IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas

Ziyu Chen^C Yilun Zhao^Y Jiashuo Sun^I Yiling Ma^Y Manasi Patwardhan^T Arman Cohan^Y

^YYale University ^CThe University of Chicago ^IUniversity of Illinois Urbana-Champaign ^TTata Consultancy Services

Paper Code

Dataset BibTeX

anchors mined from published papers

checkable criteria

expert-refined ICLR 2026 test papers

0CSR

Qwen3.5-9B + RL, above GPT-5.5 (51.0)

Abstract

Scientific research often begins by synthesizing ideas from a set of related papers to identify gaps and formulate new directions. However, training language models to perform this form of literature-grounded ideation remains challenging, as existing prompting- or feedback-based approaches lack structured supervision for how papers should be synthesized. We introduce IdeaAnchor, a paradigm for training LLMs to perform research ideation, using structured specifications as privileged signals. Each IdeaAnchor instance encodes how each input paper should be synthesized into a successful idea, including their functional roles, relationships, and target synthesis criteria. We build this paradigm by mining instances from published papers, capturing how real ideas emerge from prior literature. We then train models via demonstration, self-distillation, and reinforcement learning, and further enhance generation with retrieval at inference time. Experiments show consistent improvements in ideation quality. Our analysis reveals a functional decomposition: anchor-based training strengthens creative synthesis, retrieval enhances detail elaboration, and combining both yields the best performance.

01What is an IdeaAnchor?

Given prior papers $\mathcal{P}$, a model writes a thinking trace, a research motivation and a research plan. An IdeaAnchor, mined from a real published paper, records how its prior works led to it, in three parts:

Anatomy of an IdeaAnchor

click a tab or a paper

An example anchor, mined from a paper on Byzantine-robust distributed learning.

Anchors are privileged information, available only during training.

02Mining anchors at scale

We reverse-engineer how published papers grew out of their prior work, using an LLM to build each anchor in three stages.

Mining pipeline

STAGE 1Extract prior worksFind the 5–7 papers that shaped paper $D$; tag each with a role and what $D$ learned from it.

STAGE 2EnrichFetch each prior work’s abstract from scholarly search.

STAGE 3Generate criteriaWrite checkable criteria for motivation, method and overall synthesis.

The creator LLM reads the full text of a published paper $D$.

Training corpus

14,183 anchors from ICLR / ICML / NeurIPS and Nature Communications papers, 2023–2025.

ICLR 2026 benchmark

one square per paper

Test set from ICLR 2026, after the training cutoff. Each instance is reviewed by two expert annotators.

03Training with anchors

The same anchor becomes three different learning signals. In every case the anchor is removed from the model’s input, so the model must internalize the synthesis.

Three ways to use an anchor

Highlighted elements show the privileged anchor signal.

For RL, a judge checks each proposal against the anchor’s criteria, and the reward is the fraction satisfied:

$$r_{\text{anchor}}=\frac{1}{|\mathcal{U}_D|}\sum_{u\in\mathcal{U}_D}\mathbb{I}\big[\theta_r(\mathcal{P},y,u)=\text{satisfied}\big]$$

Anchor reward in a GRPO group

pick a sample, toggle its criteria

0.00

Violates a guideline Format error (−1)

For illustration. Group size G = 8; advantage = (r − mean) / std.

04Inference: role-aware retrieval

The model first assigns each paper a role, then retrieves the full-text sections that role calls for. It can also start from a bare topic and search for papers itself.

Ideation pipeline

Algorithm 1 in the paper.

05Results

We report the criteria satisfaction rate (CSR): the share of expert-refined criteria a proposal satisfies, judged by GPT-5.4 on 924 ICLR 2026 papers.

Main results

Reference lines show proprietary models. Retrieval results are reported for Qwen3-8B.

Every paradigm improves both backbones; RL helps most, and on Qwen3.5-9B it surpasses GPT-5.5.

Robust to the choice of judge

CSR (%) of Qwen3-8B variants under three LLM judges. Rankings are consistent across all three.

Pairwise preference against the base model

Qwen3-8B variants vs. Qwen3-8B base. Hatched bars: GPT-5.4 judge. Solid bars: human experts.

Training dynamics and cross-domain transfer

CSR (%) of Qwen3-8B. Dashed lines: trained only on natural-science anchors, tested on the ML benchmark.

Anchors from natural science still improve ML ideation: what the model learns is domain-general.

06Training decides the direction, retrieval adds detail

Retrieval barely changes CSR, yet judges strongly prefer retrieval-augmented proposals. Retrieved passages add method details, but they rarely change the research direction.

Flip the two switches

Anchor RL training Retrieval (RAG-Full)

Schematic based on the paper’s analysis; the meters are qualitative. CSR values are for Qwen3-8B.

Retrieval vs. abstract-only

Win rate of each retrieval variant against the same model without retrieval (GPT-5.4 judge).

Starting from only a topic, the full system (RL + RAG-Full) searches papers, proposes candidates and ranks them. Three ML researchers rated the results blind.

From a topic to a proposal

Mean human ratings (1–5) over six open ML topics.

07Case study

Same seven input papers, two outputs. Hover a citation to see which paper it uses.

Object-centric world models

Excerpts from Appendix case study 1.

08Keeping the anchor hidden

In self-distillation the model sees the anchor, so it may simply quote it. We compared seven ways of presenting the anchor. Framing it as the model’s own persona knowledge gives zero leakage.

Anti-leak strategies

50 samples per strategy.

Citation

@article{chen2026ideaanchor,
  title   = {IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas},
  author  = {Chen, Ziyu and Zhao, Yilun and Sun, Jiashuo and Ma, Yiling and
             Patwardhan, Manasi and Cohan, Arman},
  year    = {2026}
}

Supported by the U.S. National Science Foundation (award No. 2541654) and Tata Consultancy Services.