Beyond Confidence: Test-Time Scaling for Multi-Turn Search Agents via Retrieval Grounding
Hyunho Kook, Junhyuk So, Tianyu Fu, Haizhong Zheng, Beidi Chen
EMNLP 2026 Findings
cs.AI
2026-08-25
Retrieved docs inflate copied-token logprobs and flatten confidence; voting by answer–document overlap gains up to +5.4 accuracy and +35 on minority-correct questions.
In single-turn reasoning, weighting parallel samples by token log-probabilities (DeepConf and relatives) often beats simple majority. Once a search agent appends retrieved documents to context, that signal breaks. The paper names the mechanism copy inflation: tokens copied from the just-appended documents get systematically higher log-probabilities, within-question confidence flattens, and the weighted vote collapses toward counting heads.
On BrowseComp-Plus rollouts from Tongyi-DeepResearch, copy tokens sit 0.50 nats above non-copy tokens. The median trajectory copies 93% of its content tokens from documents it retrieved; 81% copy at least 90%. As the copy fraction rises, DeepConf's within-question share of score variance falls from 0.39 to 0.20. Even when all eight rollouts are wrong, mean DeepConf stays at 87% of its peak. Correct and wrong rollout scores overlap 54.2%.
Retrieval-Grounded Voting stops reading the model's log-probabilities and reads the retrieval log instead. For each rollout it takes the final answer prose (not the internal chain of thought) and the set of documents that rollout fetched. The weight is the maximum fraction of answer tokens that appear in any single retrieved document: intersection size over answer-token count, then max over documents. Max avoids dilution from irrelevant docs. Normalizing by the answer side avoids punishing a well-anchored short answer for retrieving a long passage.
That weight plugs into weighted majority. Answer strings are lightly normalized and clustered; the cluster with the largest summed weight wins. No log-probs, no fine-tuning, no extra model call. About 0.3 ms per rollout on one CPU thread.
Four search-agent benchmarks (BrowseComp-Plus, BrowseComp, GAIA, FRAMES) times five models (Tongyi-DeepResearch, gpt-oss-120b, MiniMax-M2.7, Kimi-K2.5, GLM-5.1), eight rollouts per question. RGV is best in all 20 cells.
The largest cell is Tongyi on BrowseComp-Plus: single 51.6, simple majority 62.0, DeepConf 65.7, RGV 71.1, oracle 74.2, +5.4 over DeepConf. Open-web BrowseComp is noisier; MiniMax moves from 46.0 DeepConf to 50.7. On FRAMES, single-rollout accuracy already sits at 84% or higher, so the lift shrinks to a fraction of a point, and no cell regresses.
Hard questions carry the gap. On 830 Tongyi/BrowseComp-Plus items, minority-correct questions (1–2 of 8) score 84.4 with RGV versus 49.2 with DeepConf, a 35.2-point spread. RGV at four rollouts reaches 65.9, matching DeepConf at eight (65.7). ROC AUC for correct versus wrong rollouts is 0.908 versus 0.837. When gold supporting documents exist, correct rollouts have gold recall 0.94 versus 0.31 for wrong ones; RGV correlates with gold recall at 0.57, DeepConf at 0.36. Digits and capitalized entities carry the score; common content words barely separate correct from wrong.
Test-time scaling for search agents does not have to read log-probs. The documents are already on the trajectory, lexical overlap is nearly free, and four rollouts match eight confidence votes. Any consumer of token-level confidence (early stopping, routing, abstention, logprob-based hallucination checks) reads the same inflated signal once retrieval sits in context.
The method is cheap to adopt. It measures whether retrieval anchored the answer, not whether the answer is true.
Grounded is not correct: among 616 questions that have a correct rollout, 26 (4.2%) lose because a wrong rollout retrieved on-topic documents and echoed them, well-grounded but wrong. Forced single-entity answers collapse the score to containment and tie DeepConf. When retrieval is weak, RGV's ceiling drops with it; when retrieval barely matters, neither does RGV. Non-retrieval tools such as code interpreters need a different grounding signal. Experiments are English. A poisoned corpus would inflate rollouts that copy from it, a weakness shared by any method that reads that log. Most cells sample at most 150 questions per benchmark; the 830-question mechanism study is Tongyi × BrowseComp-Plus only.