Skip to content

NanoResearch: Co-Evolving Skills, Memory, and Policy for Personalized Research Automation

🕒 Published (v1): 2026-05-11 00:00 UTC · Source: HuggingFace · link

Why this paper was selected

Co-evolving skills, memory, and policy for personalized multi-agent research automation

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

NanoResearch is a multi-agent research-automation framework that addresses the personalization gap in existing end-to-end AI scientist systems through tri-level co-evolution of a skill bank, a memory module, and a trainable planner policy. Unlike static pipelines, it accumulates reusable procedural knowledge, retains user-specific session history, and internalizes implicit feedback via parameter updates, growing more effective over successive research cycles.

Problem

Existing end-to-end research automation systems (The AI Scientist, EvoScientist, AI-Researcher, DeepScientist) are one-size-fits-all pipelines that discard execution knowledge between runs, lose user-specific context across sessions, and cannot convert diffuse natural-language preferences into persistent behavioral changes. Each run re-encounters the same debugging patterns, re-derives the same configurations, and ignores heterogeneous user constraints such as compute budgets, methodological preferences, and target venue styles.

Method

NanoResearch wraps a three-stage pipeline (ideation → experimentation → writing) inside a co-evolutionary harness driven by an Orchestrator \(\mathcal{O}\):

Skill Bank \(\mathcal{S}\): After each stage, \(\mathcal{O}\) reflects over the trajectory \(\tau\) and distills generalizable procedural rules (debugging patterns, configuration strategies) via \(\mathcal{S}^{(t+1)} = \mathcal{S}^{(t)} \cup \text{Distill}_\text{skill}(\tau)\). Semantically overlapping entries are merged to prevent unbounded growth. Retrieval uses a heuristic score combining keyword matching, tag alignment, and recency, weighted toward usage frequency and confidence for skills.

Memory Module \(\mathcal{M}\): Project-specific experiences (failed hypotheses, validated configs, resource constraints) are stored via \(\mathcal{M}^{(t+1)} = \mathcal{M}^{(t)} \cup \text{Summarize}_\text{mem}(\tau)\), with retrieval enforcing strict condition matching to surface only comparable prior outcomes. A persistent user profile \(U\) (queried interactively at session start) grounds all planning decisions.

Label-Free Policy Learning: Free-form user feedback \(F\) after each stage is encoded directly into the Orchestrator's planner model \(\pi_\theta\) (Qwen3-8B) via Self-Distillation Policy Optimization (SDPO). SDPO avoids reward models by using the feedback-conditioned model \(\pi_\theta(\cdot \mid x, F, y_{<t})\) as a self-teacher. The token-level advantage is: $\(A_t^\text{SDPO}(\hat{y}_t) = \log \frac{\pi_\theta(\hat{y}_t \mid x, F, y_{<t})}{\pi_\theta(\hat{y}_t \mid x, y_{<t})}\)$ and the gradient updates the student \(\pi_\theta(\cdot \mid x, y_{<t})\) to match the teacher distribution, internalizing preferences into parameters rather than into \(\mathcal{S}\) or \(\mathcal{M}\) where they risk compression or retrieval misses.

Before each stage, the Orchestrator retrieves top-\(k\) skills and memories: \(\mathcal{S}_C = \text{top-}k_{s \in \mathcal{S}} \text{score}(C, s)\), produces a stage-specific plan, dispatches specialized agents, then updates both stores from the resulting trajectory.

Key Contributions

  • Tri-level co-evolution architecture: skill bank, memory module, and planner policy co-evolve such that each component strengthens the others across cycles.
  • SDPO-based preference internalization: converts single-instance free-form feedback into dense token-level parameter updates without reward annotation or preference pairs.
  • Dual-store retrieval with differentiated weighting: skill retrieval prioritizes frequency/confidence; memory retrieval enforces condition matching—preventing cross-contamination of procedural vs. experiential knowledge.
  • Peer-review-like blueprint correction loop: an internal LLM reviewer critiques the JSON experiment blueprint iteratively before any code is written, catching infeasible designs early (Eq. 3).
  • Persistent user profile \(U\): constructed via interactive queries at session start; conditions all planning, retrieval, and agent dispatch throughout the pipeline.
  • Benchmark of 20 personalized research tasks spanning 7 domains, with Claude-simulated scientists providing persona-consistent multi-round feedback.

Results

  • NanoResearch Round 1 vs. best baseline (AI Scientist-v2 on Compliance): 8.163 vs. 6.656 Alignment score.
  • End-to-end success rate: 100% for all three NanoResearch rounds vs. 50–90% for baselines (AI-Researcher/AI Scientist-v2: 50%, EvoScientist: 75%, DeepScientist: 90%).
  • Performance (task accuracy): 0.6844 → 0.7320 → 0.7548 across Rounds 1–3 vs. best baseline 0.6238 (AI Scientist-v2).
  • Innovation score improves monotonically: 4.960 → 5.164 → 5.645 across rounds.
  • Writing quality: 5.428 → 5.859 → 6.172 across rounds.
  • Ablation: removing the Planner Model causes the largest single-component drop across all metrics; removing the Skill Bank most severely hurts Novelty (4.960 → 3.773) and Writing (5.42 → 4.75).
  • All three co-evolution components are individually necessary; no single-store variant (Memory Only, Skill Bank Only, Planner+SDPO) matches the full system.

Limitations

  • Evaluation relies entirely on LLM-simulated scientists (Claude) as both benchmark constructors and feedback providers—circular use of the same model family that powers parts of the pipeline.
  • Only 20 research tasks across 7 domains; statistical power for per-domain conclusions is limited.
  • The planner (Qwen3-8B) is the sole trainable component; all other agents are frozen commercial APIs, making the system sensitive to API availability and pricing changes.
  • SDPO is applied after each stage with single feedback instances; convergence properties with noisy or conflicting user feedback are not analyzed.
  • The autonomous debugging loop has a fixed retry budget; failure modes when that budget is exhausted are not characterized beyond noting they degrade E2E rate.
  • No wall-clock or token-cost breakdown is reported despite claiming "better research at lower cost."

Relevance to Harnesses / Meta-Harnesses

NanoResearch is a direct instance of a meta-harness: an outer control layer that wraps and persists state across multiple inner agent pipelines (ideation, coding, writing), updating shared stores (\(\mathcal{S}\), \(\mathcal{M}\), \(\pi_\theta\)) from each run's trajectory. Its tri-level co-evolution pattern—where the harness itself learns from execution traces to improve future dispatches—represents an advance over static orchestrators that merely sequence agents without accumulating cross-run knowledge. The differentiated dual-store design (procedural skills vs. episodic memory, with separate retrieval heuristics) is a reusable architectural primitive relevant to any harness that must balance generalizable tool knowledge against session-specific context. The SDPO-based policy update is particularly relevant to harness designers seeking to internalize operator preferences without requiring labeled preference datasets.