Skip to content
AI-generated research idea🟠Medium🔬Significant ResearchNo close match found

Temporal RLCF: Incorporating Citation Dynamics for Delayed-Impact Scientific Preference Modeling

A method that utilizes sequence modeling on yearly citation deltas to predict long-term impact, correcting reward model biases against delayed-impact scientific publications.

Time

2-4 weeks for longitudinal citation-delta curation, GRPO/sequence reward training across 4B-30B scales, and temporal holdout evaluation.

Compute level

Large cluster

Resources

32-128 H200-equivalent GPUs (4-16 nodes x 8 GPUs) with DeepSpeed ZeRO-2/ZeRO-3 and vLLM, plus annual citation snapshot collection across the 2.1M arXiv training pool (720,341 pairs) and sequence modeling compute.

What the paper used

H200-equivalent GPUs with DeepSpeed ZeRO-2/ZeRO-3 and vLLM, scaled by model size: 32 GPUs (4 nodes x 8 GPUs) for 1.5B, 64 GPUs (8 nodes x 8 GPUs) for 3B-7B, and 128 GPUs (16 nodes x 8 GPUs) for 14B-32B models; trained via GRPO (batch size 128, 1 epoch, 8 generations/prompt) on 720,341 pairs (from 2.1M arXiv papers) for SciJudge and 4,000 seed papers for SciThinker.

Prerequisites

Transformer architectures, reinforcement learning, time-series forecasting

Scope

modify existing code and large-scale data collection

The Gap

The current SciJudgeBench constructs pairwise preferences using static, finite-horizon cumulative citation counts. This overlooks 'sleeping beauties'—papers that initially receive few citations but become highly influential years later—causing the reward model to potentially penalize delayed-impact, highly novel research.

“Some high-potential papers may receive few citations initially but become highly influential later.”
From the paper, Section 6 (Limitations and Future Work), page 13. Matched against the PDF text.

Limitation

Citations are treated as static, aggregated values without capturing temporal dynamics, leading to an imperfect reflection of true, long-term community feedback.

Proposed Solution

Implement a time-series-aware reward model using dynamic RLCF. Instead of binary preferences based on static sums, use sequence modeling (e.g., LSTMs or time-aware Transformers) on the yearly citation deltas to predict the area under the citation curve or time-to-peak impact, explicitly adjusting for delayed-impact patterns.

The paper’s authors list this direction as future work, so they may already be pursuing it.

Why This Matters

This prevents the AI from becoming overly biased toward trend-chasing research and encourages the generation of paradigm-shifting ideas that take time to gain traction.

How to Get Started

How to test it: Test the hypothesis that training a citation-trajectory-aware reward model on yearly citation deltas improves long-term impact prediction over the static SciJudge-Qwen3-4B/30B baseline. Measure position-swap pairwise accuracy on a longitudinal holdout of delayed-impact papers and the 904-pair Temporal OOD test set, stopping if dynamic modeling fails to outperform the static baseline on delayed-impact pairs or degrades overall SciJudgeBench accuracy.

  1. Extract an extended temporal citation graph and publication metadata dataset from arXiv spanning multiple decades.
  2. Construct training targets representing the area under the citation curve or time-to-peak impact for each paper.
  3. Implement a time-series-aware reward model using an LSTM or time-aware Transformer architecture.
  4. Train and validate the dynamic reward model against static cumulative citation count baselines using correlation metrics.

Code linked to this paper on Hugging Face: github.com/tongjingqi/AI-Can-Learn-Scientific-Taste

Suggested by AI. Adapt it to your setup.

Has This Been Done?

No close match in 10 results

None of the listed papers address time-series citation dynamics or RLCF for delayed-impact modeling. The search results cover unrelated topics such as immunizations, bibliometric indicators, disease forecasting, and green AI.

We searched OpenAlex and arXiv for “A method that utilizes sequence modeling on yearly citation deltas to predict long-term impact, correcting reward model biases against delayed-impact scientific publications.” and “Temporal RLCF: Incorporating Citation Dynamics for Delayed-Impact Scientific Preference Modeling” and an AI compared the idea with the top 10 results. A small search can miss relevant work.

Source Paper

AI Can Learn Scientific Taste

Jingqi Tong, Mingzhe Li, Hangcheng Li +20 more · 2026

This paper explores whether AI models can develop 'scientific taste' to evaluate and propose high-impact research ideas. The authors propose Reinforcement Learning from Community Feedback (RLCF) to train models using large-scale citation data, successfully creating an AI that generates research proposals with higher potential impact than baseline models.

Related Ideas

Is this a good idea?

Share onXLinkedIn