PURSUhInT: In Search of Informative Hint Points Based on Layer Clustering for Knowledge Distillation
Expert Systems with Applications, vol. 213, 119040, 2023Journal article
IEEE BigData 2026
University of Maryland, Baltimore County · American University
In briefA variational framework that revises scientific ideas by preserving acceptance-worthy content and repairing rejection-worthy content on an idea graph, using peer-review feedback as weak supervision.
Large language models can generate plausible research ideas, but they lack a principled mechanism for deciding which parts of an idea should be preserved and which should be revised. We introduce GRIP (Graph-guided Review-Informed Preserve-and-repair), a variational framework that treats peer-review feedback as weak supervision over complementary latent factors: acceptance-worthy content to preserve and rejection-worthy content to repair. At the document level these factors are separate preserve and repair latents; at the idea-graph level each facet or claim node carries its own preserve-and-repair latents, enabling localised revision instead of undifferentiated rewriting. Positive reviewer feedback supervises preservation latents; negative feedback supervises repair latents; a graph-aware decoder rewrites only high-repair nodes while keeping the graph structure intact. An evidence lower bound aligns review-informed posteriors with idea-only priors, enabling revision without explicit feedback at inference time. A three-tier data representation—abstract-level corpus indexing, compact retrieval for generation, and review-grounded revision—keeps LLM cost bounded as the paper collection grows. We specify an experimental protocol with datasets, baselines, and preserve–repair metrics; empirical results are left to future instantiation of the released pipeline. Compared with checklist-grounded RL, graph-based edit selection, and facet-level interactive refinement, GRIP is the first formulation to cast scientific idea revision as dual-latent variational inference with explicit preserve–repair semantics on an idea graph.
Abstract from the authors’ manuscript. The paper was accepted at IEEE BigData 2026 in Phoenix, AZ. The proceedings record is pending.
Document- and node-level latents separate acceptance-worthy content to keep from rejection-worthy content to fix.
A graph-aware decoder rewrites high-repair facet nodes while copying high-preservation nodes, instead of regenerating the full proposal.
Training distils review supervision into idea-only priors so the model can revise without explicit reviewer input at test time.
Abstract-level indexing, compact retrieval, and review-grounded revision keep LLM cost bounded as the paper collection grows.
@inproceedings{ayanzadeh2026grip,
title = {{GRIP}: Graph-Guided Review-Informed Preserve-and-Repair for Scientific Idea Revision},
author = {Ayanzadeh, Aydin and Mousavi, Ahmad},
booktitle = {2026 IEEE International Conference on Big Data (BigData)},
year = {2026},
note = {Accepted},
address = {Phoenix, AZ, USA}
}