Atomic Calibration of LLMs in Long-Form Generations

• Conference: ACL 2025
• arXiv: 2410.13246
• Code: Not provided
• Area: LLM Evaluation / Uncertainty Estimation / Factuality Calibration
• Keywords: Atomic Calibration, Confidence Elicitation, Long-Form Generation, Hallucination, Confidence Fusion

TL;DR

This work systematically studies atomic calibration in long-form generation, categorizing confidence elicitation methods into discriminative and generative approaches. It finds these two types to be complementary and proposes a fusion strategy based on confidence consistency, revealing interesting patterns in how model confidence changes during the generation process.

Background & Motivation

• Problem: LLM confidence calibration is crucial for hallucination detection, but existing studies mainly focus on response-level calibration (macro calibration) in short-text QA tasks—assigning a single confidence score to the entire answer. In long-form generation, an answer may contain both accurate and inaccurate claims, and a single score cannot reflect fine-grained factuality.
• Key Questions: (1) Why is it necessary to evaluate calibration at the atomic claim level? (2) What factors influence atomic-level calibration? (3) What patterns can atomic-level analysis reveal that macro-level analysis cannot?
• Core Definition: Atomic-level calibration = Decomposing a long-form response into atomic claims (each containing a single fact), assigning confidence to each claim, and evaluating how well the confidence aligns with actual factuality.

Method

Overall Architecture

1. Given a query \(q\), the LLM generates a long-form response \(x\).
2. Use GPT-4o to decompose \(x\) into \(N\) atomic claims \(\{c_1, ..., c_N\}\).
3. Use GPT-4o combined with Wikipedia/Google Search to verify the factuality label \(y_i \in \{0, 1\}\) for each claim.
4. Use different confidence elicitation methods to estimate the confidence \(f(c_i)\) for each claim.
5. Use ECE, Brier Score, and AUROC to evaluate the quality of atomic calibration.

Key Designs

1. Discriminative confidence methods — let the model self-evaluate:

   • Dis-Single: Directly ask the model whether a single claim is true, taking \(P(\text{True})\) as the confidence.
   • Dis-Context: Same as above, but providing the original context to assist the judgment.
   • Dis-Rating: Prompt the model to directly output a numerical confidence score from 0 to 10.

2. Generative confidence methods — based on sampling consistency:

   • Gen-Binary: Sample an additional \(K\) responses, use an NLI model to determine if the atomic claim is supported, and compute confidence as \(\frac{|K_s|}{|K|}\).
   • Gen-Multi: Distinguish between "contradicted" and "not mentioned", computing confidence as \(\frac{|K_s|}{|K_s| + |K_c|}\).

3. Confidence Fusion Strategies (proposed in this paper):

   • AdjustedAlpha: Dynamically adjust the fusion weight \(\alpha' = \alpha + \gamma_a \cdot d\) based on the difference between the two confidence values \(d = B - A\).
   • DampedFusion: Apply damping based on consistency, defined as \(\gamma(d) = 1 - k \cdot |d|\), to reduce the overall confidence when inconsistencies exist.
   • Core Idea: Traditional weighted averaging cannot distinguish between cases like \((0, 1)\) and \((0.4, 0.6)\). The former possesses higher inconsistency and should result in a lower final confidence.

Loss & Training

No training process is involved. Evaluation is conducted using Expected Calibration Error (ECE), Brier Score (BS), and AUROC.

Key Experimental Results

Main Results — Atomic Calibration (ECE ↓, BS ↓, AUROC ↑)

Method	Llama3-8B ECE	Mistral-7B ECE	Qwen2-7B ECE
Dis-Context	35.5 / 11.9 / 12.5	24.8 / 15.7 / 20.6	26.5 / 13.9 / 17.2
Dis-Single	32.6 / 14.3 / 19.2	30.2 / 20.4 / 24.0	29.3 / 16.1 / 18.7
Gen-Binary	10.0 / 8.5 / 11.1	13.7 / 8.4 / 12.7	10.9 / 6.3 / 9.5
Gen-Multi	37.4 / 12.6 / 21.9	42.2 / 13.4 / 26.6	41.7 / 11.6 / 21.0

(The three columns correspond to Bios / LongFact / WildHallu datasets respectively)

Ablation Study — Impact of Model Size on Calibration

Method	Qwen2-7B ECE	Qwen2-57B ECE	Qwen2-72B ECE
Gen-Binary	10.9	10.5	11.2
Dis-Rating	41.5	23.2	11.4

(Bios dataset)

Confidence Fusion Results (Cross-category: Gen-Binary + Dis-Context)

Fusion Method	Llama3-8B ECE	Mistral-7B ECE
WAvg (Weighted Average)	15.2	12.8
MinConf	14.0	11.5
AdjustedAlpha	9.3	10.2
DampedFusion	9.5	10.4

Key Findings

1. Atomic calibration is significantly worse than response-level calibration: All models show a much higher ECE at the atomic level compared to the response level (data points consistently lie above the identity line). Even models that appear well-calibrated at the response level still perform poorly at the atomic level.
2. Gen-Binary is the most reliable single method: It achieves the lowest ECE across almost all models and datasets, but does not yield the highest AUROC. This demonstrates that calibration and discriminative ability are distinct dimensions.
3. Discriminative and generative methods are complementary: Cross-category fusion (Gen + Dis) significantly improves calibration, while intra-category fusion (Dis + Dis) yields limited gains.
4. Larger models are not necessarily better calibrated: Generative methods are insensitive to model size. However, in discriminative methods, larger models perform significantly better (Qwen2-7B Dis-Rating ECE drops from 41.5 to 11.4 on Qwen2-72B).
5. Confidence changes during the generation process: In discriminative methods, model confidence decreases as generation progresses. In generative methods, confidence is lowest during the middle of the generation. This suggests that different methods capture different types of uncertainty.

Highlights & Insights

• Systematic Framework: For the first time, a formal definition of atomic-level calibration (Definition 2) is presented, clearly distinguishing between macro vs. atomic calibration and proving that the two are not interchangeable.
• Methodology Taxonomy: Categorizes confidence elicitation methods into discriminative and generative approaches, revealing their complementarity and providing clear methodological guidance for future research.
• Profound Analytical Insights: Analyzing the patterns of confidence variation across generation positions and alignment across different methods provides a fresh understanding of the intrinsic uncertainty in LLMs.

Limitations & Future Work

• Atomic claim decomposition and factuality verification rely on GPT-4o, introducing pipeline errors.
• Post-processing calibration methods (e.g., temperature scaling, Platt scaling) are not considered, evaluating only "raw" calibration.
• Only 7 models from 3 families were evaluated, lacking analysis on larger-scale models (e.g., GPT-4, Claude).
• The hyperparameters \(\gamma_a\) and \(k\) in the confidence fusion strategies (AdjustedAlpha, DampedFusion) require tuning on a validation set.
• Focus is limited to the factuality dimension; atomic-level calibration for other quality dimensions, such as coherence or creativity, remains unexplored.

Related Work & Insights

• Atomic Fact Decomposition: FActScore (Min et al., 2023), VeriScore (Song et al., 2024), D-FActScore (Chiang & Lee, 2024)
• Uncertainty Estimation: Semantic Entropy (Kuhn et al., 2022), P(true) (Kadavath et al., 2022), Self-Rating (Tian et al., 2023)
• Long-Form Calibration: Luq (Zhang et al., 2024b), Linguistic Calibration (Band et al., 2024)
• Confidence Fusion: Weighted averaging by Rivera et al. (2024)

Rating

• Novelty: 8/10 — The formal definition of atomic-level calibration and the discovery of the complementarity between discriminative/generative methods carry significant value.
• Technical Depth: 7/10 — The fusion strategies are simple and effective though lacking in high technical complexity; the formal definition is rigorous.
• Experimental Thoroughness: 8/10 — Incorporates 7 models, 3 datasets, 5 methods, and multiple fusion strategies, offering rich analysis dimensions (model size, generation position, method alignment).
• Writing Quality: 8/10 — Concept definitions are clear, experimental results are presented systematically, and diagrams aid understanding.
• Overall Score: 7.5/10

Related Papers

• [ACL 2025] A Conformal Risk Control Framework for Granular Word Assessment and Uncertainty Calibration of CLIPScore Quality Estimates
• [ACL 2025] CuLEmo: Cultural Lenses on Emotion - Benchmarking LLMs for Cross-Cultural Emotion Understanding
• [ACL 2025] EcomScriptBench: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association
• [ACL 2025] AndroidLab: Training and Systematic Benchmarking of Android Autonomous Agents
• [ACL 2025] Browsing Lost Unformed Recollections: A Benchmark for Tip-of-the-Tongue Search and Reasoning