一键在 Manus 中运行任何 Skill

revere-reflective-research-agent

星标6

分支0

更新时间2026年3月26日 15:00

Enable LLM agents to autonomously improve on research-code tasks through reflective learning from execution trajectories. Distill recurring failure patterns into actionable heuristics applied via targeted prompt edits, improving performance 3.51%-4.89% while maintaining 10x better cost-efficiency.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

ADu2021

ADu2021/skillXiv

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Capability Gap Addressed

Existing prompt-optimization techniques fail at research-coding tasks because they rely on local signals (single-batch feedback) and full-prompt rewrites, causing poor generalization and knowledge loss. Research-coding workflows demand handling of heterogeneous repositories, underspecified environments, and weak feedback—conditions where static prompts and multi-agent systems struggle.

Core Abstractions

1. Execution Trajectory Reflection - Record failure patterns across repositories: dependency resolution issues, environment configuration errors, data formatting mismatches, incomplete library specifications.

2. Heuristic Distillation - Convert recurring failure patterns into reusable heuristics (e.g., "always check Python version compatibility before importing," "validate data encoding before processing").

3. Prompt Field Edits - Apply heuristics as surgical Python programs rather than full-prompt regeneration, targeting three configurable fields:

System prompt (agent role and capabilities)
Task-prompt template (structured task specification)
Cumulative cheatsheet (repository-specific and cross-repository remedies)

Design Decisions

Global Training Context - Maintains across adaptation iterations:

Reflection history (what failures were observed and how they were addressed)
Auxiliary context (environment specs, common libraries, known issues)
Cumulative cheatsheet (growing repository of solutions)

This enables learning beyond immediate batch feedback and knowledge retention across iterations.

Code-Based Field Updates - Generates Python programs that apply surgical edits to prompts rather than regenerating prompts entirely. Benefits:

Prevents semantic drift from full rewrites
Enables precise, targeted modifications
Preserves coherent system-state understanding

Unified Reflector Module - Single agent diagnoses failures and performs edits rather than splitting responsibilities across multiple agents. Advantages:

Maintains coherent understanding of system state
Reduces inter-agent communication overhead
Enables holistic learning across failure types

Self-Improvement Patterns

REVERE autonomously identifies and accumulates:

Infrastructure issues - Dependency resolution, environment setup, library version mismatches
Code-specific patterns - Common error patterns in particular codebases, optimization opportunities
Cross-repository remedies - General solutions applicable across multiple projects

The paper demonstrates that "agents equipped with mechanisms for continual learning...can meaningfully evolve their capabilities," with improvements persisting across new repositories through the evolving cheatsheet.

Integration Patterns

Continuous Adaptation Loop:

Execute agent on batch of research-code tasks
Record execution trajectories and identify failures
Reflector analyzes patterns and generates prompt edits
Apply edits via code-based modification
Re-execute on same tasks to validate improvements
Evolving cheatsheet transfers to new repositories

Generalization Mechanism - The cumulative cheatsheet acts as transferable knowledge, allowing improvements on Repo A to benefit performance on Repo B without explicit transfer learning.

Performance Profile

3.51%-4.89% improvement over expert-crafted baseline instructions
10x better cost-efficiency than alternative adaptation methods (fewer API calls, fewer full rewrites)
Improvements measured across three research-coding benchmarks
Knowledge retention: heuristics persist and apply to new repositories

同仓库更多 Skills

同仓库

meaningful-kebab-case-name

ADu2021/skillXiv

Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.

2026-03-266

action-quantization-behavior-cloning

ADu2021/skillXiv

Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.

2026-03-266

adaptive-lora-personalized-ranks

ADu2021/skillXiv

Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.

2026-03-266

additivellm2-domain-adaptation

ADu2021/skillXiv

Adapt general LLMs to specialized manufacturing domains via domain-adaptive pretraining on open-access journals and visual instruction tuning. Extract 50M tokens and 24K images from peer-reviewed papers, achieve >90% accuracy on domain knowledge tasks, and enable real-time defect identification from manufacturing images.

2026-03-266

agentic-ai-intelligence-explosion

ADu2021/skillXiv

Future intelligence explosions will be plural, social, and entangled with humanity through distributed collaborative systems rather than singular superintelligence. Intelligence is inherently social, demanding infrastructure matching agent development; integrate governance, institutional frameworks, and constitutional checks across hierarchies of autonomous agents and human-AI centaurs in shifting configurations.

2026-03-266

animalclap-taxonomy-aware-pretraining

ADu2021/skillXiv

Build taxonomy-aware audio-text pretraining systems for species recognition from animal vocalizations. Train contrastive models that augment text prompts with hierarchical taxonomic structure (scientific/common names, phylogenetic sequences), evaluate on unseen species via rare-species test sets, and predict ecological traits directly from audio.

2026-03-266

name	revere-reflective-research-agent
title	REVERE: Reflective Evolving Research Engineer for Scientific Workflows
version	0.0.3
engine	skillxiv-v0.0.3-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2603.20667
keywords	["Agentic AI","Self-Adaptation","Research Code","Prompt Optimization","Reflective Learning"]
description	Enable LLM agents to autonomously improve on research-code tasks through reflective learning from execution trajectories. Distill recurring failure patterns into actionable heuristics applied via targeted prompt edits, improving performance 3.51%-4.89% while maintaining 10x better cost-efficiency.

Capability Gap Addressed

Core Abstractions

3. Prompt Field Edits - Apply heuristics as surgical Python programs rather than full-prompt regeneration, targeting three configurable fields:

System prompt (agent role and capabilities)
Task-prompt template (structured task specification)
Cumulative cheatsheet (repository-specific and cross-repository remedies)

Design Decisions

Global Training Context - Maintains across adaptation iterations:

Reflection history (what failures were observed and how they were addressed)
Auxiliary context (environment specs, common libraries, known issues)
Cumulative cheatsheet (growing repository of solutions)

This enables learning beyond immediate batch feedback and knowledge retention across iterations.

Code-Based Field Updates - Generates Python programs that apply surgical edits to prompts rather than regenerating prompts entirely. Benefits:

Prevents semantic drift from full rewrites
Enables precise, targeted modifications
Preserves coherent system-state understanding

Unified Reflector Module - Single agent diagnoses failures and performs edits rather than splitting responsibilities across multiple agents. Advantages:

Maintains coherent understanding of system state
Reduces inter-agent communication overhead
Enables holistic learning across failure types

Self-Improvement Patterns

REVERE autonomously identifies and accumulates:

Infrastructure issues - Dependency resolution, environment setup, library version mismatches
Code-specific patterns - Common error patterns in particular codebases, optimization opportunities
Cross-repository remedies - General solutions applicable across multiple projects

Integration Patterns

Continuous Adaptation Loop:

Execute agent on batch of research-code tasks
Record execution trajectories and identify failures
Reflector analyzes patterns and generates prompt edits
Apply edits via code-based modification
Re-execute on same tasks to validate improvements
Evolving cheatsheet transfers to new repositories

Generalization Mechanism - The cumulative cheatsheet acts as transferable knowledge, allowing improvements on Repo A to benefit performance on Repo B without explicit transfer learning.

Performance Profile

3.51%-4.89% improvement over expert-crafted baseline instructions
10x better cost-efficiency than alternative adaptation methods (fewer API calls, fewer full rewrites)
Improvements measured across three research-coding benchmarks
Knowledge retention: heuristics persist and apply to new repositories