Daily AI

OpenTerminal

Aggregates multiple free public market endpoints into a self‑hosted, terminal‑style web trading workspace: live quotes, candlesticks, screener, options, crypto, portfolio tracking and an optional Claude AI assistant. No paid API keys and automatic provider fallbacks.

Zi Wei Dou Shu · Open-Source Charting Engine

Charting engine and interpretation dataset for Zi Wei Dou Shu (Purple Star Astrology): deterministic natal chart algorithm, Four Transformations, pattern rules, and large v3/v4 interpretation datasets. Code MIT and data CC BY 4.0; judgment library and AI backend are closed.

UniMate: One Unified Model to Animate Diverse Skeletons

Generates articulated motion for arbitrary rigged 3D skeletons from text prompts with a single unified model and no per-skeleton retraining. Uses topology-aware diffusion-transformer architecture and a large UniML3D motion dataset to generalize across bipeds, quadrupeds, and other topologies; suitable for text-to-animation workflows and rapid prototyping.

Headroom

Compresses tool outputs, logs, files and RAG chunks locally before they reach the LLM to reduce token usage while caching originals for on-demand retrieval. Provides proxy/library/wrap modes, JSON- and code-aware compressors, and reversible compression.

Muse Spark Open 816B (NVFP4)

Generates text conditioned on multimodal inputs (text, image, audio, video) with a 1M-token context window; an open-weights mixture-of-experts model aimed at long-horizon coding and agentic workflows, with NVFP4 checkpoints to reduce memory footprint.

NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4

Quantized multimodal vision-language model for reasoning, coding, tool use, document and video understanding; 120B total (12B active) with NVFP4 quantization, native Multi-Token Prediction for speculative decoding, and up to 1M-token context — designed for deployment on NVIDIA GPU stacks.

Beam-1.0-501B-A23B (Reflection)

Text-only sparse Mixture-of-Experts language model for long-context reasoning, coding, and agentic tool use; 501B total / ~23B active parameters, 1M-token context, and BF16/FP8/NVFP4 checkpoints released under Apache 2.0.

Timestral-209M

Performs zero-shot probabilistic multivariate time-series forecasting with up to 32,768-step history and 21 quantile outputs; implemented as a 209M-parameter patch-based transformer supporting past and known-future covariates and CPU-friendly inference.

SharpaDex v1.0

Multimodal real-robot dataset of 28,993 teleoperated bimanual manipulation demonstrations (~300 hours) with synchronized multi-view RGB/tactile videos, 65D robot state/action, 60D fingertip tactile signals, and temporally aligned language segments; intended for imitation, visuomotor and visual‑tactile research, CC BY‑NC 4.0 (non‑commercial).

VDEBench

Benchmark dataset for bilingual (Chinese/English) visual document image editing with 3,506 examples covering 1,578 text edits and 1,928 table edits. Each example pairs input and target images with editing instructions, region masks, and structured annotations for OCR-level evaluation.

Embeds an agent's reasoning together with its search query to improve retrieval. Trains a 4B embedding model on synthesized deep-research trajectories, helping preserve intent across multi-step searches.

Trains an 8B model to search, synthesize evidence, and write cited research reports. Its reinforcement-learning rewards use evolving rubrics grounded in retrieved evidence and contrasting model responses.

Trains long-horizon web researchers by combining agentic mid-training, synthetic trajectories, and reinforcement learning. A 30.5B-parameter model activates 3.3B parameters per token and learns to plan, search, and synthesize evidence.

Tests research agents on 100 expert-written tasks across 22 fields. Separates report quality from citation reliability, using adaptive reference-based grading and checks that cited pages support individual claims.

Tests whether browsing agents can uncover obscure facts across the web through 1,266 short-answer questions. Answers are easy to check even when finding them requires sustained, creative search.

Explains how base-model behavior and GRPO normalization distort claims about reasoning gains. Introduces Dr. GRPO to reduce length bias, with a 7B training recipe reaching 43.3% on AIME 2024.

Improves language-model problem solving by generating candidate answers, retaining verified successes, and fine-tuning on them. Repeats this loop with an expectation-maximization interpretation instead of collecting more human-written solutions.

Generates computer-use tasks with executable setups and verifiable rewards. A generator and independent reward writer produce 32,112 verified examples across 110 desktop and mock-web environments for reinforcement learning.

Aligns language models by generating candidate outputs, filtering them with a reward model, and reusing the retained data for repeated fine-tuning. Translation experiments test this alternative to continual online sampling.

Turns human desktop demonstrations into training data for agents that reason over screenshots and control mouse and keyboard. Releases a cross-OS collection tool, AgentNet data, and models; the paper reports 45.0% on OSWorld-Verified for its 72B model.

Tests whether agents can finish 369 tasks in real desktop applications, using executable checks of the final state. The original study found a wide gap between agent and human success rates.

Tests whether web agents can use visual evidence to finish 910 tasks across shopping, classifieds, and social forums. Execution-based checks distinguish successful actions from plausible explanations.

Tests whether browser agents complete 812 tasks across realistic shopping, forum, development, and content-management sites. Evaluates final outcomes rather than matching a reference sequence of clicks.

Trains one model to delegate work recursively to copies of itself. Joint reinforcement learning teaches task splitting and communication, improving harder-task generalization while trading extra computation for parallel execution.

Helps agents decide when to retrieve, test, or commit by supplying calibrated uncertainty estimates alongside action costs. Experiments show more adaptive exploration and better cost-adjusted rewards than prompting or RL alone.

Trains web agents to gather information through longer interaction, using an online RL curriculum that expands rollout horizons. Tests on WebVoyager and WebArena show how exploration can complement longer reasoning before each action.

Separates web-agent planning from action execution and trains the planner on plans inferred from successful trajectories. Reports 57.58% success on WebArena-Lite and 81.36% on text-only WebVoyager.

Grounds language-based robot planning in learned skill value functions, selecting actions that are both useful and feasible. PaLM-SayCan reports 84% planning and 74% execution success across 101 real-world instructions.

Turns language commands into executable robot policies that combine perception, control APIs, loops, and geometry libraries. Hierarchical code generation recursively fills in missing functions to support new behaviors without policy retraining.

Splits complex tasks into reusable subtasks, each with its own prompt or symbolic handler. Recursive decomposition and retrieval improve GPT-3 reasoning on symbolic tasks and multi-hop question answering.

Breaks hard reasoning problems into simpler subproblems, then solves them using earlier answers. With 14 exemplars, code-davinci-002 reaches at least 99% accuracy on SCAN splits, including the length split.

Distills reusable reasoning strategies from successful and failed agent runs. Retrieved lessons guide later tasks, while memory-aware test-time scaling turns extra exploration into better memories.

Turns successful web-agent trajectories into executable, verified Python skills that expand the action space. On WebArena, success reaches 40.4%, versus 32.7% for a static agent and 36.3% for textual skills.

Extracts reusable workflows from past task trajectories to guide future web actions. Supports offline demonstrations and online experience, improving long-horizon navigation without manually written routines.

Lets language models move information between their context window and external storage. Memory editing, retrieval, and function chaining support conversations and document analysis beyond a single prompt's capacity.

Measures whether reusable skills improve agent task completion through paired tests. Across 87 tasks and 18 model–harness configurations, curated skills raise average pass rate by 16.6 percentage points; compact skill bundles work best.

Enforces output grammars during LLM decoding while reducing CPU validation overhead. Precomputed token masks and persistent stacks support nested formats without checking the full vocabulary at every step.

Lets language-model agents compose tools through executable Python and revise actions from interpreter feedback. Evaluates 17 models and releases 7,000 interaction trajectories for training agents that can debug their own code.

Explains when language models benefit from external programs, comparing tool selection, use, and creation. Evaluates performance gains against compute costs and identifies gaps in tool-use benchmarks.

Teaches a language model when and how to call external APIs using a few demonstrations and self-generated training data. Calls are retained only when their results improve prediction of subsequent text.

MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion

Improves stability and consistency for multi-turn image editing by training autoregressive diffusion editors with Single-Turn Distribution Matching Distillation and Teacher-Forcing Guidance. Combines a single-turn teacher for visual-quality supervision and trajectory-based teacher-forcing to preserve edits across successive turns.

Self-Evolving Coding Agents via Harness-Policy Optimization

Alternates harness and policy optimization so a coding agent uses its own policy to propose harness revisions that target evolving failure modes. Combines harness search (H-step) with RL-based policy updates (P-step), improving resolve rates on multiple coding benchmarks and generalizing out-of-distribution.

ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation

Learns executable language-conditioned local navigation by pretraining on instruction-conditioned Internet egocentric videos, aligning learned visual dynamics with simulated action labels, then jointly fine-tuning for navigation. Achieves high success rates on short-horizon benchmarks (86.5% OVON, 94.9% HabitatGS-LN) with real-robot validation and released code/data.

OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory

Provides a training-free multimodal reasoning harness that combines heterogeneous visual tools, persistent multimodal memory, and code execution so MLLMs can iteratively acquire and compose grounded evidence. Improves accuracy on 22 benchmarks by about 9.7–12.6 points while cutting comparative token costs.

Learning from Fleet Experience: A Collective Learning Paradigm for Robots

Shares execution traces across deployed robot fleets to improve a single manipulation policy using failure-guided experience retrieval fine-tuning (FERT) and a fault-tolerant asynchronous training system (FATS). Demonstrates large gains in real-world tasks and higher success with larger fleets, without extra expert corrective actions.