2026-07-24·Daily digest

🔬 Research Pulse

Daily Digest

July 24, 2026

🤖 AI

🧠 LLMs

1. OpenForgeRL: Train Harness-native Agents in Any Environment

Authors: Xiao Yu, Baolin Peng, Ruize Xu... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can open-source infrastructure train agents end-to-end inside complex, stateful inference harnesses (Claude Code, Codex, OpenClaw) that existing SFT/RL stacks cannot natively express?

Summary: OpenForgeRL is an open-source framework that trains harness-native agents end-to-end by proxying the harness's model calls into a standard RL stack and orchestrating rollouts in per-task Kubernetes containers. It demonstrates competitive results on tool-use and GUI benchmarks with modest task budgets, and offers analysis of how harness choice and RL shape agent behavior.

Key Results: OpenForgeClaw reaches 31.7 pass^3 and 55.9 pass@3 on ClawEval, and 33.7 on QwenClawBench. OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager — using only hundreds to a few thousand training tasks. Both match or surpass open baselines of similar size, with GUI models rivaling models several times larger.

Key Findings:

Training directly inside real harnesses is feasible without rewriting them, using a proxy + container orchestration design.
Some harnesses (e.g., OpenClaw vs. ZeroClaw vs. Codex) are substantially harder to learn than others, meaning harness choice itself is a training-relevant variable.
RL improves self-verification, tool coverage, and multi-step plan completion, but error-recovery ability remains weak.

Technical Novelty: A lightweight model-call proxy that transparently intercepts and records harness LLM calls as RL training data, paired with a Kubernetes orchestrator giving each rollout its own remote container — decoupling training from inference so any harness in any environment can be trained without modifying the harness itself.

What's New: Prior open RL stacks require agents to fit a fixed inference loop; OpenForgeRL is the first to treat arbitrary production harnesses as first-class training targets by intercepting model calls rather than reimplementing the harness.

Extension Opportunities:

Add explicit error-recovery training signals (e.g., adversarial failure injection) since the paper identifies error recovery as a remaining weakness after RL.
Extend the proxy+K8s architecture to multi-agent harnesses where several harness instances cooperate or debate within a single rollout.
Use the harness-difficulty analysis to design a curriculum that starts with easier harnesses (ZeroClaw) and transfers policies to harder ones (OpenClaw/Codex).

Replicability: Framed as open-source ('OpenForgeRL', built on veRL). Reproduction requires a Kubernetes cluster capable of running many concurrent containerized rollouts (GUI/browser/computer-use environments are heavy), plus GPUs sufficient for RL on models of at least the sizes benchmarked; exact compute is not specified in the abstract.

Research Gaps:

Error recovery in long-horizon agent trajectories remains unsolved even after RL.
No principled account of why certain harness designs are systematically harder to optimize.

2. MedGame: Storytelling Gamification Empowered by Large Language Models for Medical Education

Authors: Qian Wu, Xinrong Zhou, Zizhan Ma... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can LLMs organize entire clinical cases into decision-centered learning trajectories for medical education, rather than being limited to localized interactions like Q&A or single-turn feedback?

Summary: MedGame transforms static clinical cases into interactive storytelling games via a dual-engine LLM design that generates clinical narratives and orchestrates multimodal decision-point rendering. The authors release a 5,000-case benchmark and show that fine-tuning closes the gap between open-source and commercial LLMs, while a student pilot confirms improved engagement over text-only formats.

Key Results: Introduced MedGame Bench, a 5,000-case benchmark with evaluation protocols for Medical Narrative Generation and Story Direction. Demonstrated that task-specific fine-tuning substantially improves open-source LLMs on the benchmark, narrowing the gap with commercial models. A pilot student study showed learners perceived MedGame as more engaging and useful than text-only alternatives.

Key Findings:

Dual-engine (Narrative Designer + Story Director) decomposition enables converting static cases into executable decision trajectories
Task-specific fine-tuning substantially closes the performance gap between open-source and commercial LLMs on MedGame Bench
Students rate gamified case-based learning as more engaging and useful than text-only alternatives in pilot study

Technical Novelty: Dual-engine architecture separating a Medical Narrative Designer (case-grounded storyline synthesis with states and decision nodes) from a Story Director (dependency-aware multimodal orchestration planning). This decouples clinical reasoning structure from rendering logic, unlike prior single-pass generation approaches.

What's New: Shifts LLM medical education from localized Q&A to whole-case decision trajectories, and introduces the first benchmark (5,000 cases) evaluating both narrative generation and story-direction orchestration as distinct sub-tasks.

Extension Opportunities:

Add multimodal patient inputs (medical imaging, lab results, vital signs streams) to decision nodes for more realistic simulation
Extend to multi-agent scenarios where students collaborate as care teams (nurse, resident, attending) within the same storyline
Integrate adaptive difficulty using learner performance history to personalize case complexity and branching depth

Replicability: Authors state they released an interactive platform and MedGame Bench (5,000 cases). Compute requirements not specified in abstract, but fine-tuning open-source LLMs suggests standard GPU setup (likely multi-GPU for 7B–70B models). Full code/data availability would need verification from the paper repo.

Research Gaps:

Pilot student study is small-scale; long-term learning outcomes and clinical skill transfer are not measured
Evaluation focuses on narrative quality and orchestration rather than downstream diagnostic accuracy or patient safety impacts

⚡ CUDA/Systems

1. SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation

Authors: Junsong Chen, Jincheng Yu, Yitong Li... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can video diffusion transformers achieve softmax-level generation quality while avoiding the quadratic cost of full attention, enabling efficient high-resolution long-video generation on a single GPU?

Summary: SANA-Video 2.0 is a 5B/14B hybrid video diffusion transformer that interleaves gated linear attention with periodic softmax anchors (3:1) and adds Block Attention Residuals to propagate anchor features across depth, trained from scratch. It matches full-softmax video DiT quality (VBench 84.30) while running 3.2x faster than matched softmax baselines and, with Sol-Engine, 120x faster than Wan 2.2-A14B at 720p/5s on one H100.

Key Results: SANA-Video 2.0 (5B/14B) achieves VBench 84.30 in 13.2s at 480p on a single H100 with 40-step sampling. Hybrid 3:1 linear-to-softmax ratio (25% softmax) is empirically optimal. Block Attention Residuals boost deep-layer effective rank by ~12%. Compiled DiT forward pass is 3.2x faster than full-softmax baseline at 720p/60s (gap grows with duration). Sol-Engine adds 3.58x speedup, reaching 13.06s at 720p/5s — 120x faster than Wan 2.2-A14B on one H100.

Key Findings:

25% softmax anchors (3:1 hybrid) is the optimal quality-efficiency trade-off, established via reduced-resolution proxy studies
Block Attention Residuals raise deep-layer effective rank by ~12%, mitigating the rank collapse of pure linear attention
From-scratch hybrid training beats linearizing pretrained softmax models and scales cleanly to 14B and 720p on a single H100

Technical Novelty: Two coupled contributions: (1) Hybrid Linear-Softmax Attention at a fixed 3:1 ratio using gated linear attention with periodic gated-softmax anchors, trained from scratch rather than linearized post-hoc; (2) Block Attention Residuals (AttnRes) that route completed block summaries into later linear layers to propagate softmax-anchor information across depth and restore effective rank.

What's New: Prior linear-attention video work typically linearizes pretrained softmax models or uses uniform linear layers; SANA-Video 2.0 instead trains a native hybrid with periodic softmax anchors and introduces AttnRes to reuse anchor features downstream — a structural, not post-hoc, remedy for linear attention's rank deficiency.

Extension Opportunities:

Apply the hybrid linear-softmax + AttnRes recipe to audio or 3D/point-cloud diffusion where long-sequence scaling similarly bottlenecks softmax attention
Search the softmax-anchor ratio and placement adaptively per layer/timestep rather than fixed 3:1, potentially via learned gating or NAS
Extend AttnRes routing to cross-modal conditioning (text/image/audio residuals) to enrich anchor features without extra full-attention cost

Replicability: Abstract does not mention code/weight release. Reproduction requires substantial compute — from-scratch training of 5B/14B video DiTs plus reduced-resolution proxy ablations, likely multi-node H100 clusters. Sol-Engine kernel fusion/caching/sparse attention stack would also need to be reimplemented for the reported inference numbers.

Research Gaps:

No principled theory for choosing anchor ratio/placement; 3:1 is empirical and may not transfer to other modalities or resolutions
Quality gap versus the largest full-softmax DiTs at very long durations and 4K resolutions remains unquantified

🦾 ROBOTICS

1. AXIS: A Growable Community-Driven Data Engine for Scalable Robot Manipulation

Authors: Mengfei Zhao, Dihong Huang, Yikai Tang... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can we scale robot manipulation data collection beyond the bottlenecks of specialized teleoperation hardware, centralized operator pools, and fixed task suites — while ensuring the resulting community-sourced data is actually usable for training policies?

Summary: AXIS is a growable, community-driven data engine that lowers the barrier to robot-manipulation data collection via browser teleoperation and an automated task-generation, validation, and augmentation pipeline. It ships a 207-task, 50K+ trajectory dataset and a unified VLA evaluation suite, and demonstrates that continual pretraining on AXIS improves π₀.₅ by 5.8% overall and by 37.3% over a RoboCasa365-pretrained baseline.

Key Results: AXIS assembled 207 diverse tasks and 50K+ trajectories via browser-based teleoperation. Continual pretraining on AXIS improved π₀.₅ overall success rate by 5.8%, and beat a model pretrained on RoboCasa365 by 37.3%. Scaling was consistent with data volume, with the largest gains under layout, sensor-noise, and camera perturbations.

Key Findings:

Browser-based teleop plus automated validation can produce training-usable data at scale without specialized hardware
Continual pretraining on AXIS yields consistent scaling with data volume for VLA policies
The largest gains from AXIS pretraining appear under distribution shift — layout, sensor noise, and camera perturbations — suggesting the diversity from community collection is doing real work

Technical Novelty: The novelty is systems-level rather than algorithmic: a browser-based teleop stack that removes the hardware barrier to contribution, combined with an automated pipeline (task generation + success checking + quality filtering + trajectory smoothing + visual/physics augmentation) and a task-snapshot + held-out evaluation protocol that turns crowd-collected data into reproducible benchmarks.

What's New: Prior scaling efforts (Open X-Embodiment, RoboCasa, DROID) relied on centralized operators or fixed simulated task suites. AXIS is the first to combine browser-native contribution, automatic task synthesis + validation, and a snapshot-based held-out benchmark into a single continually-growing engine.

Extension Opportunities:

Add cross-embodiment support (bimanual, humanoid, mobile bases) to the browser teleop layer to broaden the task distribution beyond single-arm manipulation
Build an active-learning task generator that prioritizes synthesizing tasks where current policies fail on the held-out protocol, closing the data-collection loop
Layer in reward-model or preference-based filtering on top of the existing success checker to distinguish 'succeeded but sloppy' trajectories from high-quality ones for imitation learning

Replicability: Abstract does not confirm code/dataset release, though the 'community-driven' framing strongly implies public availability. Reproducing continual pretraining of π₀.₅ on 50K+ trajectories would require multi-GPU compute (likely 8+ A100/H100-class GPUs for days); running the eval suite alone would be much cheaper.

Research Gaps:

Quality-vs-quantity trade-offs from anonymous crowd contributors are not deeply characterized — how much filtering is actually needed and what fails silently?
Currently single-embodiment single-arm; no evidence yet that the pipeline generalizes to bimanual, mobile, or dexterous hands where teleop is much harder in a browser

2. Beyond Episodic Evaluation: Memory Architectural Bottlenecks in Sequential Embodied Question Answering

Authors: Zikui Cai, Kaushal Janga, Tan Dat Dao... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How do memory architectures in Embodied Question Answering (EQA) agents behave when evaluated sequentially (multiple questions per scene with carried-forward memory) rather than under the standard episodic formulation where state resets between tasks?

Summary: The paper argues that standard episodic evaluation of EQA agents masks a critical architectural weakness: memory that doesn't persist visual-semantic evidence across queries. By evaluating sequentially in the same scene, the authors show that only spatially-grounded 3D visual memory breaks the accuracy-efficiency tradeoff, validated in both simulation and on a real mobile robot.

Key Results: Through extensive experiments in simulated environments plus real-world mobile robot validation, the authors demonstrate that structured, spatially-grounded memory (mapping persistent visual observations onto metric 3D geometry) breaks the accuracy-efficiency tradeoff in sequential EQA, simultaneously achieving higher answer accuracy and lower navigation costs. They show 2D occupancy maps preserve traversability but not visual-semantic evidence, and episodically-trained agents suffer severe temporal mismatch when exposed to continuous multi-query histories. (Specific numerical benchmarks not disclosed in abstract.)

Key Findings:

2D occupancy maps preserve traversability but discard the visual-semantic evidence later questions depend on, causing sequential EQA failures
Agents trained on short-horizon episodic data exhibit severe temporal mismatch when their context is extended with continuous multi-query histories
Mapping persistent visual observations onto metric 3D geometry simultaneously improves answer accuracy AND reduces navigation cost — breaking the usual tradeoff
Findings transfer from simulation to a real-world mobile robot, indicating the memory bottleneck is architectural rather than a simulation artifact

Technical Novelty: Reframes EQA evaluation from episodic to sequential/continuous, exposing that widely-used episodic training regimes produce agents whose inherited context suffers 'temporal mismatch' under multi-query histories. The prescriptive contribution is identifying that persistent visual observations must be projected onto metric 3D geometry (not just 2D occupancy) to preserve reusable visual-semantic evidence across queries.

What's New: Prior EQA work almost universally uses episodic evaluation with per-task memory reset; this paper is among the first to isolate sequential memory as the bottleneck and to comparatively characterize how different memory representations (occupancy, episodic context, 3D-grounded visual) degrade or succeed under continuous operation.

Extension Opportunities:

Build a hybrid memory architecture combining 3D spatial grounding with a learned semantic retrieval index (e.g., CLIP-embedded voxels) to enable open-vocabulary queries over accumulated scene memory
Introduce memory decay/consolidation policies that prune stale visual evidence over long horizons, testing whether bounded-memory agents can match unbounded ones on hour-long or day-long sequential deployments
Extend to dynamic scenes where objects move between queries, requiring the spatial memory to update rather than just accumulate, and benchmark against static-scene assumptions this paper likely relies on

Replicability: Abstract does not mention released code, data, or specific compute requirements. Experiments span simulation (likely Habitat/HM3D-class) plus a real mobile robot platform. Full reproduction of the real-world component would require physical hardware; simulation reproduction likely feasible on a single GPU workstation if code is released.

Research Gaps:

No standardized benchmark or metrics specifically designed for sequential/continual EQA with cross-query memory reuse
Lack of investigation into memory scalability — how spatially-grounded 3D memory scales to long horizons, large scenes, or dynamic environments

3. RL-MACRO: A Cybernetic Closed-Loop Intelligence Framework for Multimodal Adaptive Robotic Craniotomy

Authors: Xiao Zhang, Jiaxuan Li, Renzhen Le... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can an autonomous surgical robot safely perform craniotomy when critical states (cutting temperature) cannot be directly measured and tissue properties vary unpredictably, requiring closed-loop adaptation under partial observability?

Summary: RL-MACRO is a closed-loop framework for autonomous robotic craniotomy that infers unmeasurable cutting temperature from force and sound via a CNN-LSTM observer, then uses an offline IQL policy with a dual-head Actor to adaptively modulate feed rate, spindle speed, and cutting depth. It demonstrates safe, efficient bone cutting on ex vivo goat skulls and bovine ribs despite partial observability and tissue heterogeneity.

Key Results: A CNN-LSTM observer reconstructs hidden cutting temperature from fused force and sound signals with R^2=0.939 and MAE=1.717 deg C. The full RL-MACRO pipeline was validated on bovine ribs and six ex vivo goat skulls, demonstrating robust perception, adaptive recovery from force/temperature excursions, and smooth execution on irregular anatomical surfaces.

Key Findings:

Force+sound fusion via CNN-LSTM accurately reconstructs hidden cutting temperature (R^2=0.939, MAE=1.717 deg C), enabling thermal-aware control without direct sensing
Offline IQL with a dual-head Actor can jointly optimize three cutting parameters while respecting hard safety bounds on force and temperature
The system recovers adaptively from force/temperature excursions and executes smoothly on irregular skull surfaces via online trajectory re-planning and velocity servoing

Technical Novelty: Novel combination of (1) a CNN-LSTM soft-sensor that infers unmeasurable cutting temperature from force+acoustic fusion, (2) using this reconstructed state as a belief input to an offline IQL policy, and (3) a dual-head Actor that jointly coordinates feed rate, spindle speed, and cutting depth under hard safety bounds — integrated into a cybernetic perception-decision-execution loop for bone cutting.

What's New: Prior autonomous bone-cutting work typically assumes measurable states or uses fixed cutting parameters; this paper is the first to combine deep state estimation of an occluded thermal variable with offline RL policy optimization under safety constraints for craniotomy, framed as a cybernetic closed loop.

Extension Opportunities:

Add vision/thermal imaging as an additional modality to the CNN-LSTM observer for tumor boundary detection and cross-validation of the inferred temperature state
Replace offline IQL with online fine-tuning during procedures (e.g., IQL-to-online transitions) so the policy adapts to per-patient bone density in real time
Extend the dual-head Actor framework to other constrained surgical tasks (spinal drilling, dental implantology) where thermal necrosis and mechanical overload share similar dynamics

Replicability: No explicit code/data release mentioned in the abstract. Reproduction would require a surgical robot arm with force/torque and acoustic sensors, ex vivo bone specimens (goat skulls, bovine ribs), and modest GPU compute for training the CNN-LSTM observer and offline IQL policy — the bottleneck is hardware and biological samples, not compute.

Research Gaps:

Validation is limited to ex vivo tissue — in vivo dynamics (blood flow cooling, irrigation, patient motion) are not addressed
Offline IQL cannot adapt to distributional shifts encountered on unseen bone types without retraining

🏥 HEALTHCARE

1. GraphVid: Interactive Graph-Controllable Video Generation

Authors: Vedant Shah, Onkar Susladkar, Tushar Prakash... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can we enable precise, flexible multi-subject control in video generation without requiring users to draw ambiguous per-object motion trajectories that break down under occlusion and scene complexity?

Summary: GraphVid introduces graph-conditioned image-to-video generation, using structured interaction graphs instead of text or trajectories to specify multi-object relationships. Paired with the new GraphVid-Bench dataset, it achieves substantial gains over Motion-I2V (up to 39.9% FID reduction, 37.6% FVD reduction) while using less data and fewer parameters, demonstrating that structured semantic interfaces are a viable control paradigm.

Key Results: GraphVid, a graph-conditioned I2V model, outperforms Motion-I2V despite using less data and fewer trainable parameters: FID reduced by up to 39.9%, FVD reduced by 37.6%, PSNR improved from 9.87 to 15.98, and SSIM improved from 0.38 to 0.61. The authors also curate GraphVid-Bench, a large-scale interaction-centric video dataset with structured relational annotations.

Key Findings:

Interaction graphs are a more scalable and less ambiguous control signal than per-object trajectories, particularly under occlusion and overlap
Graph conditioning yields large quality gains over Motion-I2V across FID, FVD, PSNR, and SSIM despite reduced training budget
Interaction-centric supervision (GraphVid-Bench) is a key enabler for training controllable multi-subject video models

Technical Novelty: Replaces pixel-space trajectory or text conditioning with a structured interaction graph (nodes = subjects, edges = relations) as the control interface for I2V diffusion. Introduces a paired dataset (GraphVid-Bench) with explicit relational annotations to train interaction-aware models — a semantic-graph conditioning path prior methods lacked.

What's New: Prior controllable video generation relies on text prompts (imprecise for multi-object interactions) or trajectory maps (poorly scaling, occlusion-ambiguous). GraphVid pivots to structured relational graphs, contributing both the conditioning mechanism and a matched dataset with explicit interaction annotations.

Extension Opportunities:

Extend the interaction graph schema to support temporal/hierarchical relations (e.g., sequential sub-events, causal chains) for longer-horizon multi-shot video generation
Build an LLM-driven front-end that auto-generates interaction graphs from natural language prompts, letting non-expert users control complex scenes without manual graph editing
Adapt the graph-conditioning paradigm to 3D/4D scene generation or embodied simulation, where structured object-relation control is critical for physics-plausible interactions

Replicability: The abstract does not mention a public code or dataset release, though GraphVid-Bench is described as a curated resource that would presumably be shared. Compute is described as lower than prior motion-control baselines (fewer trainable parameters, less training data), suggesting reproduction is feasible on a modest multi-GPU setup typical for I2V fine-tuning.

Research Gaps:

Lack of interaction-aware, relationally-annotated video datasets for training multi-subject controllable generators
Absence of semantic (as opposed to pixel- or trajectory-level) control interfaces that scale with scene complexity

2. 3D-Aware VLMs with Implicit and Explicit Geometries

Authors: Wenhao Li, Xueying Jiang, Quanhao Qian... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can 2D-trained VLMs be equipped with 3D spatial awareness for fine-grained spatial understanding and reasoning tasks without requiring additional 3D sensor inputs (e.g., depth, point clouds)?

Summary: VLM-IE3D augments 2D vision-language models with 3D spatial awareness by extracting two complementary token streams from RGB video alone — implicit geometric priors and explicit reconstructed-geometry tokens — fused with 2D features via a dedicated adapter. The RGB-only design avoids the need for depth sensors or point-cloud inputs while claiming state-of-the-art results across four disparate 3D tasks.

Key Results: The abstract claims VLM-IE3D achieves superior performance consistently across four 3D task categories: 3D video detection, 3D visual grounding, 3D dense captioning, and spatial reasoning. No specific numerical results, benchmark scores, or dataset names are cited in the abstract itself — quantitative claims are deferred to the paper body.

Key Findings:

Combining implicit and explicit geometric tokens outperforms either representation alone for 3D VLM tasks
Strong 3D inductive biases can be injected into VLMs using only RGB video, without depth or LiDAR inputs
A single unified architecture generalizes across 3D detection, grounding, dense captioning, and spatial reasoning

Technical Novelty: The dual representation of Implicit Geometry Tokens (high-level priors from raw video) alongside Explicit Geometry Tokens (fine-grained tokens from reconstructed 3D attributes), fused via a dedicated 3D-aware adapter — a departure from prior work that typically injects either depth features or point-cloud tokens alone, or requires actual 3D sensor inputs.

What's New: Prior 3D-aware VLMs typically rely on explicit 3D inputs (point clouds, depth maps) or single-modality geometric injection. VLM-IE3D is novel in unifying implicit priors with explicit reconstructions from RGB alone and fusing them via a purpose-built 3D adapter within a single framework spanning four downstream 3D tasks.

Extension Opportunities:

Extend the IGT/EGT dual-token design to dynamic 4D scene understanding (temporal 3D reasoning) for robotics manipulation and embodied navigation tasks
Replace the RGB-video reconstruction backbone with monocular foundation models like DUSt3R or VGGT to improve explicit geometry quality without added training cost
Distill the 3D-aware adapter into smaller open-source VLMs (e.g., Qwen2-VL-2B) to enable on-device spatial reasoning for AR/robotics applications

Replicability: Code and models are publicly released at github.com/Vegetebird/VLM-IE3D. Reproduction likely requires multi-GPU training (8+ A100/H100s) given the video-based 3D reconstruction pipeline plus VLM fine-tuning, though inference should be feasible on a single high-end GPU.

Research Gaps:

Abstract does not disclose performance on real-world embodied tasks (navigation, manipulation) where robustness of RGB-only 3D reconstruction is critical
Unclear how the approach handles dynamic scenes, occlusions, or textureless surfaces where video-based 3D reconstruction typically degrades

🔥 GitHub Trending

1. youben917/CUDA-MOEA

⭐ 12 stars | Jupyter Notebook

A full CUDA-Stack Infrastructure for Multiple/Many Objective Evolutionary ALgorithms

cuda eda evolutionary-algorithms gpu-accelerated moea

2. tonquoc0407/capybara

⭐ 8 stars | Go

Terminal trace debugger for AI agents.

agent-tools ai-agents cli-tool debugging golang llm

3. Pinvou/pinvou-agent

⭐ 7 stars | Rust

Open-source desktop AI agent for tools, files, knowledge, workflows, and real deliverables.

ai-agent automation desktop-app llm local-first mcp

4. Morteza-Asadi-Shalmaiy/Superres-fr

⭐ 6 stars | Jupyter Notebook

python notebook super resolution

ai artificial-intelligence basicsr cctv cctv-cameras cctv-detection

5. Morteza-Asadi-Shalmaiy/Reid-fr

⭐ 4 stars | Jupyter Notebook

python notebook ReId project

ai artificial-intelligence cctv cctv-cameras cctv-detection cctv-monitoring

6. olaflaitinen/laga

⭐ 3 stars | Python

A fast, conservative JSON repair library for malformed model output, hand-written config, and almost-JSON text. Repairs common syntax issues and returns clean Python objects.

ai cli configuration data-cleanup json json-parser

7. programmersd21/frame_studio

⭐ 2 stars | TypeScript

🏃 Frame Studio - AI-powered motion graphics generator. Write a prompt, get production-ready Remotion videos with Gemini. TypeScript-validated, direct MP4 download.

ai-video animation creative-tool framer-motion gemini-api generative-ai

8. OlehDatsyk/ai-assistant-platform

⭐ 2 stars | Python

An enterprise-grade AI assistant platform built with FastAPI, the OpenAI Responses API, Anthropic Claude, Google Gemini, LangChain, RAG, MCP, AI Agents, Voice AI, Vision AI, and automation integration

ai artificial-intelligence automation chatbot claude-code llm

9. nagameTW/humanizer-zh-tw

⭐ 2 stars | Python

AI 寫作痕跡清除技能，台灣在地強化版。把 AI 生成的文章改成道地的台灣中文，涵蓋 46 種模式，含中國用語偵測、中文標點保護、語域對照。是 Claude skill，也能用在任何支援 skill 格式的 agent。

agent-skills ai-detection ai-writing anthropic chatgpt claude-code

10. 2654400439/Paper2Seminar

⭐ 2 stars | Python

An Agent Skill for turning research papers into complete, editable, seminar-ready PPTX decks.

academic-presentation agent-skills ai claude-code codex llm

11. DeveloperXG/Portfolio

⭐ 1 stars | TypeScript

My interactive ML/AI engineering portfolio, built with Next.js, TypeScript, and a cinematic multiverse-inspired visual system.

artificial-intelligence framer-motion frontend machine-learning nextjs portfolio

12. AnamShaikh-code/Telecom-Customer-Churn-Prediction

⭐ 1 stars | Jupyter Notebook

End-to-End Machine Learning project for predicting telecom customer churn using Python, Scikit-learn and Power BI with business-driven insights.

classification customer-churn data-analysis machine-learning matplotlib pandas

13. Azimml/gptq-from-scratch

⭐ 1 stars | Python

From-scratch PyTorch implementation of GPTQ post-training quantization (2/3/4-bit) with multi-architecture support

gptq llm machine-learning python pytorch quantization

14. KhangNguyen08/House-Price-Prediction-Model

⭐ 1 stars | Jupyter Notebook

Ensemble solution (XGBoost, LightGBM, CatBoost, Extra Trees, Neural Net) for Kaggle House Prices, with a data-driven feature selection and target-encoding search process.

catboost ensemble-learning feature-engineering house-price-prediction kaggle-competition lightgbm

15. wangzifan396-wzf/transformer-viz-lab

⭐ 1 stars | HTML

Transformer 可视化实验室 · Transformer Architecture Visualizer — 纯 HTML/CSS/JS + SVG · 零外部依赖

attention deep-learning education html interactive machine-learning

Generated by Research Pulse on 2026-07-24 06:05