cs.CLAug 27, 2026

Evaluating Confidence-Gated Retrieval with Matched Trajectory Replay

Authors: Prateek Chhikara

Organizations: University of Southern California, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Interactive language-model agents use confidence signals to decide whether to answer immediately, retrieve additional evidence (from memory or external knowledge), or defer. Yet confidence is usually evaluated in isolation, without measuring the trajectory-level consequences of the actions it triggers. We propose matched trajectory replay, a controlled protocol for comparing confidence-to-action mappings. The protocol holds candidate answer states, evidence points, budgets, and action costs fixed. We use it to compare raw verbalized confidence with post-hoc isotonic calibration in a multi-hop question-answering system using Mistral, GPT, and Qwen models on HotpotQA and MuSiQue datasets. At the same numerical commitment threshold, calibration changes which questions agents ultimately commit to answering. Across all six model-dataset pairs, it increases accuracy among committed answers by up to 41 percentage points. However, it can reduce coverage and increase retrieval use. Overall accuracy improves by up to 15 percentage points on HotpotQA but falls by up to 17 percentage points on MuSiQue. These effects reflect a shift to a more selective, lower-risk operating point, not improved answers or confidence ranking. A calibration map fitted before retrieval improves held-out calibration through retrieval depths one and two, but is worse than raw confidence at depth three for all three models. Additional evidence helps on average, but this aggregate effect does not establish whether confidence identifies which individual episodes will benefit from another retrieval. Taken together, these results show that calibration can make commitment risk interpretable, but it does not estimate the expected benefit of another retrieval. Retrieval therefore requires a separate value-of-information or utility estimate. Evaluations should report held-out calibration, risk-coverage, and retrieval cost.

Figures & tables

Figure 1: In this example, a raw confidence of 0.750.75 exceeds the commitment threshold even though comparable answers are correct only about half the time. Mapping the score to 0.500.50 triggers retrieval instead; the action has a cost, while its benefit is not implied by calibration.
Figure 2: Matched trajectory replay. Candidate answer–confidence states are collected once for each fixed evidence plan and replayed through raw and calibrated policies.
  HotpotQA ( N=1,852N=1{,}852 )MuSiQue ( N=605N=605 )
ModelSystemOACACov.CostOACACov.Cost
Mistral Small 4 (2603)Parametric only39.139.1100.00.0010.610.6100.00.00
Fixed RAG@166.666.6100.01.0034.434.4100.01.00
Fixed RAG@266.466.4100.02.0041.241.2100.02.00
Fixed RAG@383.083.0100.03.0054.954.9100.03.00
Adaptive proxy72.472.4100.02.3252.452.4100.02.88
Table 1: Controlled-trajectory test results at the fixed operating point. All rows use fixed plans that interleave dataset-marked supporting passages with ranked distractors. OA ↑\uparrow is the percentage of all episodes ending in a correct commit; CA ↑\uparrow is accuracy conditional on commitment; Cov. ↑\uparrow is coverage; and Cost ↓\downarrow is the mean number of retrieval actions. Boldface marks the highest OA or CA within each model–dataset block. Blue highlight marks the higher OA or CA value between Calibrated-Gated and Raw-Gated . Fixed-depth and adaptive baselines always commit and therefore have 100% coverage.
Figure 3: Fixed numerical thresholds on the risk–coverage plane for the micro-pooled controlled-trajectory tests ( N=2,457N=2{,}457 : 1,852 HotpotQA and 605 MuSiQue). Raw and calibrated lines apply the same set of up to seven numeric thresholds; they are sampled operating points, not full or matched-coverage frontiers. Circles mark τ=0.7\tau=0.7 and squares show confidence-blind baselines.
 Depth 0Depth 1Depth 2Depth 3
ModelRawCal.RawCal.RawCal.RawCal.
Mistral Small 4 (2603)44.22.527.418.127.316.715.423.0
GPT-OSS-120B27.72.512.29.512.78.66.18.0
Qwen3-235B-A22B-250725.52.022.05.521.07.012.814.6
Table 2: Held-out 10-bin ECE (percentage points) on test states. Depth zero is invariant to retrieved evidence; later depths pool all collected evidence paths and micro-average HotpotQA and MuSiQue. Calibration improves ECE through depth two, but at depth three calibrated ECE exceeds raw ECE for all three models.
Figure 4: OA–cost trade-off on the micro-pooled controlled-trajectory test set. Curves vary the maximum retrieval budget from 0–3; realized cost can be lower when the controller commits early. Squares denote no retrieval, Fixed RAG@1/@2, the adaptive proxy, and Fixed RAG@3.
ModelHelpfulHarmfulNet (pp)
Mistral Small 4 (2603)18.43.7+14.7+14.7
GPT-OSS-120B16.23.4+12.8+12.8
Qwen3-235B-A22B-250718.43.6+14.9+14.9
Average17.73.6+14.1+14.1
Table 3: Adjacent retrieval-transition diagnostics on micro-pooled controlled-trajectory test traces. Helpful denotes wrong →\rightarrow correct, harmful denotes correct →\rightarrow wrong, and Net (pp) is the helpful-minus-harmful percentage-point difference. Each model has 7,367 retrieval transitions that reveal new evidence; rates are percentages of transitions.
Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Model in paperProvider identifierTotal paramsActive paramsArchitecture typeTraining / alignment strategyContext limit
Mistral Small 4 (2603)mistralai/ mistral-small-2603119B6.5B per tokenMultimodal sparse MoE transformer; 128 experts, 4 active per tokenOpen-weight hybrid instruct, reasoning, and coding model with configurable reasoning effort256k tokens
GPT-OSS-120Bopenai/ gpt-oss-120b116.8B5.1B per tokenAutoregressive MoE transformer with alternating dense and locally banded sparse attention, grouped multi-query attention, and RoPEOpen-weight reasoning model; text-only pretraining followed by supervised fine-tuning and high-compute reinforcement learning for instruction following, tool use, and reasoning128k tokens
Qwen3-235B -A22B-2507qwen/ qwen3-235b- a22b-2507235B22B per tokenCausal-language MoE transformer; 94 layers, GQA, 128 experts, 8 active expertsPretraining plus post-training for multilingual instruction following, non-thinking generation, reasoning, coding, and tool use262,144 tokens
Appendix
Table 4: Evaluated model details: context limits and model descriptions are the provider-reported limits for the checkpoint or hosted identifier; the experiment itself caps generation at 4096 output tokens and uses temperature zero.
Dataset / artifactCalibrationTest
HotpotQA3,7021,852
MuSiQue answerable development1,208605
Complete-artifact total4,9102,457
Appendix
Table 5: Unique source-example counts used for every model under the 2:1 calibration–test split. Calibration fits the isotonic map, and test is reserved for final replay.
Figure 5: Calibration-partition reliability diagrams for the three evaluated answer models. Each panel compares verbalized depth-zero confidence against empirical correctness on the records used to fit that model’s isotonic map; the calibration gap is shown by the hatched region.
More in this topic

AI Agent Evaluation

AI agent evaluation measures whether tool-using systems complete tasks reliably, safely, and efficiently across interactive settings.

See all 516 papers

Explore similar work

CardsList
Sep 1, 2026

DualStake: Dual-Path Confidence Calibration in Deep Research Agents

Deep Research agents tackle knowledge-intensive tasks through multi-round retrieval and decision-oriented generation. However, these agents suffer from severe overconfidence, making their expressed confidence unreliable for user trust and downstream abstention. To address this, we augment the Deep Research pipeline with step confidence elicitation after each retrieval, building on the commonly used post-answer verbalized confidence. Interestingly, we find that Evidence Confidence (E-Conf), elicited after the final retrieval step, provides a stronger uncertainty signal than Answer Confidence (A-Conf), elicited after answer generation, and that A-Conf is largely shaped by E-Conf. Based on these findings, we propose DualStake, a dual-path calibration method that applies margin-clipped, confidence-dependent stake rewards to jointly align E-Conf and A-Conf with answer correctness while limiting extreme confidence optimization. Experiments on Qwen2.5-7B, Qwen2.5-7B-Instruct, and Qwen3-4B across 8 QA benchmarks demonstrate that DualStake consistently improves calibration without sacrificing answer accuracy. The code is available at https://github.com/FloXXXt/DualStake.

Yinuo Xu, Yuwei Liang, Jianjie Cheng +4

NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA) · School of Artificial Intelligence, University of Chinese Academy of Sciences · Meituan Inc.

Deep Research AgentsLanguage Model CalibrationLLM Uncertainty EstimationUncertainty Calibration
AskPDF
Oct 19, 2026

Measurement Validity and Representation Dynamics of Confidence Signals in Large Language Models

Whether a language model "knows what it knows" matters for trustworthy deployment and for deciding when it should abstain. Prior work usually measures this through one channel: verbalized confidence, a linear probe on hidden states, or agreement among sampled answers (semantic entropy), and often treats these as interchangeable. We compare all three in a pre-registered audit of five Qwen2.5 models (0.5B-7B, including a base/instruct pair at 7B) on factual retrieval (PopQA, SimpleQA) and multi-step reasoning (GSM8K, MATH). First, probes trained on prefill states to predict whether sampled answers will disagree become informative earlier for retrieval (about 7% of layer depth on average) than for reasoning (about 28%). The pre-registered five-layer test supported this only marginally; a finer 21-layer follow-up, designed after those results, supports it more clearly. In most retrieval cells, however, a prompt-only TF-IDF classifier matches the probe, so the early signal may reflect surface features of the question. Second, verbalized confidence is unstable: slightly different prompt templates yield scores that barely correlate. Third, Expected Calibration Error is misleading: verbal confidence looks well calibrated because it clusters near the base rate while discriminating poorly, whereas sampling agreement has worse ECE but about twice the resolution. Of five pre-registered hypotheses, two were falsified, one was partly supported, one was not supported, and one could not be tested because the base model's verbal confidence was degenerate.

Debojeet Ghosh, Shreyansh Kumar, Ayushman Garg

Confidence Estimation in Language ModelsLanguage Model CalibrationVerbalized Confidence EstimationLanguage Model ProbingSemantic Entropy
AskPDF
Jul 28, 2026

Beyond Self-Knowledge: Propagating Uncertainty Across Reasoning and Retrieval in LLMs

Retrieval-augmented generation improves knowledge-intensive question answering, but indiscriminate retrieval can introduce irrelevant evidence and unnecessary computation. We investigate whether verbalized confidence from black-box language models can serve as an actionable signal for retrieval routing. Our method, BeyondUncertainty, first elicits a structured provisional answer and confidence estimate, then applies a model-specific threshold selected on held-out validation data and frozen before test evaluation. Low-confidence questions receive top-5 TF-IDF retrieval followed by a second answer call, whereas high-confidence questions return the provisional answer directly. We evaluate 27,000 policy instances across six QA benchmarks, three model families, and three retrieval policies. BeyondUncertainty achieves 0.483 mean token-level F1, compared with 0.467 for always retrieval and 0.401 for no retrieval, while reducing retrieved passages by 20.4% relative to always retrieval. When matched on the number of questions routed to retrieval within each dataset-model cell, it outperforms a post-hoc random allocation in 17 of 18 settings, with an average gain of 0.024 F1. Although poorly calibrated as an absolute probability, probe uncertainty modestly predicts question-level retrieval benefit (AUROC = 0.628). However, the additional probe increases total token usage by 28.2%, revealing a trade-off between more selective evidence acquisition and end-to-end token efficiency.

Chandan Kumar Sah, Li Zhang, Xiaoli Lian

Beihang University, Beijing, China

Retrieval-Augmented GenerationConfidence Estimation in Language ModelsLLM Uncertainty EstimationRAG EvaluationEfficient Language Model Inference
AskPDF