#Coding

SPOT enhances the speculative decoding of diffusion language models by predicting the submission order.

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.

Qwen-Image-PE: Enhances text-to-image prompt enhancement capabilities through a two-stage training process.

Qwen-Image-PE enhances text-to-image prompt enhancement capabilities through a two-stage training process (based on title reverse distillation and group design rewards). When using Qwen-Image-2.1 on Qwen-Image-Bench, Qwen-Image-PE-397B improves by 7.81 points compared to the original request, and Qwen-Image-PE-9B surpasses all evaluated PE baselines.

Databricks improves document extraction accuracy through ai_extract and MLflow Prompt Registry

Databricks improves document extraction accuracy by building a document extraction process that includes a review step, using ai_extract combined with JSON schema and global instructions, and leveraging MLflow Prompt Registry for versioned configuration. Finally, it validates the improvement effect on a fixed set of manually labeled data, enhancing the overall performance of the system.

Mirendil released its first model, Fern 0.1, supporting a 10M-Token context

Mirendil released its first AI model, Fern 0.1, supporting a context window of up to 10M-Tokens, equivalent to approximately 23 MB of code or 500,000 lines, achieving the first complete loading of a Megatron-LM level codebase.

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.

OpenAI releases GPT-6.1 Terra: API prices halved and performance improved

OpenAI releases GPT-6.1 Terra, pushing the cost-performance boundary to new heights, API prices halved compared to GPT-5.6 Terra to 1 dollar per million input tokens, 5 dollars per output token

ByteDance's Imagination AI Short Drama Studio has launched, offering self-developed script, image, and storyboard models.

The Imagination AI Short Drama Studio has launched, offering self-developed script models DreamScript (supporting multi-genre generation, hit diagnosis, and 100,000-character long story analysis), image models Anycook (supporting differentiated character settings and multi-style scene construction), and storyboard models DreamShot (dual-mode creation, 3D director's desk dynamic preview).

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.

Stanford and Berkeley researchers launch TokenGate: Boosting the speed of long-context pre-filling by up to 11.3 times

Researchers from Stanford and Berkeley have introduced the TokenGate module, which uses learnable token pruning techniques to improve the speed of long-context pre-filling by 1.4 to 11.3 times without losing accuracy, and reduces peak pre-filling memory consumption by 1.5 to 3.4 times. It outperforms 6 baseline models on 5 tasks, with the best performance showing an 11.3-fold speed increase and a 3.42-fold memory reduction on the NIAH 16k task. The code and checkpoints are open-sourced.

Cognition has open-sourced the Agent plugin Devin Vulnerability Fixes, used to fix recent cybersecurity threats.

Cognition has released a free, open-source Devin Vulnerability Fixes plugin, containing three skills to fix three critical remote code execution (RCE) vulnerabilities in PHP2Shell, Node.js, and GitLab CVE-2026-110264.

The FRED framework enhances the future video representation capabilities of world action models through self-distillation and temporal conditioning.

The FRED framework enhances the future video representation capabilities of world action models through self-distillation and temporal conditioning, consistently outperforming competing models in real-world operational scenarios, achieving significant gains in zero-shot generalization on long-term and out-of-domain tasks, and increasing the success rate of human video skill transfer from 19.4% to 50.0%.

LitPivot: Developing research ideas through dynamic context in the literature landscape

LitPivot helps researchers adjust their research direction by reading previous literature, dividing research ideas into problems, contributions, and evaluation plans, and providing literature-based criticism and revision suggestions for each section. In a laboratory study involving 17 researchers, domain experts rated the literature foundation of novel ideas developed using LitPivot better than ideas developed using text editors and AI chat tools. Participants also reported stronger literature comprehension and improved understanding of the novelty and practicality of their ideas.

IQuest-Q1 Open Source Model: Enhancing the Basic Capabilities of Agent CLI Systems through Multi-Stage Pre-training

IQuest-Q1 is an open-source sparse expert mixture (MoE) agent base model with a total of 320B parameters (15B activation parameters), achieving powerful encoding and general capabilities through multi-stage pre-training, mid-training, and post-training.

LaLa-World: Using Natural Language Actions to Extend Latent World Models

LaLa-World is a language-conditioned latent world model that achieves efficient model prediction planning and strategy initialization by combining simulation, robotics, and video data. The model adopts a language-conditioned interface of video generation models, enabling JEPA-style latent world models (LWMs) to learn across different domains and data sources.

Cola2: Adaptive Hierarchical Language Model improves hierarchical language generation through continuous semantic compression and balanced discrete generation.

Cola2, through dynamic span splitting, masked diffusion decoding, and stream matching priors jointly optimizes continuous semantic compression, reconstruction-generation balance, and encoder-decoder adaptation, verifying its core design on five downstream tasks. Using 2100 billion training tokens, it demonstrates competitive non-saturated scalability, with a 3460 billion source token checkpoint showing an average of 60.10 points on eight tasks and surpassing Llama-3-8B on HellaSwag and StoryCloze.

Different tasks, different time steps: temporal task-aware quantization for diffusion language models T²Quant

T²Quant framework predicts accuracy sensitivity through minimal unlabeled calibration prompts, allocating activation precision within a fixed bit-width budget, improving existing quantization methods by 6.7 percentage points across 3 diffusion language models and 6 benchmark tests without increasing average accuracy budget.

AgenticRoboRL framework enhances robot planner performance through post-training

AgenticRoboRL framework improves robot planner performance via supervised fine-tuning and multi-round reinforcement learning, boosting the success rate of a 9B parameter planner from 18.1% to 71.0% on LIBERO-PRO without external memory, approaching Codex (GPT-5.5)'s 72.1% with external memory. The framework adapts to different execution mechanisms, linking recorded planner interactions with task outcomes to achieve a universal reinforcement learning training process through adapters.

Recursive releases the open-source 40B fuzzy model Flint, designed specifically for agents.

Recursive releases the open-source 40B fuzzy model Flint, 2.5B active parameters, performs best among MoE models below 50B, supports inference and OpenAI compatible tool calls, 4-bit weights about 23GB, SWE-bench Verified 74.2%.

How to train Qwen3 on AWS Trainium using torchtitan

AWS introduces three methods to train Qwen3 on AWS Trainium with torchtitan, including using Transformers v5, torchtitan native configuration, and Neuron Kernel Interface (NKI) mixture-of-experts (MoE) kernels.