#Hugging Face

Extended Research on Looped Transformers with Truncated Backpropagation: Should the training loop be truncated?

Looped Transformers train better than truncated backpropagation with the same computational cost through full backtracking, extended regularity shows that when the model size is adjustable, two loops are the best. Looped Transformers increase computational depth by reusing shared layers without increasing the number of trainable parameters, usually using backtracking training (BPTT), but are often truncated to the last few loops to save training computational cost, which may lead to gradient bias and weaken the loop advantage.

HABIT-Bench released: Evaluating the ability of agents to learn habits from longitudinal weak evidence

HABIT-Bench evaluates an agent's ability to learn habits from synthetic user histories involving hundreds of interactions, requiring the agent to integrate scattered evidence and choose appropriate actions for new requests. Using Qwen3-235B-A22B and GLM-5.3-Flash, eight memory methods were evaluated, with the best configuration achieving a macro-average accuracy of 33.61% in a four-choice task, outperforming random guessing at 25%. Even with labeled evidence, there were still numerous errors, indicating that understanding how to guide decisions is more important than simply recalling history.

CodeCap-Bench: Can executable code serve as high-quality video subtitles?

CodeCap-Bench demonstrates that code captions outperform natural language captions in terms of reasoning accuracy and reconstruction fidelity on video question answering and reconstruction tasks. The benchmark includes 323 structured 2D content videos, covering five categories. Evaluation finds that mainstream closed-source MLLMs perform well in code caption generation but struggle to capture complex temporal dynamics, while open-source models have a significant gap.

Test-Time Memory Evolution method outperforms Test-Time Training in scientific discovery tasks

The Test-Time Memory Evolution (TTME) method guides scientific discovery more efficiently than Test-Time Training (TTT) by updating external storage without changing model weights. On five mathematical tasks, TTME outperformed TTT while reducing FLOPs computational cost by 82.8-84.6% and runtime by 47.7-58.4%.

ChronoReplay training framework enhances the spatiotemporal reasoning ability of streaming visual language models.

ChronoReplay training framework organizes heterogeneous spatiotemporal supervision into time-ordered interaction segments, preserves specific task answer formats, and balances learning from different supervision signals, enabling a single model to learn complementary spatiotemporal capabilities while maintaining temporal consistency. It achieves improvements on OVO-S-Bench, OST-Bench, and UCS-Bench benchmarks, with a 13.41-point increase on OST-Bench, while optimizing streaming inference performance through adaptive frame selection.

HiP3R: Hierarchical Pose Graph for Long-Term 3D Reconstruction

HiP3R uses hierarchical pose graphs to reconstruct and merge local videos into globally consistent camera trajectories and 3D geometry without retraining. It first optimizes robust block-level Sim(3) graphs to establish global layout and scale, then fixes this layout and estimates shared SE(3) cameras for each frame, converting overlapping window predictions into a single trajectory. Experiments on multiple benchmarks prove HiP3R's effectiveness in improving camera trajectories and 3D reconstruction, and it is applicable to both block-based and streaming backends without additional training or fine-tuning.

Hugging Face releases Old Ideas, Novel Problems study, evaluating the reliability of large language models in judging the reliability of research innovation.

Hugging Face releases Old Ideas, Novel Problems study, evaluating the reliability of six LLMs (GPT-5.1 to 5.4, Claude Sonnet 4.5 to 4.6) in judging the reliability of research innovation.

Δτ-Bench Released: Evaluating the Ability of Agents to Adapt to Dynamic Tool Ecosystems

Δτ-Bench evaluates the ability of agents to adapt to dynamic tool ecosystems, finding that the current strongest model only achieves a 42.3% task success rate. Δτ-Bench includes 1,588 multi-round tasks across 42 domains and 6 ecosystem states. Steward achieves higher task success rates in ecosystem changes through incremental skill maintenance and dynamic resource allocation.

AV-FABLE: Audio-Visual Tampering Forensic Benchmark for Long Videos

AV-FABLE introduces an audio-visual benchmark for sparse tampering in long videos, demonstrating that frozen audio-visual large language models' perceptual features contain forensic evidence that the best recovery is achieved by separating time decoders rather than language outputs. Local AI tampering in long videos often affects brief events that impact the authenticity of other content, making video-level classification insufficient for forensic analysis.

Goldman Sachs: AI Reshaping Market Pricing, Semiconductor Sector Influence Exceeds Oil Prices

Goldman Sachs notes that the influence of the semiconductor sector on the price trends of various assets has exceeded that of oil prices, explaining about 54% of the asset performance divergence over the past three weeks, while oil prices only explained 2%. Investor sentiment is closely tied to tech stocks, with Goldman's risk appetite indicator moving in sync with the performance of large-cap tech stocks. After Anthropic submitted its IPO application, the Nasdaq, S&P 500, NVIDIA, Philadelphia Semiconductor Index, and South Korean stocks all rose.

Work Agents: Systems, Capabilities, and Optimization: Work intelligent agents: systems, capabilities, and optimization

Propose a unified framework to integrate foundational models with runtime components for reliably executing complex digital tasks, introducing dimensions of work persistence and openness, and identifying five core capabilities and optimization methods.

Transformers releases v5.21.0: Added MiniCPM-OCR, Flint, UltraVBert, Gradium Phonon TTS, Pocket TTS and other models

OpenBMB's MiniCPM-OCR model released v5.21.0 version, using SigLIP visual encoder and Qwen3.5 language model, specifically for reading scanned pages and screenshots.

Tsinghua University open-sourced AKV, providing a cache engineering server for LLM agents on SGLang.

Tsinghua University and HKU researchers open-sourced AKV, providing KV cache engineering services for LLM agents. By discarding selected cache entries and reusing the rest, an average cost reduction of 27.7% and an average task score improvement of 1.7 points were achieved in 3 models and 4 agent benchmark tests.

RSIHub: A composite framework for agent evolutionary methods

RSIHub is an open-source framework that standardizes research on agent evolution by providing reusable components and a consistent evaluation infrastructure, used to build, compare, and extend related methods. Researchers can focus on evolutionary strategies while keeping evaluation external, with results associated with candidates and evaluation conditions.

UserArena: An Interactive User Simulation Benchmark for Long-Term Shopping Intelligent Agents

UserArena introduces an e-commerce interaction benchmark containing 64,729 session trajectories, covering 11,010 users, 566,102 items, and 13 action types, supporting five-level evaluation; proposes On-Policy Trajectory Distillation technology to improve simulator fidelity, experiments show it works best under the UserArena evaluation protocol, closest to offline simulation effects of production strategies.

How to make Claude Code agents use a logged-in Chrome browser.

Claude Code agents can work in a logged-in Chrome browser by installing a browser extension or enabling Chrome remote debugging. Using the Claude in Chrome extension (requires a paid Claude plan) allows for quick connection.

Hugging Face releases Carbon-X: the first open-source 14B model combining DNA and natural language

Hugging Face has released Carbon-X, an open-source DNA language model with 14.8B parameters, which is the first to decode DNA into English answers. It achieved an average score of 66.7% on 5 DNA benchmark tests, outperforming all test models on BioReason Coding (74.9%) and NextBase (39.4%). The model can run on 1 x 80 GB GPU, and the official GGUF Q8_0 version is 15.8 GB.

WeVisOne-VL: A unified visual perception model from object to pixel

WeVisOne-VL predicts structured tokens to achieve unified object-level and pixel-level visual perception, supporting tasks such as detection, localization, counting, OCR, segmentation, and matting. It employs a lightweight pixel decoder and multi-stage training, leading in over 30 benchmark tests while maintaining robust rejection of non-existent objects.

Factor-Aware Predictive Learning is released: a multi-modal large language model representation method for cinematic video generation.

Factor-Aware Predictive Learning (FAPL) enables video diffusion Transformers to utilize multi-modal large language model (MLLM) representations for sparse temporal reasoning and cinematic control, achieving this through staged grounding and optimization. FAPL consistently outperforms baselines on three cinematic video generation tasks (sparse-to-video, factor-aware editing, and multi-shot generation), validating its contributions in MLLM feature utilization, temporal control, and factor coordination.

TTR-Compressor reshapes sequence compression technology