AutoResearchClaw v0.5.0

Highlights

• Multi-Domain Architecture: Expanded beyond ML to support HEP Physics, Biology, Quantum Computing, and Statistics domains with profile-driven deployment
• ARC-Bench Evaluation Framework: Standardized benchmark suite with 50+ topics across 5 domains (ML01-ML25, P01-P10, Q01-Q10, B01-B07, S01-S03), rubric-based judging, and baseline adapters for AIDE, Agent Laboratory, and AI-Scientist-v2
• ColliderAgent Integration: Full HEP physics simulation pipeline support (MadGraph → Pythia → Delphes) with incremental experiment mode and Stage-12 re-entry
• Biology-Agent Integration: Metabolic modeling with COBRApy/Biopython skills, FBA simulation, and GSMM validation
• Quantum-Qiskit Skill: Qiskit-based quantum computing experiment support for quantum topics
• Statistics Domain Agent: Statistical method design, experiment evaluation, and theory analysis
• Requirements Gate: LLM capability validation before pipeline execution
• Profile-Driven Deployment: Interactive CLI for domain profile creation and management
• Incremental Experiment Mode: Resume experiments at Stage-12 with delta-prompt assembly
• Expanded Test Suite: New tests for HEP prompt hygiene, incremental experiments, and domain integrations

Breaking Changes

• Topic IDs renamed: T01-T25 → ML01-ML25 in ARC-Bench
• researchclaw/prompts.py refactored into researchclaw/prompts/ package (domain-aware prompt banks)

Documentation

• Domain Integration Guide for adding new scientific domains
• Tester guides in English, Chinese, and Japanese
• ARC-Bench experiment design docs and run guides
• Showcase papers demonstrating pipeline outputs

Full Changelog: v0.4.0...v0.5.0

v0.4.0 — Human-in-the-Loop Co-Pilot System

AutoResearchClaw is no longer purely autonomous. The new HITL Co-Pilot system transforms the pipeline into a human-AI collaborative research engine.

Highlights

• 6+ Intervention Modes: full-auto, gate-only, checkpoint, step-by-step, co-pilot, custom, express
• Idea Workshop: Brainstorm and refine hypotheses collaboratively (Stages 7-8)
• Baseline Navigator: Review and customize experiment designs (Stage 9)
• Paper Co-Writer: Section-by-section collaborative drafting (Stages 16-19)
• SmartPause: Confidence-driven dynamic intervention
• ALHF Intervention Learning: Learns from your review patterns
• Claim Verification: Inline fact-checking against collected literature
• Cost Guardrails: Budget monitoring with threshold alerts
• Pipeline Branching: Fork to explore multiple research directions
• CLI Commands: attach, status, approve, reject, guide
• 3 Adapters: CLI, WebSocket, MCP

New Files

• researchclaw/hitl/ — 34 modules (7,500+ lines)
• tests/test_hitl_*.py — 9 test files (242 tests)
• docs/HITL_GUIDE.md — 620-line guide
• 3 new builtin skills

Testing

• 2,753 tests passed, 0 failures

Full Changelog: v0.3.2...v0.4.0

What's New

Cross-Platform Support

• ACP-compatible agent backends: Claude Code, Codex CLI, Copilot CLI, Gemini CLI, Kimi CLI
• OpenClaw bridge: messaging platform integration (Discord, Telegram, Lark, WeChat)
• CLI-agent code generation backend: delegates Stages 10 & 13 to external CLI agents with budget control and timeout management

Anti-Fabrication System

• VerifiedRegistry: ground-truth whitelist from experiment results with tolerance matching
• Experiment diagnosis & repair loop: 13 deficiency categories, auto-repair with best-result selection
• Always-on sanitization: unverified numbers replaced in paper tables

Stability & Quality

• 100+ bug fixes across 8 deep audit rounds
• Modular executor refactoring (10K → 400-line facade)
• --resume auto-detection for interrupted runs
• LLM retry hardening with exponential backoff
• Community-reported fixes (macOS M3, math/theoretical topics)

New Subsystems

• Assessor (paper quality scoring + venue recommendation)
• Calendar (conference deadline tracking)
• Collaboration (multi-user research coordination)
• Copilot (interactive steering modes)
• Dashboard (real-time metrics broadcasting)
• Knowledge Graph (entity extraction + visualization)
• Memory (cross-run experiment/ideation/writing memory)
• MCP (Model Context Protocol server)
• Overleaf (live sync with conflict resolution)
• Project Manager (multi-project scheduling)
• Remote Servers (SSH/SLURM/cloud execution)
• Skills Library (12 built-in domain/tooling skills)
• Trends (daily arXiv digest + opportunity finder)
• Voice (speech-to-text commands)
• Wizard (guided project setup)

Testing

• 1,935 tests passing

Full Changelog: v0.3.1...v0.3.2

What's New

OpenCode Beast Mode

New "Beast Mode" routes complex code generation to OpenCode with automatic 6-signal complexity scoring and graceful fallback to CodeAgent.

Universal Cross-Domain Support

• Domain detector for 7 research domains (ML, physics, chemistry, economics, math, biology, security)
• 25+ domain-specific experiment profiles with tailored datasets, metrics, and evaluation protocols
• Domain-aware prompt adapters and Docker images

Code Searcher Agent

GitHub-integrated code search for experiment design reference — query generation, pattern extraction, and result caching.

Web Integration Layer

Web search, crawling, PDF extraction, and Google Scholar support for enhanced literature discovery.

Community Contributions

• Novita AI provider — added as built-in LLM provider preset (#80)
• Thread-safety hardening — all module-level globals protected with locks (#77)
• Figure agent config — fixed 6 missing config fields (#75)
• Robust LLM output parsing — 4-strategy JSON extractor, ACP-aware YAML extraction (#69)
• Test collection fix — safe skipif guard for Anthropic tests (#53)

Bug Fixes

• 20+ edge-case fixes from 3-round deep audit
• Fixed sleep-inside-lock contention in Semantic Scholar batch API
• Fixed false-positive regex on Markdown links in thinking-tag stripper
• Caught RecursionError in JSON parser for deeply nested payloads
• Convergence evaluator with early stopping recommendations

Full Changelog: v0.3.0...v0.3.1

What's New in v0.3.0

MetaClaw Cross-Run Learning Integration

• New researchclaw/metaclaw_bridge/ module: skill injection, lesson-to-skill conversion, PRM quality gates, session lifecycle management
• Pipeline failures → structured lessons → reusable skills, injected into all 23 stages
• +18.3% pipeline robustness in controlled experiments
• Opt-in via metaclaw_bridge.enabled: true, fully backward-compatible

CodeAgent v2 — Enhanced Code Generation

• Enhanced Blueprint: deep implementation specs with per-file pseudocode, tensor shapes, generation order
• Sequential File Generation: dependency-ordered with AST-based CodeMem
• Hard Validation Gates: block identical ablations, hardcoded metrics, cross-file import errors
• Targeted Error Repair: parse traceback to fix surgically instead of full regeneration

BenchmarkAgent & FigureAgent Improvements

• BenchmarkAgent: domain-aware benchmarks, import validation, pretrained resize
• FigureAgent: LLM output type safety, Paul Tol colorblind-safe palette, heatmap/ablation chart types
• visualize.py full rewrite: academic styling, 300 DPI, 6 enhanced chart types

50+ Pipeline Bug Fixes (BUG-06 through BUG-51)

• Metric direction, citation verify, CodeGen guard, condition drift, RL stability
• BST ordering, raw metrics, bracket citations, arXiv categories
• Docker HF permission, KD stability, ablation detection
• references.bib fallback generation, ICML LaTeX template fix

Paper Quality Hardening (4 Rounds)

• Post-compilation quality checks, weasel/duplicate word lint, NeurIPS checklist
• LaTeX escaping, 7-dim AI-Scientist-style review, AI-slop detection (50+ phrase blocklist)
• Related work depth checker, stats rigor validator, anti-boilerplate prompts
• Cross-discipline support for 7 domains

Docker Sandbox & Infrastructure

• Network-policy-aware code generation & sandbox execution
• Rate-limit defense for literature search APIs (OpenAlex → Semantic Scholar → arXiv cascade)

Full Changelog: v0.2.0...v0.3.0

Highlights

This release introduces three multi-agent subsystems, a hardened Docker sandbox, and 4 rounds of paper quality auditing — significantly improving the end-to-end quality of generated research papers.

New Multi-Agent Subsystems

CodeAgent (4-phase architecture)

• LLM generates multi-file experiment code (main.py + setup.py + requirements.txt)
• Static analysis & deep validation (AST-based class/method checks)
• LLM-guided code review with structured JSON feedback
• Iterative repair loop (up to 3 rounds) with automatic UnboundLocalError fix

BenchmarkAgent (4 sub-agents: Surveyor → Selector → Acquirer → Validator)

• Domain-aware dataset and baseline selection from 13-domain knowledge base
• Automatic benchmark acquisition with Docker compatibility validation
• Integrated at Stage 9 (experiment_design), output injected into Stage 10

FigureAgent (5 sub-agents: Planner → CodeGen → Renderer → Critic → Integrator)

• Academic-quality chart generation with SciencePlots, 300 DPI, colorblind-safe palette
• 6 built-in chart templates + LLM fallback for custom visualizations
• Tri-modal critic review (data accuracy, aesthetics, academic convention)

Docker Sandbox Enhancements

• Network-policy-aware code generation: none | setup_only | pip_only | full
• Dynamic dependency installation via requirements.txt
• Pre-cached datasets: CIFAR-10/100, MNIST, FashionMNIST, STL-10, SVHN
• Extended ML stack: torch, torchvision, timm, einops, transformers, etc.

Paper Quality Hardening (4-round audit)

• Post-compilation quality checks, weasel/duplicate word lint
• 7-dimension AI-Scientist-style review scoring
• AI-slop detection (50+ phrases), statistical rigor validator
• Cross-discipline support for 7 research domains (ML/physics/chem/econ/math/eng/bio)
• NeurIPS checklist integration

Bug Fixes (15+)

• Fix baselines dict-to-list crash in BenchmarkAgent
• Fix Gymnasium environment versions (v4 → v5)
• Fix experiment condition drift in iterative refinement (anchor to exp_plan.yaml)
• Fix compute budget constraint for experiment design
• Fix metric direction mismatch, citation verification batching
• Fix LaTeX output sanitization, figure plan format handling
• Add RL stability guidance (gradient clipping, NaN guard)
• And more — see full commit message for details

Compatibility

All changes are backward-compatible with v0.1.0 configuration files.

Full Changelog: v0.1.0...v0.2.0

AutoResearchClaw v0.1.0

Fully autonomous research pipeline: one message in, full conference paper out. 🦞

Highlights

• 23-stage pipeline: Research Scoping → Literature Discovery → Knowledge Synthesis → Hypothesis Generation → Experiment Design → Self-Healing Execution → Analysis & Decision → Paper Writing → Citation Verification
• Multi-agent debate: 3 agents (Innovator, Pragmatist, Contrarian) argue over hypotheses; adversarial analysis panel reviews results
• Self-healing executor: autonomous crash diagnosis, code repair, and Pivot/Refine decisions
• Cross-run evolution: time-decayed lesson store that improves future runs
• Citation verification: 4-layer pipeline (arXiv, DOI, Semantic Scholar, LLM relevance check)
• OpenClaw integration: trigger full runs from a chat message

Results (6 end-to-end runs)

• 100% pipeline completion (124/124 steps)
• 94.3% citation integrity
• Mean quality 6.2/10 on conference review scale

Requirements

• Python 3.9+
• OpenAI-compatible LLM API