Large language models (LLMs) often hallucinate in long-form generation. Existing approaches mainly improve factuality through post-hoc revision or reinforcement learning (RL) with correctness-based rewards, but they do not teach the model to estimate which parts of its generation are reliable. As a result, models may still state incorrect claims confidently in their responses. Recent advances in reasoning have significantly improved LLM performance, and have been leveraged to estimate confidence by incorporating calibration into RL objectives. However, existing approaches remain limited to a single scalar confidence for the entire response, which is insufficient for long-form generation where uncertainty varies across individual claims. To mitigate this problem, we propose CURE, a framework that improves long-form factuality by teaching LLMs to reason about uncertainty at the claim level. We first introduce a Claim-Aware Reasoning Protocol, which structures outputs into atomic claims paired with explicit confidence estimates. We then develop a multi-stage training pipeline that aligns model confidence with claims' correctness and then optimizes on factuality. The resulting calibrated confidence further enables selective prediction, allowing the model to abstain from uncertain claims at inference time. Experiments on four long-form factuality benchmarks show that CURE consistently improves factual accuracy over competitive supervised and RL baselines, while maintaining factual recall. In particular, it improves claim-level accuracy by up to 39.9% on Biography generation. These gains are accompanied by improved calibration, as reflected by a 16.0% increase in AUROC on FactBench.
Figures & tables
Figure 1: Illustration of the claim-aware reasoning protocol.The model performs structured reasoning with explicit uncertainty, then decomposes the output into atomic claims with calibrated confidence. These confidences enable selective prediction (Figure 3 ).
Figure 2: Stages 2 and 3 of the multi-stage training pipeline. In Stage 2 (left), confidence correction constructs preference pairs for DPO-based calibration, aligning confidence estimates with empirical correctness. In Stage 3 (right), GRPO optimizes claim-level factuality under feasibility constraints, with rewards explicitly restricted to claim tokens. This design decouples calibration from factual optimization.
Figure 3: Selective prediction via confidence thresholding. Low-confidence claims are filtered, while retained confidence scores provide explicit transparency into the final output.
Method
FactBench
LongFact
Biography
Acc.
ECE
Brier
AUROC
Acc.
ECE
Brier
AUROC
Acc.
ECE
Brier
AUROC
( ↑ )
( ↓ )
( ↓ )
( ↑ )
( ↑ )
( ↓ )
( ↓ )
( ↑ )
( ↑ )
( ↓ )
( ↓ )
( ↑ )
Base LLM
73.6
0.162
0.219
0.541
74.5
0.158
0.193
0.591
45.7
0.124
0.239
0.691
L2RF
77.1
–
–
–
79.4
–
–
–
47.1
–
–
–
LitCab
73.6
0.212
0.276
0.575
74.5
0.156
0.190
0.601
45.7
0.125
0.239
0.693
SFT
82.0
0.137
0.184
0.563
85.1
0.118
0.146
0.655
51.5
0.361
0.395
0.620
Table 1: Main results on three long-form factuality benchmarks. Stages in our training pipeline are cumulative: starting from SFT, we sequentially apply Feasibility RL (F-RL), Calibration Optimization (CO), and finally Factuality Optimization (FO). For L2RF, we report only accuracy following the results reported in their paper, since their outputs are not available for computing calibration metrics. CURE achieves the highest accuracy across all datasets and the best AUROC on FactBench and LongFact. AUROC measures the ability to distinguish correct from incorrect claims and is more informative and pratical than ECE and Brier.
Figure 4: Effect of feasibility RL on claim verifiability, reasoning faithfulness, and answer relevance on FactBench (Llama3.1-8B-Instruct). It improves all three metrics, with the largest gain in relevance.
Table 2: Comparison of joint optimization and decoupled training on FactBench using Llama3.1-8B-Instruct. Joint GRPO augments the reward with the Brier score and is evaluated with varying group sizes n (Rows 2-4), with weighted Brier in Row 5. Row 6 applies GRPO for calibration only, while Rows 7-8 use DPO-based calibration followed by factuality optimization. Joint GRPO (Rows 2-5) fails to achieve strong calibration despite larger n and reweighting. GRPO-based calibration alone (Row 6) remains insufficient, whereas DPO-based calibration (Row 7) substantially improves performance, with further gains in both factual accuracy and discriminative calibration (AUROC) after factuality optimization (Row 8).
Figure 5: Pareto frontier of factual accuracy versus recall. The curves illustrate the trade-off shaped by the training penalty ϵ and navigated by the inference confidence threshold τ . Different values of ϵ define distinct Pareto frontiers, while τ enables movement along each frontier by filtering claims based on calibrated confidence.
Method
FactBench
LongFact
Biography
Acc.
ECE
Brier
AUROC
Acc.
ECE
Brier
AUROC
Acc.
ECE
Brier
AUROC
Base
53.2
0.262
0.315
0.512
58.3
0.225
0.302
0.542
21.7
0.539
0.509
0.513
SFT
70.8
0.220
0.246
0.560
58.3
0.352
0.360
0.592
22.4
0.657
0.619
0.568
CURE
80.4
0.143
0.205
0.661
70.2
0.201
0.267
0.625
24.6
0.256
0.288
0.672
Table 3: Generalization results on Qwen3-4B across FactBench, LongFact, and Biography.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Stage
Model
# Training Instances
SFT
Llama3.1-8B-Instuct
2,987
SFT
Qwen3-4B
2,632
RL
All Models
4,000
Appendix
Table 4: Data statistics for SFT and RL training. Prompts are provided by Chen et al. (2025) .
Figure 6: Training reward dynamics under GRPO. Left: factuality reward. Right: calibration reward. Solid lines denote per-step rewards, and dashed lines denote moving averages (window size 10). The calibration reward closely follows the factuality reward across training, indicating that it is largely driven by claim correctness rather than confidence–correctness alignment. This entanglement leads to weak and noisy learning signals for calibration.
Figure 7: Pareto frontier of factual accuracy versus recall across extended configurations. Curves represent the trade-off shaped by different training-time penalties ϵ∈{0.0,0.2,0.4,0.6,0.8,1.0} , while markers indicate the navigation along each frontier via the inference-time confidence threshold τ∈{0.0,0.2,0.4,0.6,0.8} . Our framework consistently dominates the Base LLM and the accuracy-only baseline across all viable settings.
Query
Give the next three items in this series: hydrogen, lithium, boron, …
Response
<think> The user is asking for the next three items in a series that starts with hydrogen, lithium, and boron. I need to identify the pattern. First, I’ll establish the basic sequence. The series appears to be the first three elements in the periodic table, listed in order. <reason> I feel very confident about this because the sequence provided matches the well-known order of the first three elements in the periodic table. This is a fundamental fact in chemistry that I’ve seen consistently across many sources. My confidence is extremely high, a 0.99. ([Claim 1], Confidence: 0.99)… </think> <decompose> • Claim 1: The series starts with the first three elements in the periodic table. Confidence: 0.99 … </decompose>
Appendix
Table 5: Case study on claim verifiability. The claim marked in red does not constitute an independent fact, as it lacks sufficient contextual information for proper interpretation. This makes its correctness difficult to verify and may introduce noise during factuality optimization.
Query
What heavy metal song has the fastest tempo?
Response
<think> The user is asking about a specific heavy metal song with the fastest tempo. First, I’ll consider some well-known heavy metal bands and their songs. Bands like Slayer, Megadeth, and Death Angel are known for their fast tempos. <reason> I’m not certain which song has the absolute fastest tempo. I’ll assign a confidence of 0.7 that the song is from one of these bands. This is a moderate level of confidence, indicating that I’m fairly certain about the general category but not the specific details. ([Claim 1], Confidence: 0.7)… </think> <decompose> • Claim 1: A song from a well-known heavy metal band like Slayer, Megadeth, or Death Angel is likely to have the fastest tempo. Confidence: 0.70 … </decompose>
Appendix
Table 6: Case study on reasoning faithfulness. The claim highlighted in red introduces additional information not supported by the reasoning process. Specifically, the reasoning only states that Death Angel is known for their fast tempo, whereas the extracted claim further asserts that their songs are likely to be the fastest. This mismatch makes the correctness assigned to the claim unreliable, and thus affects the calibration optimization.
Large Reasoning Models achieve strong performance on complex tasks but remain prone to hallucinations, particularly in long-form generation where errors compound across reasoning steps. Existing approaches to improving factuality, including abstention and factuality-driven optimization, follow a coupled exploration-commitment paradigm, in which intermediate reasoning is unconditionally propagated to the final output, limiting fine-grained control over information selection and integration. In this paper, we propose an Exploration-Commitment Decoupling paradigm that disentangles knowledge exploration from final commitment, enabling models to explore with awareness while answering cautiously. We instantiate the paradigm with Calibration-Aware Generation (CAG), a framework that equips models with end-to-end, calibration-aware generation capabilities, by augmenting intermediate reasoning with calibrated reliability estimates and prioritizing reliable content in final outputs. Across five long-form factuality benchmarks and multiple model families, CAG improves factuality by up to 13%, while reducing decoding time by up to 37%. Overall, our work highlights decoupling as a principled approach for more reliable long-form generation, offering directions for trustworthy and self-aware generative systems.
Wen Luo, Guangyue Peng, Liang Wang +7
State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University Research Asia · MSRA · Microsoft
Large Language Models (LLMs) generate fluent long-form text, however, often add unsupported factual claims. Existing verification techniques improve factuality by grounding generation in external evidence. However, the same verification policy usually applies to all claims despite being differences in hallucination risks. We propose FACTOR (FACTuality-Oriented Risk-aware Verification), an inference-time model that adapts verification criteria according to claim-level uncertainty. FACTOR combines uncertainty estimation, adaptive language inference verification, and candidate re-ranking to allocate verification effort where it is most needed. We evaluate FACTOR on FactScore benchmark showing that adaptive verification improves factuality while reducing verification cost simultaneously. We further perform different ablation studies to identify the primary driver of these gains. Our results show the effective and model-agnostic performance of FACTOR for improving factuality in long-form generation.
Despite the rapid advancement of Large Language Models (LLMs), uncertainty quantification in LLM generation is a persistent challenge. Although recent approaches have achieved strong performance by restricting LLMs to produce short or constrained answer sets, many real-world applications require long-form and free-form text generation. A key difficulty in this setting is that LLMs often produce responses that are semantically coherent yet factually inaccurate, while the underlying semantics are multifaceted and the linguistic structure is complex. To tackle this challenge, this paper introduces Interrogative Uncertainty Quantification (IUQ), a novel framework that leverages inter-sample consistency and intra-sample faithfulness to quantify the uncertainty in long-form LLM outputs. By utilizing an interrogate-then-respond paradigm, our method provides reliable measures of claim-level uncertainty and the model's faithfulness. Experimental results across diverse model families and model sizes demonstrate the superior performance of IUQ over two widely used long-form generation datasets. The code is available at https://github.com/louisfanhz/IUQ.
Haozhi Fan, Jinhao Duan, Kaidi Xu
University of Pennsylvania · 2UNC Chapel Hill · 3City University of Hong Kong