acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

LLM-powered multi-agent systems can now automate the full research pipeline from ideation to paper writing, but a fundamental question remains: automation for whom? Researchers operate under different resource configurations, hold different methodological preferences, and target different output formats. A system that produces uniform outputs regardless of these differences will systematically under-serve every individual user, making personalization a precondition for research automation to be genuinely usable. However, achieving it requires three capabilities that current systems lack: accumulating reusable procedural knowledge across projects, retaining user-specific experience across sessions, and internalizing implicit preferences that resist explicit formalization. We propose NanoResearch, a multi-agent framework that addresses these gaps through tri-level co-evolution.A skill bank distills recurring operations into compact procedural rules reusable across projects. A memory module maintains user- and project-specific experience that grounds planning decisions in each user’s research history. A label-free policy learning converts free-form feedback into persistent parameter updates of the planner, reshaping subsequent coordination. These three layers co-evolve: reliable skills produce richer memory, richer memory informs better planning, and preference internalization continuously realigns the loop to each user. Extensive experiments demonstrate that NanoResearch delivers substantial gains over state-of-the-art AI research systems, and progressively refines itself to produce better research at lower cost over successive cycles.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Research Automation, Multi-Agent Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RpBrOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents NanoResearch, a multi-agent system for personalized research automation across idea generation, experimental implementation, and manuscript preparation. Its proposed tri-level co-evolution combines a skill bank of reusable procedural rules, memory of user- and project-specific experience, and a planner adapted to natural-language feedback through SDPO with user-specific LoRA updates. An orchestrator retrieves relevant skills and memories, coordinates specialized agents, and updates accumulated knowledge throughout the research workflow. The evaluation compares existing research systems on 20 tasks across seven domains and includes ten real research tasks supplied by PhD researchers. Reported validation covers execution success, task-native predictive performance, preference alignment, novelty, writing quality, component ablations, and changes in performance and cost over three rounds.

Critical strengths
  1. A substantial system contribution with a carefully controlled static comparison. NanoResearch connects literature retrieval, experiment planning, code execution and debugging, manuscript production, procedural skills, contextual memory, and an adaptable planner into one research workflow. Appendix C explicitly gives every system the complete researcher profile and matches model backends by shared role, inference settings, tools, and task budgets. Consequently, the Round-1 comparison in Table 1 provides meaningful evidence about orchestration: the reported benchmark E2E success is 1.000 versus at most 0.900 for the baselines, while objective task performance is 0.6844 versus the strongest baseline's 0.6238. The improvement is therefore broader than a subjective preference score. This is a useful empirical contribution even if the learning mechanisms require stronger attribution.

  2. The human evaluation goes beyond simulated users and separates preference satisfaction from independent quality assessment. Appendix N describes ten PhD researchers contributing ten frozen, real research tasks with heterogeneous constraints, including reproducibility, interpretability, compute efficiency, and patient-wise generalization. Task owners provide feedback and rate Alignment; three independent, domain-matched reviewers per task assess Novelty and Writing under anonymization, format normalization, and randomized presentation. Review assignments are correctly averaged within task rather than counted as independent experimental replicates. Table 1 reports an already strong human-study Round 1, followed by increases in Alignment from 8.1 to 9.1 and objective performance from 0.8012 to 0.8564. These design choices make the evidence more persuasive than an evaluation relying exclusively on an LLM feedback-and-judging loop.

Critical weaknesses
  1. The central ablation configurations and aggregation rules are unresolved. Table 2's “Nano w/o Planner Model” row is numerically identical on all five shared metrics to Table 6's “Planner Model Only” row, including E2E = 0.835 and Perf. = 0.642, although the labels describe opposite planner configurations. A labeling error could explain this, but the actual component masks and experimental conditions need to be identified before these results can establish component contributions. E2E is defined as a fraction of completed runs, yet values such as 0.849, 0.935, and 0.835 cannot arise from one binary outcome for each of twenty tasks. Repeated runs or another aggregation could explain them, but the success counts, denominators, and aggregation rules are not supplied. The full-system ablation row also reproduces Round-1 results without clearly specifying the adaptation state of each variant. These ambiguities are consequential because the ablations provide the main evidence for the proposed three-layer design; they prevent a reliable interpretation of its incremental benefits.

  2. The incremental contribution of persistent parameter learning is not adequately separated from feedback-conditioned revision. NanoResearch combines retrieved procedural skills, user/project memory, and a planner updated using existing SDPO. Its learning contribution therefore depends on establishing what the parameter updates and their interaction with retrieval add. Table 6's Planner+SDPO versus Planner Only comparison removes skill/memory retrieval, which is informative, but does not establish that feedback messages, retained context, execution history, and experiment budgets are matched. Appendix Q acknowledges the missing exact same-feedback frozen-planner control. The specific benefit of parameter persistence therefore remains uncertain. Component removal also does not by itself establish the proposed directional co-evolution between skills, memory, and policy. R1–R3 revisit the same task, and Appendix Q explicitly leaves cross-project transfer untested. The results provide useful evidence of feedback-driven project improvement, while the broader claims about persistent preferences and reusable cross-project knowledge need stronger support. This scope limitation compounds the unresolved ablation evidence; it would not alone determine my recommendation.

Questions for authors
  1. Clarification: What exact component masks, initialization states, and evaluated rounds generated Tables 2 and 6? Please resolve the identical “w/o Planner Model” and “Planner Model Only” results and identify which labels and values are correct. When, relative to each evaluated artifact, are feedback and SDPO updates applied? This determines what the planner and preference-alignment ablations measure.

  2. Clarification: For each ablation E2E value, what are the success count, number of runs, and aggregation rule? How are failed tasks included in Perf., and what are the paired task-level differences or uncertainty estimates for the principal component comparisons? These details are necessary to distinguish robust improvements from aggregation effects or stochastic variation.

  3. Clarification and validation of current claims: In Planner+SDPO versus Planner Only, are feedback messages, feedback retained in context, execution history, and inference/experiment budgets matched? Under otherwise matched feedback and research history, what improvement remains when only the learned adapter differs? This would directly affect my assessment of persistent parameter learning. Please also distinguish demonstrated within-task adaptation from the proposed cross-project reuse and preference persistence that Appendix Q leaves untested.

Reviewer_jJMUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NanoResearch, a personalized multi-agent framework for end-to-end research automation, covering idea generation, experimental implementation, and manuscript writing. The framework introduces three complementary mechanisms: a skill bank that accumulates reusable procedural knowledge, a memory module that retains user- and project-specific experiences, and an adaptive planner updated through self-distillation policy optimization using natural-language feedback. Compared with several existing automated research systems, NanoResearch reports improvements in execution success, task performance, preference alignment, and research output quality. The results also suggest improved efficiency across successive research rounds.

Critical strengths
  1. Relevant and well-motivated problem. Personalization is an important but relatively underexplored aspect of autonomous research systems.
  2. Comprehensive end-to-end system. The proposed architecture integrates ideation, experimentation, debugging, and paper writing into a unified workflow.
  3. Encouraging empirical results. The evaluation covers multiple research domains and includes comparisons with four existing systems.
Critical weaknesses
  1. One central claim is that SDPO internalizes user preferences into the planner parameters and provides persistent benefits. However, improvements across successive rounds may also arise from additional feedback, accumulated context, and repeated experimentation. It lacks a direct comparison against a frozen planner receiving the same feedback through prompting or memory.
  2. The paper emphasizes the accumulation of reusable procedural knowledge across research projects. However, the longitudinal experiments primarily evaluate repeated refinement of the same task. The growth of the skill bank and memory module does not directly demonstrate that knowledge learned from one project improves performance on a different project. Consequently, an important aspect of the claimed co-evolution remains insufficiently validated.
  3. Fairness of longitudinal comparisons and efficiency claims. The baseline systems are evaluated in a single round, whereas NanoResearch benefits from multiple rounds of feedback and refinement. The Round-3 results are not directly comparable to one-shot baselines under equivalent interaction budgets. Similarly, Table 3 reports decreasing per-round costs, but obtaining the Round-3 result requires completing the preceding rounds.
  4. Most simulated tasks are relatively constrained, incremental ML problems generated using Claude. Furthermore, the novelty evaluation primarily compares generated ideas with specified baselines, which may not adequately establish novelty relative to the broader scientific literature. A
  5. The paper reports aggregate performance without sufficient uncertainty estimates or detailed per-task variability. In addition, Table 2 reports end-to-end success rates such as 0.849 and 0.935, which cannot arise from a single binary outcome on each of the 20 benchmark tasks. The number of repeated trials and aggregation procedure should be clarified.
Questions for authors
  1. Is the SDPO parameter update necessary for personalization? Could the authors compare against a frozen planner that receives exactly the same user feedback through prompting and/or memory, with matched interaction and computation budgets?
  2. How well do the learned skills and preferences transfer across projects? For example, after completing one research task, does retaining the learned skill bank, memory module, or personalized planner improve performance on a different but related task?
  3. How should the reported efficiency gains be interpreted? Could the authors provide cumulative costs across all rounds, including planner fine-tuning, and compare against baselines given equivalent opportunities for iterative refinement?
  4. How robust are the experimental improvements? Could the authors report variance or confidence intervals across repeated runs, and provide more detailed per-task performance results?
  5. How is the scientific validity of generated research verified? Beyond execution success and LLM-based novelty assessments, are there independent evaluations of literature originality, experimental correctness, and whether the generated papers make scientifically well-supported contributions?
Reviewer_7DGfOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how an automated research system can adapt to a researcher’s preferences, resource constraints, and working history. The proposed NanoResearch system stores reusable procedures in a skill bank and user/project-specific experience in a memory module. An orchestrator retrieves relevant entries before each stage and updates the stores from the resulting execution record. The planner also learns from natural-language feedback through Self-Distillation Policy Optimization. A feedback-conditioned version of the model supplies token-level learning signals for the planner without that feedback in its input. The implementation updates user-specific LoRA adapters on Qwen3-8B-Instruct. Evaluation covers 20 simulated research tasks and ten tasks contributed by PhD researchers.

Critical strengths

The separation of skills, memory, and planner updates makes the design easy to follow. Each has a different job: retaining procedures, recording experience under particular conditions, or learning preferences in the planner’s parameters. Appendix M makes these distinctions concrete through examples of user profiles, retrieved records, and their effects on experiment design.

Critical weaknesses

The experiments do not isolate the benefit of updating the planner’s parameters from the benefit of using feedback in context. Planner+SDPO outperforms Planner Only without skill or memory retrieval, but that does not tell us whether a frozen planner given the same feedback would perform just as well. The comparison I would like to see is an equal-feedback frozen-planner baseline.

Several inconsistencies make the experimental results a bit difficult to interpret. Table 2 assigns the same five shared metrics to “Nano w/o Planner Model” that Table 6 assigns to “Planner Model Only.” These labels describe different interventions. Figure 3(b)’s human-study tasks also differ from those listed in Table 8.

Questions for authors

Why Table 2’s “without Planner Model” and Table 6’s “Planner Model Only” have identical results?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.