EvoScientist Documentation

Research in retrieval and multimodal access is categorized into four main method families. Use this classification to identify the core mechanism and trade-offs of different retrieval approaches:

1. Dense retrieval supervision refinement: Focuses on improving supervision quality and semantic discrimination through hard-negative relabeling, latent reasoning, and semantically aware pooling.
2. Token-level efficiency and index compression: Focuses on making retrieval suitable for repeated calls in agent/RAG settings via token pruning, filtered ANN, and multi-vector compression.
3. Multimodal alignment and local grounding: Focuses on stronger local evidence modeling and multimodal interpretability through region-aware alignment, graph-text quality estimation, and listwise reranking.
4. Retrieval interfaces for downstream generation: Focuses on aligning retrieval with end-to-end task value (e.g., code generation, fake news detection, RAG) using task-specific retrievers.

EvoScientist provides a systematic survey of research trends based on arXiv papers. For the SIGIR 2026 context, the research landscape is categorized into five primary themes:

1. Recommendation & Advertising: Focuses on restructuring recommendation under LLM, generative modeling, and industrial deployment constraints.
2. Retrieval, Ranking & Multimodal Access: Focuses on improving quality, efficiency, and grounding across text, visual, code, and hybrid multimodal scenarios.
3. Search Agents, RAG & Reasoning: Focuses on multi-step trajectory planning, query refinement, reasoning routing, and uncertainty-aware retrieval.
4. Evaluation, Fairness & Alignment: Focuses on redefining evaluation, alignment, and fairness in the era of LLM-integrated information access.
5. Graph Structure, Representation & General IR: Focuses on general methodological significance outside the four frontiers.

EvoScientist is a self-evolving AI Scientist platform. Recent versions have introduced significant capabilities:

• v0.2.4: Support for Claude Opus 5 (Anthropic/OpenRouter), Gemini 3.6 Flash/3.5 Flash Lite, and Kimi K3 via Anthropic-protocol channels. Includes fixes for tool-call history and OpenRouter SSE streaming.
• v0.2.3: Kimi K3 support on Moonshot/OpenRouter, async sub-agent cleanup, Telegram slash commands, and OpenAI reasoning_effort support.
• v0.2.1: AutoSkills feature (EvoMemory drafts reusable skills via /autoskills) and a new --output-format stream-json for headless/SDK clients.
• v0.2.0: Scheduled tasks via /schedule, self-linking memory (knowledge graph), and a GET /api/models endpoint.
• v0.1.8: OpenRouter Anthropic prompt caching enabled by default (openrouter_anthropic_prompt_cache=false to opt out).
• v0.1.5: Dangerous mode (real-filesystem access with safety checks) and EvoSci deploy --tunnel for Cloudflare tunnels.
• v0.1.2: Browser WebUI mode and EvoSci deploy standalone LangGraph server.

Research in recommendation and advertising is categorized into four main subfamilies based on their core mechanisms and objectives:

1. Generative recommendation foundations: Focuses on item tokenization, differentiable semantic IDs, and joint optimization of generators and symbolic spaces to bridge item identity and autoregressive generation.
2. LLM-enhanced sequential / multimodal recommendation: Uses LLMs, adapters, multimodal reasoning, or intent abstraction to mitigate semantic poverty in interaction histories and address cold-start issues.
3. Alignment, control, and satisfaction modeling: Employs beam-aware training, test-time scaling, reward modeling, and reason-verify pipelines to align models with real decision objectives and reasoning reliability.
4. Unified modeling, systems, and industrial deployment: Focuses on CTR re-design, GPU systems, and shared search-recommendation parameter spaces to move toward deployable systems and unified task modeling.

Current research in retrieval and multimodal access has shifted from improving model capacity (encoders) to addressing three coupled sub-problems:

1. Supervision Quality: Moving beyond simple encoders to focus on clean, high-quality supervision signals (e.g., addressing hard negatives and training noise).
2. Index Efficiency: Treating token-level efficiency as an algorithmic necessity rather than just a deployment concern, especially for Agentic/RAG systems where retrievers are called repeatedly.
3. Evidence Grounding: Shifting from global embedding alignment to fine-grained, fragment-level grounding so that downstream systems (RAG, Agents) can locate and consume specific evidence.

Key research trends include:

• Dense Retrieval: Focus on answer-centric relabeling and latent reasoning.
• Token-level Efficiency: Using Voronoi pruning and multi-vector compression for RAG/Agent scenarios.
• Multimodal Access: Moving toward region-aware alignment and graph-text QA.
• Downstream-oriented Interfaces: Designing retrievers specifically for code generation, fake news detection, or RAG.

EvoScientist provides a comprehensive suite of features for autonomous scientific research:

• Multi-Agent Collaboration: 6 specialized sub-agents (Planning, Research, Coding, Debugging, Analysis, and Writing) work together.
• Self-Evolving Memory: Automatically refines and connects knowledge into a knowledge graph that evolves across sessions.
• AutoSkills: Periodically extracts repetitive patterns from memory and drafts them as reusable skills, available for user review via /autoskills.
• Multi-Model Support: Configure and switch between providers like Anthropic, OpenAI, Google, MiniMax, and NVIDIA.
• Multi-Channel Access: Centered around a CLI, but shares the same agent sessions across Telegram, Slack, Feishu, and WeChat.
• Desktop WebUI: Launch a web application with a workspace panel using the --ui webui flag.
• Scientific Workflow: Follows a structured process: Requirement Gathering → Planning → Execution → Evaluation → Writing → Verification.
• Scheduled Tasks: Supports cron-style scheduling for repetitive research tasks.
• Code Generation Mode: Includes a 'More Effort' mode for iterative refinement of code quality.
• Adaptive Tools: Dynamically filters relevant tools during each conversation to improve efficiency.
• Context Editing: Dynamically rewrites system prompts based on conversation state.
• MCP & Skills: Supports plug-and-play MCP (Model Context Protocol) servers and one-click installation of skill packs from GitHub.

Based on the SIGIR 2026 survey, research in Recommendation and Advertising is categorized into four primary methodological families. Understanding these helps in identifying the core mechanism and trade-offs of different approaches:

1. Generative Recommendation Foundation Layer: Focuses on item tokenization, differentiable semantic IDs, and joint optimization of generators and symbolic spaces. It aims to bridge item identity with autoregressive generation but suffers from training complexity.
2. LLM-Enhanced Sequential/Multimodal Recommendation: Uses LLMs, adapters, multimodal reasoning, or intent abstraction to strengthen state representations. This mitigates semantic poverty in interaction history but has high computational costs.
3. Alignment, Control, and Satisfaction Modeling: Employs beam-aware training, test-time scaling, reward modeling, and reason-verify mechanisms. It aligns closer to real decision-making goals but faces a gap between offline metrics and online value.
4. Unified Modeling, Systems, and Industrial Deployment: Focuses on CTR structure optimization, GPU systematic implementation, and unified parameter spaces for search and recommendation. It aims to move from 'trainable' to 'deployable' in industrial environments.

To use Discord, install the Discord extra. You must create an application in the Discord Developer Portal, enable Message Content Intent, and invite the bot to your server using an OAuth2 URL with bot scope and permissions for Send Messages, Read Message History, Attach Files, and Add Reactions.

Installation:

pip install evoscientist[discord]

Configuration: Set channel_enabled: "discord" and provide the discord_bot_token. You can restrict access to specific users via discord_allowed_senders or specific channels via discord_allowed_channels.

Environment Variables:

• EVOSCIENTIST_DISCORD_BOT_TOKEN
• EVOSCIENTIST_DISCORD_ALLOWED_SENDERS
• EVOSCIENTIST_DISCORD_ALLOWED_CHANNELS
• EVOSCIENTIST_DISCORD_PROXY

channel_enabled: "discord"
discord_bot_token: "MTIzNDU2Nzg5.xxxx.xxxxx"
discord_allowed_senders: ""
discord_allowed_channels: ""
discord_proxy: ""