, we adopt a capability-oriented perspective and formalize four operators: Anchor ($f$), Disentangle ($g$), Apply ($\oplus$), and Compose ($C$). Any multi-reference prompt can then be represented as a compositional formula over these operators, whose structural complexity is quantified by the number of operator slots. Building on this formulation, we construct TRACE-Bench, comprising approximately 1,600 evaluation cases across slot counts 1--8, built from 631 formula templates and around 4,000 reference images spanning diverse artistic styles and real-world subjects. The formula structure directly drives an operator-aligned evaluation protocol for per-capability scoring and a diagnostic tree analysis for recursive failure localization. Evaluating 9 leading models reveals insights invisible to holistic scoring: the primary bottleneck lies in disentanglement ($g$) and attribute binding ($\oplus$) rather than scene-level composition ($C$), with even the best model scoring only 0.74 on attribute fidelity. Project page: this https URL

Large-scale pre-training has made robot policy fine-tuning increasingly data-efficient, but this progress has largely been driven by datasets and embodiments built around simple parallel-jaw grippers. Dexterous, multi-fingered hands remain comparatively data-starved because real teleoperation is costly to scale, while human hand video is off-embodiment and requires lossy pose estimation and retargeting. We introduce Simulation Pre-training for Dexterity (SPD), a pre-training framework for dexterous manipulation that uses data entirely collected in simulation. In SPD, humans manipulate virtual objects inside a VR headset, enabling on-embodiment trajectories and robot-free collection. With the help of five operators, we collect 75 hours of multi-task dexterous manipulation over one week, and use it to pre-train a causal transformer on a sequence modeling objective. We study the benefits of simulation pre-training on real-world tasks by fine-tuning on 1-2 hours of physical demonstrations on a 56-DoF bimanual dexterous setup. We find that our approach outperforms training behavior cloning policies from scratch, showing that simulation teleoperation is a viable pre-training source for real-world dexterous manipulation. We perform ablation studies, measuring the benefits of history conditioning and short action chunks for reactive control.