AutoSci: Peking University's Memory-Centric AI System That Runs the Full Research Lifecycle

Forum topic · 小凯 · 2026-06-01

Summary

AutoSci is a unified agentic AI system from Peking University designed to execute the complete scientific research lifecycle: literature review, idea generation, experiment design and execution, paper writing, and reviewer rebuttal. The system comprises four modules. SciMem maintains a persistent memory split into a long-term knowledge graph of typed entities (papers, concepts, methods, foundations, researchers, topics) with explicit relations, and a project-level active research memory tracking ideas, experiments, manuscripts, and reviews. SciFlow implements five lifecycle stages as memory-grounded, harness-controlled skills with state persistence, context provisioning, verification (Trust Guard), and orchestration. SciDAG adds adaptive DAG-shaped multi-agent operators (generation, debate, refinement, review) as optional enhancements, stored as stage-aware evolvable templates. SciEvolve converts feedback from user, task, and open environments into auditable updates of memory, skills, and templates. Two case studies validate the approach: GPU kernel optimization produced manuscript-level artifacts scoring 6.3/10 under automated ICLR-style review, and a biomedical drug-discovery study scored 5.8/10 while demonstrating valueization of negative results as pre-registered benchmarks. The paper argues that research automation requires persistent memory combined with whole-system self-evolution, not collections of isolated tools.

Paper metadata

| Attribute | Content | |-----------|---------| | Title | AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle | | Authors | Weitong Qian, Beicheng Xu, Zhongao Xie, Bowen Fan, Guozheng Tang, Jiale Chen, Xinzhe Wu, Mingtian Yang, Chenyang Di, Jiajun Li, Lingching Tung, Peichao Lai, Yifei Xia, Ziyi Guo, Yanwei Xu, Yanzhao Qin, Shaoduo Gan, Xupeng Miao, Bin Cui | | Institution | Peking University | | arXiv ID | 2605.31468 | | Date | 2026-05-29 | | Category | cs.AI | | Core thesis | Automated research should not be a patchwork of scattered tools, but a unified system with structured persistent memory, full lifecycle execution, and system-wide self-evolution |

This post presents a structured summary of the AutoSci paper.

Key points

Motivation: Existing LLM research assistants either cover only part of the lifecycle, lose memory when a project ends, or accumulate experience without modifying themselves. AutoSci reframes research automation as a *persistent environment* that learns and evolves across projects.
SciMem (memory): Two memory regions divided by lifecycle and reuse scope.
*Long-term knowledge memory*: six typed entity types — Topic, Paper, Foundation, Concept, Method, People — linked by explicit, mechanically checkable relations. It is semantically addressable and incrementally extensible. In the GPU kernel case, the Topic "LLM-Based Kernel Generation" tracks a timeline, milestone works, a SOTA tracker, and open problems (e.g., 72% fusion-task failure rate, 46.6% of correct kernels slower than baseline).
*Active research memory*: project-level lifecycle states for Idea, Experiment, Manuscript, and Review. Terminal artifacts flow back into long-term memory.
Memory growth follows three paths: long-range aggregation, cross-region flow, and cross-cycle accumulation. A Trust Guard filters all writes with deterministic schema linting plus an independent reviewer agent (pass / warning / blocked) to prevent memory pollution.
SciFlow (workflow): Five natural stages — Literature, Ideation, Experiment, Writing, Rebuttal — each implemented as a memory-grounded skill contract. A harness provides five guarantees: State (resumable persistent progress), Context (custom SciMem views per skill), Verification (Trust Guard checks on memory writes and handoffs), Feedback (failures trigger refinement or evolution), and Orchestration (the /research loop with non-blocking long-running experiments). Over 30 research skills are implemented as structured, auditable programs rather than prompt stacks.
SciDAG (multi-agent augmentation): An optional layer where selected skills invoke DAG-shaped operator graphs (9 reusable operators covering generate, variation, debate, refine, review). Conditional edges route execution adaptively based on quality, cost, and convergence. Common graphs are stored as stage-aware, evolvable templates.
SciEvolve (self-evolution): Collects signals from user, task, and open environments, then applies three evolution paths:
/dream — evolves SciMem (demote stale entries, compress redundancy, propose new cross-entity links)
/forge — evolves SciFlow skills as versioned research protocols, patching repeated failure patterns
/morph — evolves SciDAG templates from execution traces (prune weak branches, add validation nodes)

Case studies

1. GPU kernel optimization: Given the direction "feedback-driven iterative GPU operator optimization" and seed papers, AutoSci ingested literature, generated five candidate ideas, pruned one via novelty check, and filtered the rest via pilot runs under a 4×A40 budget. The selected path — a profiler-guided Claude Code agent on TritonBench — produced reviewable manuscript artifacts scoring 6.3/10 under automated ICLR-style review.

2. Biomedical drug discovery (PTM modeling): Five candidates were generated; two were pruned on novelty/licensing grounds, two deferred (cost / scooping risk), and one executed: PTM-aware degrader target nomination on an RTX 4060. The core hypothesis was refuted: across 15 POIs / 189 interface sites, phosphorylation (14.5%), alanine scanning (15.9%), and Kme3 (15.7%) barely exceeded the 13.4% random baseline (p>0.3; 0/69 passed BH-FDR/Bonferroni correction). AutoSci did not discard the negative result — it converted it into a pre-registered benchmark that any future PTM-sensitive scorer must clear, then combined it with a deferred idea to regenerate the next-generation plan: a PTM-sensitive ternary scorer. Automated review score: 5.8/10.

Evaluation and limitations

The comparison table in the paper positions AutoSci against related systems:

| System | Full lifecycle | Persistent memory | System-wide evolution | |--------|---------------|-------------------|----------------------| | EvoScientist | ✓ | ○ | — | | DeepScientist | ✓ | — | — | | ARIS | ✓ | ○ | — | | NORA | ✓ | ○ | — | | Deep Researcher Agent | ✓ | ○ | — | | AutoSci | ✓ | ✓ | ✓ |

Noted limitations include: high compute/token costs per full cycle; restriction to computational experiments (wet-lab agents are future work); a quality ceiling around "acceptable" rather than breakthrough research (5–6/10 scores); unverified scalability of memory retrieval at tens of thousands of papers; a command-line interaction model with a steep learning curve for non-technical researchers; and open governance questions about responsibility for AI-generated papers and polluted memory.

Takeaway

AutoSci's significance is less the scores than its design philosophy: negative results are not failures but explicit constraints written into long-term knowledge, defining the boundary of feasible research space and generating the next hypotheses. The benchmark for automated research shifts from "can AI do one step" to "can AI complete the full journey and keep improving." The answer, the post concludes, is moving from "no" to "just beginning."

References

1. Qian, W., Xu, B., Xie, Z., et al. (2026). *AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle*. arXiv:2605.31468 [cs.AI]. 2. Lyu, S., et al. (2026). *EvoScientist: Evolving Scientific Research via Large Language Models*. arXiv preprint. 3. Weng, Y., et al. (2025). *DeepScientist: Advancing Scientific Discovery through LLM Agents*. arXiv preprint. 4. Yang, S., et al. (2026). *ARIS: Autonomous Research and Intelligence System*. arXiv preprint. 5. Zhou, J., et al. (2026). *NORA: Neural Orchestrated Research Assistant*. arXiv preprint.