IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas

1 Introduction

• The paper frames ideation as creative synthesis—finding gaps and directions—and detail elaboration—developing actionable methodological specifics from prior work.A strong idea requires both a literature-grounded direction and enough technical depth to be actionable.Source: §1 · p. 2
• Existing approaches leave a gap: they may assume a research goal, require domain-specific execution environments, or omit multi-paper synthesis.Given related papers, models may produce fluent overviews or shallow blends rather than meaningful gaps and concrete, literature-grounded methods.Source: §1 · p. 2
• IdeaAnchor encodes each input paper’s role, the gaps to identify, and synthesis patterns, supporting demonstration, self-distillation, and reinforcement-learning training.The framework also uses retrieval at inference time to augment research ideation.Source: §1 · pp. 2–3
• The authors mine published papers’ intellectual genealogies to create structured training instances and build an expert-annotated benchmark for literature-grounded ideation.The benchmark evaluates whether models can identify non-obvious gaps and formulate novel directions from multiple papers.Source: §1 · pp. 2–3
• Training improves ideation quality, with reinforcement learning producing the largest gains in creative synthesis; retrieval mainly enhances detailed content elaboration.The analysis separates gap recognition and direction formulation, which need to be internalized through training, from grounding ideas in methodological details, which retrieval can enhance.Source: §1 · p. 3

2.1 Literature-Conditional Research Ideation

• The ideation task takes related papers as input and produces a thinking trace, a literature-derived motivation, and a plan grounded in those papers.The motivation is discovered rather than supplied as input, and the plan must draw on the provided papers.Source: §2.1 · p. 3

2 IdeaAnchor

• IdeaAnchor is a structured specification derived from a ground-truth paper and its input literature.The section presents these instance-specific specifications as the backbone of the training framework.Source: §2.2 · p. 3, §2 · p. 3

2.2 IdeaAnchor Specification

• Each anchor combines functional roles for prior works, analysis of how their relationships lead to a contribution, and checkable criteria for a successful synthesis.The criteria cover motivation, methods, and overall coherence; anchors are instance-specific and grounded in how real papers led to a published contribution.Source: §2.2 · pp. 3–4

2.3 Mining IdeaAnchor from Research Papers

• The mining pipeline reconstructs published papers’ intellectual genealogies, and the authors validate the mined prior works against author accounts.It extracts 5–7 influential prior works, enriches them with abstracts, and generates criteria across motivation, method, and overall dimensions. The corpus covers machine learning and natural science; the evaluation set comprises 924 ICLR 2026 papers with expert-reviewed prior works and criteria. Of 20 responding authors, 16 were satisfied with the mined sets; authors credited them with 93.7% of the works cited for their ideas.Source: §2.3 · pp. 4–5

3 Boosting Research Ideation with IdeaAnchor

• Training uses anchors as prompting context, privileged input, and reward signal, with retrieval and topic-driven literature discovery added at inference.Source: §3 · p. 5, §3.1 · p. 5, §3.2 · p. 6

3.1 Anchor-Driven Training

• Anchor-guided demonstrations embed anchor criteria in teacher outputs, while self-distillation trains on the model’s own anchor-conditioned outputs after removing the anchor.Source: §3.1 · p. 5
• Anchor-privileged reinforcement learning scores checkable, instance-specific anchor items and updates the policy toward higher-scoring candidates, with general quality guidelines also enforced.Source: §3.1 · p. 5

3.2 Inference-Time Enhancements

• Training supports creative synthesis, gap recognition, and direction formulation, but proposals still lack specific detail for actionable plans.Role-aware retrieval assigns each input paper a functional role and fetches corresponding full-text passages, such as methods, results, or problem statements.Source: §3.2 · p. 6
• For a high-level topic, the model can retrieve candidate papers and apply the same synthesis pipeline, optionally combined with full-text retrieval.Source: §3.2 · p. 6

4.1 Experimental Setup

• Training instances pair 5 to 7 prior papers with privileged IdeaAnchor specifications, and evaluation combines rubric-based GPT-5.4 CSR with blind expert comparisons.The policies start from Qwen3-8B or Qwen3.5-9B; experts assess grounding, novelty, specificity, and feasibility.Source: §4.1 · p. 6
• Inference tests abstract-only generation against RAG-Full, which retrieves role-relevant passages, and RAG-Summary, which compresses retrieved evidence into targeted summaries.RAG-Summary keeps input length and style closer to training while adding task-directed methodological information.Source: §4.1 · p. 6

4.3 Analysis and Case Studies

• Natural-science-only anchor training improves all three paradigms on the ICLR benchmark, supporting transfer of a general gap-identification and cross-paper synthesis procedure.In-domain performance improves with training scale; self-distillation gains are smaller but consistent.Source: Figure 5 · p. 7, §4.3 · p. 8

4.2 Main Results

• All three training paradigms exceed 75% pairwise win rate over Qwen3-8B base across judge configurations, with RL leading.The ranking captures holistic synthesis quality, complementing criterion-by-criterion CSR.Source: §4.2 · p. 7, Figure 4 · p. 7
• 54.2% CSR for Qwen3.5-9B RL versus 31.1% base; Qwen3-8B RL reaches 24.6% (+13.7), its best trained variant.Every training paradigm improves its base model; RL has the largest gain at both scales and surpasses proprietary references at Qwen3.5-9B.Source: Figure 3 · p. 7, §4.2 · p. 7

4.3 Analysis and Case Studies

• Retrieval improves preference more than CSR by adding operational detail, while rarely changing the synthesized direction.It can hurt the base model when unreliable role assignment amplifies weak premises; RAG-Full slightly outperforms RAG-Summary for the strongest RL policy.Source: §4.3 · pp. 7–8, Table 1 · p. 7
• In a six-topic open-ML study, the full system improves literature grounding and methodological specificity while maintaining feasibility; weaker proposals combine relevant papers without a sharp gap.The workflow retrieves recent papers, selects 6 to 8 per topic, then uses RL with RAG-Full and a pairwise ranker to choose a proposal.Source: §4.3 · p. 8, Table 2 · p. 8

AI for Scientific Discovery

• Scientific LLMs assist literature synthesis [Skarlinski et al., 2024, Zhao et al., 2025], code generation [Wijk et al., 2024, Nathani et al., 2025], and peer review [D’Arcy et al., 2024, Liang et al., 2024].Source: §5 · p. 8
• Si et al. [2024] found generated ideas novel but weaker on feasibility; later systems add review, retrieval, plan search, and idea-facet recombination [Baek et al., 2025, Wang et al., 2024, Hu et al., 2024, Pu et al., 2025].Source: §5 · p. 8
• Most ideation methods rely on prompting or single-paper inference rather than explicit training for multi-paper synthesis.Source: §5 · p. 9

Training with Privileged Information

• Privileged-information training includes rubric-based self-grading and paper-derived SFT/RL for research-plan and idea generators [Goel et al., 2025, Li et al., 2024].Source: §5 · p. 9

6 Conclusion and Discussion

• IdeaAnchor mines instance-specific specifications from published papers for privileged training, while analysis separates creative synthesis from retrieval-enhanced detail elaboration.The authors identify gap recognition and direction formulation as capabilities to internalize through training, and methodological detail as a target for retrieval.Source: §6 · p. 9
• Future work includes selecting and ranking candidate gaps, evaluating diverse ideas from shared prior works, and validating proposals through execution where feasible.The evaluation focuses on written proposals; execution-level validation may not be feasible in every domain.Source: §6 · p. 9

A.1 Stage 1: Prior Work Extraction & Role Assignment

• Stage 1 extracts a paper’s core idea, selects 5–7 proximal prior works using counterfactual, specificity, and proximity tests, and assigns functional roles.The selection process favors recent works and distinguishes direct predecessors, inspiration sources, gap-and-motivation papers, and methodological ingredients.Source: §A.1 · p. 14

A.2 Stage 2: Synthesis Narrative & Reference Proposal

• Stage 2 reconstructs reasoning from prior works to a new idea in a synthesis narrative, then develops a reference proposal with motivation and method.The proposal uses an overview and numbered method components that explain their roles and connections to prior work.Source: §A.2 · p. 15

A.3 Stage 3: Checkable Criteria Generation

• Stage 3 generates pass/fail criteria for motivation, method, and overall quality that assess engagement with prior-work gaps without requiring one specific solution.The criteria are designed to admit multiple valid proposals rather than demand reproduction of the reference proposal’s method.Source: §A.3 · p. 16

B Prompts for Training Data Synthesis

• Appendix B describes prompts for the three anchor-driven training paradigms, with SFT and self-distillation sharing a proposal-generation prompt.The shared prompt conditions on input papers and IdeaAnchor and produces a thinking trace, motivation, and research plan.Source: B Prompts for Training Data Synthesis · p. 16, §B.1 · p. 16

B.1 Demonstration Generation Prompt (Shared by SFT and Self-Distillation)

• SFT uses an external teacher to generate demonstrations, whereas self-distillation uses the target model to produce them.The self-distillation prompt adds an anti-leakage persona to prevent privileged anchor content from appearing in the proposal or thinking trace.Source: §B.1 · p. 17
• For self-distillation, anchor criteria are rewritten as first-person observations and normalized to avoid verbatim disclosure of privileged guidance.Motivation and method criteria are transformed into prose-like observations rather than shown as explicit rubrics.Source: §B.1 · p. 17

B.2 RL Reward Judging Prompt

• The RL judge scores each generated proposal against anchor criteria item by item, returning hit-or-miss judgments and a total.The prompt instructs the judge to assess whether and how each desideratum is met, rather than accepting a proposal’s claim that it satisfies one.Source: §B.2 · pp. 17–18, §C.1 · p. 19

C.1 Criteria-Based Evaluation Prompt

• Criteria-based evaluation rewards concrete technical reasoning and penalizes keyword lists that do not explain techniques’ roles or connections.A proposal can satisfy a criterion through a different valid approach if it explains the mechanism and reasoning clearly.Source: §C.1 · pp. 18–19

D Annotation Protocol and Inter-Annotator Agreement

• The ICLR 2026 benchmark’s candidate prior works, role assignments, and criteria are reviewed by two experienced annotators, with unresolved disagreements adjudicated by a senior researcher.Reviewers check prior-work relevance, adjust roles, and edit criteria for clarity and faithful representation.Source: D Annotation Protocol and… · p. 19

E Experimental Configuration

• Experiments use a node with 4 × NVIDIA B200 GPUs.Source: E Experimental Configuration · p. 19
• SFT and self-distillation share a fine-tuning configuration because they differ only in the model generating training data.Table 3 summarizes the shared hyperparameters.Source: E Experimental Configuration · p. 19
• RL rollouts use temperature 0.7, top-k 50, top-p 0.9, and a 2048-token maximum; SFT/self-distillation evaluation uses greedy decoding and at most 64 new tokens.Source: E Experimental Configuration · p. 19

F Training and Evaluation Data Statistics

• The training corpus contains 14,183 instances, each with 5–8 input prior works and 6–7 checkable criteria.Table 5 summarizes training-corpus statistics.Source: F Training and Evaluation Data… · p. 19, Table 5 · p. 20
• Each ICLR 2026 evaluation-benchmark instance has 7–8 expert-refined criteria across three dimensions.Table 6 provides detailed benchmark statistics.Source: F Training and Evaluation Data… · p. 20, Table 6 · p. 20
• Across the two training domains, motivation and method criteria are balanced, with exactly one overall criterion per instance on average.Table 7 compares criteria-category distributions across the domains.Source: F Training and Evaluation Data… · p. 20

• Annotators compare two proposals for the same input papers, with randomized order to reduce position bias.Source: G Human Evaluation Details · p. 20
• The five dimensions are overall quality, literature grounding, synthetic novelty, methodological specificity, and feasibility.Source: G Human Evaluation Details · pp. 20–21
• The accompanying tables describe criteria distributions across training domains and counts of instances by rubric size.Source: Table 7 · p. 21, Table 8 · p. 21
• For each dimension, annotators choose whether Proposal A or B is better, or select a tie.Source: G Human Evaluation Details · p. 21

• Anchor leakage risks training–inference mismatch; detection combines phrase matching, tag validation, manual inspection, leakage rate, rubric compliance, and ROUGE-L.Source: H Self-Distillation Data Leakage… · pp. 21–22
• Six anti-leak strategies vary how anchor information is presented, including weighted rubric emphasis without an explicit content signal.Source: H Self-Distillation Data Leakage… · p. 21
• 100% of naive-baseline outputs leaked; persona knowledge achieved 0% leakage, 1.000 rubric compliance, and 0.193 ROUGE-L.Source: Table 9 · p. 22, H Self-Distillation Data Leakage… · p. 22
• The detected 4% leakage for critique–edit and uncertainty steering was attributed to false positives; five of six strategies had effectively 0–2% true leakage.Source: H Self-Distillation Data Leakage… · p. 22
• The near-zero-leakage frontier separates priorities: quality_chars has the best ROUGE-1 at 0.543, persona_knowledge has 0% leakage and 0 tag errors, and critique_edit has perfect compliance with self-contained reasoning.Source: H Self-Distillation Data Leakage… · p. 22

I Additional Judge Evaluation

• CSR is reevaluated with alternative judges to test dependence on a particular LLM judge; Table 10 covers 924 ICLR 2026 instances in the abstract-only setting.Source: I Additional Judge Evaluation · p. 22, Table 10 · p. 22

J Case Studies

• In qualitative comparisons, SFT produces structured traces that synthesize across papers using citation handles, while Qwen3-8B Base tends toward generic, list-based proposals.Source: J Case Studies · p. 23

J.1 Case Study 1: Object-Centric World Models with Latent Actions

• The initial UA-OCWM proposal combines object-centric representations with temporal uncertainty, latent action control, and environmental adaptability.Source: §J.1 · pp. 23–24
• Unlike the base model's broad synthesis, SFT identifies missing entity-aligned latent actions and proposes particle-space actions to generate controllable futures.Source: §J.1 · pp. 24–25

J.2 Case Study 2: GFlowNet-Based Formulaic Alpha Factor Mining

• For alpha mining, the stated inefficiency is that RL evaluates completed factor sequences, while factor quality also depends on structural dependencies and composition.Source: §J.2 · pp. 26–27
• The proposal samples factor sequences with terminal probabilities proportional to portfolio return and uses discounted immediate returns to provide earlier training signals.Source: §J.2 · p. 27

J.3 Case Study 3: Safety-Aware Low-Rank Adaptation

• ID-SaLoRA targets global safety constraints that may sacrifice efficiency or performance by selectively constraining an intrinsic low-dimensional subspace.The proposal identifies that subspace with SVD or random projection, applies SaLoRA-style constraints there, and uses low-rank adaptation; quantization is suggested to reduce memory further.Source: §J.3 · pp. 28–29
• The motivation for adaptive constraints is that fixed safety projections can suppress task learning, whereas task-specific estimates may preserve task-relevant update directions.The discussion frames the goal as suppressing harmful directions while leaving task-relevant directions available for learning.Source: §J.3 · p. 29
• Safety LoRA estimates safety directions from the model’s safety-aligned outputs and constrains downstream LoRA updates to their null space.The proposed method uses SVD to identify dominant directions and removes safety-irrelevant directions, preserving low-rank adaptation efficiency.Source: §J.3 · p. 29
• The SFT model traces a specific design path and four coupled steps, while the base model proposes ID-SaLoRA at a higher level of abstraction.The SFT model identifies fixed safety projections as over-regularizing; its method estimates a safety subspace from model outputs, extracts dominant directions via SVD, constrains LoRA updates to their null space, and trains under that constraint.Source: §J.3 · p. 30