@inproceedings{gao-etal-2026-surprise,
title = "When Surprise Is Not Enough: Validation-Aware Rewards for {LLM}-Guided Scientific Discovery",
author = "Gao, Yifan and
Li, Haoyue and
Yuan, Feng and
Gao, Xin",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.245/",
pages = "5596--5611",
ISBN = "979-8-89176-459-0",
abstract = "LLM-based scientific exploration agents need feedback signals that decide which experimental outcomes should guide future search. Recent systems use Bayesian surprise, rewarding results that change an LLM{'}s belief. This can create a noisy-TV failure mode: unstable artifacts, leakage, proxy effects, or gaps in the model{'}s prior may appear more valuable than reliable mechanisms. We introduce Reducible Information Gain (ReIG), a validation-aware discovery method. ReIG keeps LLMs for hypothesis proposal and interpretation, and uses evidence that remains stable under resampling and null calibration as the search feedback. ReIG then reports final candidates along separate axes of reliability, novelty, and non-triviality. Across Rep-60, Stress-72, and faceted report evaluations, ReIG consistently improves discovery quality under matched search policies. Under MCTS, ReIG-MCTS improves ValidDisc@10 from 0.57 to 1.55 over Surprise-MCTS, raises reward AUROC from 0.44 to 0.83, and reduces false flags on no-signal tasks from 3.8 to 0.38 at budget 20. Ablations, same-candidate reranking, and the semantic audit show that the gains come from guiding exploration with validated, reusable evidence and keeping reliability, novelty, and non-triviality as distinct criteria."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gao-etal-2026-surprise">
<titleInfo>
<title>When Surprise Is Not Enough: Validation-Aware Rewards for LLM-Guided Scientific Discovery</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yifan</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haoyue</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feng</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>LLM-based scientific exploration agents need feedback signals that decide which experimental outcomes should guide future search. Recent systems use Bayesian surprise, rewarding results that change an LLM’s belief. This can create a noisy-TV failure mode: unstable artifacts, leakage, proxy effects, or gaps in the model’s prior may appear more valuable than reliable mechanisms. We introduce Reducible Information Gain (ReIG), a validation-aware discovery method. ReIG keeps LLMs for hypothesis proposal and interpretation, and uses evidence that remains stable under resampling and null calibration as the search feedback. ReIG then reports final candidates along separate axes of reliability, novelty, and non-triviality. Across Rep-60, Stress-72, and faceted report evaluations, ReIG consistently improves discovery quality under matched search policies. Under MCTS, ReIG-MCTS improves ValidDisc@10 from 0.57 to 1.55 over Surprise-MCTS, raises reward AUROC from 0.44 to 0.83, and reduces false flags on no-signal tasks from 3.8 to 0.38 at budget 20. Ablations, same-candidate reranking, and the semantic audit show that the gains come from guiding exploration with validated, reusable evidence and keeping reliability, novelty, and non-triviality as distinct criteria.</abstract>
<identifier type="citekey">gao-etal-2026-surprise</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.245/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>5596</start>
<end>5611</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Surprise Is Not Enough: Validation-Aware Rewards for LLM-Guided Scientific Discovery
%A Gao, Yifan
%A Li, Haoyue
%A Yuan, Feng
%A Gao, Xin
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F gao-etal-2026-surprise
%X LLM-based scientific exploration agents need feedback signals that decide which experimental outcomes should guide future search. Recent systems use Bayesian surprise, rewarding results that change an LLM’s belief. This can create a noisy-TV failure mode: unstable artifacts, leakage, proxy effects, or gaps in the model’s prior may appear more valuable than reliable mechanisms. We introduce Reducible Information Gain (ReIG), a validation-aware discovery method. ReIG keeps LLMs for hypothesis proposal and interpretation, and uses evidence that remains stable under resampling and null calibration as the search feedback. ReIG then reports final candidates along separate axes of reliability, novelty, and non-triviality. Across Rep-60, Stress-72, and faceted report evaluations, ReIG consistently improves discovery quality under matched search policies. Under MCTS, ReIG-MCTS improves ValidDisc@10 from 0.57 to 1.55 over Surprise-MCTS, raises reward AUROC from 0.44 to 0.83, and reduces false flags on no-signal tasks from 3.8 to 0.38 at budget 20. Ablations, same-candidate reranking, and the semantic audit show that the gains come from guiding exploration with validated, reusable evidence and keeping reliability, novelty, and non-triviality as distinct criteria.
%U https://aclanthology.org/2026.emnlp-main.245/
%P 5596-5611
Markdown (Informal)
[When Surprise Is Not Enough: Validation-Aware Rewards for LLM-Guided Scientific Discovery](https://aclanthology.org/2026.emnlp-main.245/) (Gao et al., EMNLP 2026)
ACL