Automatic Evaluation of Attribution by Large Language Models

• Only 52% of statements from generative search engines such as New Bing and PerplexityAI are fully supported by their cited references.Source: §1 · p. 1
• Human evaluation of attribution is expensive and time-consuming, costing about $1 per annotated example in one study.Source: §1 · p. 1
• AttrScore categorizes attribution as attributable, extrapolatory, or contradictory to identify specific error types.Source: §1 · pp. 1–2
• The framework evaluates prompting LLMs and fine-tuning LMs on simulated and repurposed data from QA, fact-checking, NLI, and summarization.Source: §1 · p. 2
• The authors curate the first real-life-interaction evaluation set spanning 12 New Bing domains, helping avoid data contamination.Source: §1 · p. 2
• Both approaches show reasonable performance, but failures remain common when comparing contextual cues, numerical values, and symbolic operations.Source: §1 · p. 2

• The single-reference setting is treated as a unit task that can support later handling of multiple references or segments.Source: §2 · p. 3
• Attributable means the reference fully supports the generated answer.Source: §2 · p. 3
• Extrapolatory means the reference lacks sufficient information to validate the generated answer.Source: §2 · p. 3
• Contradictory means the generated answer conflicts with information presented in the reference.Source: §2 · p. 3
• Fine-grained error labels help users distinguish error types and guide targeted training mechanisms, although category boundaries can blur in rare complex cases.Source: §2 · p. 3
• Attribution evaluation maps a query, generated answer, and reference to an attributable, extrapolatory, or contradictory class.Source: §2 · p. 3

3 Automatic Evaluation of Attribution

• AttrScore explores prompting LLMs and fine-tuning LMs on simulated and repurposed data for automatic attribution evaluation.Source: §3 · p. 3

3.1 Prompting LLMs

• Prompted LLMs receive definitions of attribution errors and a query-answer-reference triple for evaluation.Source: §3.1 · p. 3

3.2 Fine-tuning LMs on Repurposed Data

• Fine-tuning is constrained by a lack of training data, while collecting real examples through annotation can be prohibitively expensive.Source: §3.2 · p. 3
• Fact-checking, NLI, and summarization datasets are adapted because they model consistency or logical relations between evidence and generated statements.Source: §3.2 · pp. 3–4
• QA simulations use original question-answer-context triples as attributable examples, substitute answers or context spans for contradictory cases, and retrieve answer-free alternatives for extrapolatory cases.Source: §3.2 · p. 4, Figure 2 · p. 4
• Short QA answers are converted into longer sentences so simulated training data better matches attributed LLMs’ long-answer setting.Source: §3.2 · p. 4, Figure 2 · p. 4

4.1 Datasets

• Training data combines QA, fact-checking, NLI, and summarization datasets, using all summarization examples and 20K samples from each of the other task groups.The main training set combines the simulated datasets from these related tasks.Source: §4.1 · pp. 4–5
• AttrEval-Simulation uses six out-of-domain QA datasets to test generalization, while AttrEval-GenSearch contains approximately 250 manually annotated New Bing examples across 12 domains.The New Bing samples treat each reference or answer segment as a separate attribution example and label it attributable, contradictory, or extrapolatory.Source: §4.1 · p. 5, Figure 3 · p. 5
• The annotated New Bing set is created with new queries across 12 domains to reduce training-data contamination and assess real-life attribution evaluation.Queries are sent to New Bing in balanced mode, and annotators verify the first answering sentence with its reference.Source: §4.1 · p. 5

4.2 Implementation Details

• Prompted experiments evaluate Alpaca, Vicuna, ChatGPT, and GPT-4, averaging results over four prompts and using three demonstrations for few-shot settings.GPT-4 is evaluated on 500 randomly sampled AttrEval-Simulation examples because of its cost and inference speed.Source: §4.2 · pp. 5–6
• Fine-tuning covers Roberta, T5, GPT2, LLaMA, Alpaca, and Vicuna models, with F1 reported per class and micro-F1 used as overall accuracy.Training uses Huggingface implementations on four A100 80GB GPUs, with model-specific batch sizes and epoch counts.Source: §4.2 · p. 6

5.1 Overall Performance

• GPT-4 reaches 81–83% overall accuracy on AttrEval-GenSearch and outperforms other evaluated models, while fine-tuning related-task data substantially improves smaller models.Vicuna (13B) rises from 34.6% to 66.0% on AttrEval-Simulation and from 41.4% to 71.3% on AttrEval-GenSearch after fine-tuning.Source: §5.1 · pp. 6–7
• Contradictory errors are especially difficult, and even GPT-4 and fine-tuned models often classify contradictory inputs as attributable.The difficulty requires comparing one piece of information in the statement against another in the reference.Source: §5.1 · p. 7

5.2 Qualitative Analysis

• 30.6% of GPT-4’s errors arise from insensitivity to fine-grained information such as numbers, dates, and times.The qualitative analysis examines GPT-4’s zero-shot errors using representative examples in Table 3.Source: §5.2 · p. 7, Table 3 · p. 7
• Additional GPT-4 errors involve misunderstanding task definitions or label relations (22.2%), symbolic operators (13.9%), and failure to condition judgments on reference context.The model may confuse equal with approximately equal and may rely on parametric knowledge rather than reference evidence.Source: §5.2 · p. 7

5.3 Ablation Study

• Simulated QA and fact-checking examples provide the largest individual-task improvements, while combining related datasets generally helps, especially on out-of-domain GenSearch instances.Figure 4 reports the influence of individual task data and dataset combinations.Source: §5.3 · p. 8, Figure 4 · p. 8
• Fact-checking and NLI prompts generally outperform attribution-specific and summarization prompts in the evaluated prompting settings.The authors attribute this pattern to possible exposure to similar tasks during instruction tuning.Source: §5.3 · p. 8

Attributed LMs.

• Prior work augments generative LMs with retrievers, search engines, and other external tools to reduce hallucinations, but the validity of reference-supported attribution remains questionable.The related-work discussion frames attribution as an unresolved consequence of reference-augmented generation.Source: §6 · p. 8

Evaluation of attribution.

• Human audits have found unsupported statements and inaccurate citations in generative search responses, but human evaluation is expensive and time-consuming.This motivates automatic attribution evaluation beyond manual verification.Source: §6 · p. 8
• Existing automatic approaches rely largely on NLI models, whereas this work studies additional related tasks and evaluates both benchmark-derived and real examples.The broader setting includes fact-checking, NLI, summarization, QA-derived data, and examples from a generative search engine.Source: §6 · p. 8

7 Conclusion

• AttrScore defines attribution errors and evaluates them through prompted LLMs and fine-tuned smaller language models.The experiments use simulated examples and manually curated examples from a real-life generative search engine.Source: §7 · p. 8
• Results on simulated and curated testbeds show promising signals but also substantial challenges for automatic attribution evaluation.The paper positions its formulation, testbeds, and findings as a foundation for future studies.Source: §7 · p. 8

• Simulated and repurposed training data may not reflect real attribution scenarios, and simplistic, noisy labels can limit models’ handling of varied errors.The authors identify higher-quality training data at scale as a major direction for future work.Source: §8 · p. 9
• AttrEval-GenSearch may be biased because GPT-4 generated the New Bing examples and also evaluated them, while its moderate size limits representativeness.Source: §8 · p. 9
• Future evaluations should diversify generative search sources and annotate larger datasets spanning broader topics, styles, and perspectives.Source: §8 · p. 9

• The authors mitigate potential harm by carefully selecting evaluation data and analyzing risks during the research process.Source: §9 · p. 9

A.1 Simulation - QA

• QA simulations label answer-supported evidence as attributable, replace answers or references to create contradictory cases, and retrieve answer-free documents for extrapolatory cases.Source: §A.1 · p. 14

A.2 Simulation - Fact Checking

• Fact-checking datasets are mapped into attributable, contradictory, and extrapolatory labels using their original evidence and claim annotations.Source: §A.2 · p. 14

A.3 Simulation - NLI

• NLI datasets map entailment, contradiction, and neutrality to attributable, contradictory, and extrapolatory attribution categories.Source: §A.3 · pp. 14–15

A.4 Simulation - Summarization

• Summarization datasets use factuality and hallucination annotations to construct attributable, contradictory, and extrapolatory examples.Source: §A.4 · p. 15

B Label and Subset Distributions of Training and Test Sets

• The authors report label and data-source distributions for the training and AttrEval-Simulation sets in Figures 5 and 6.Source: B Label and Subset Distributions of… · p. 15, Figure 5 · p. 18, Figure 6 · p. 18

C Prompts for LLMs as AttrScore

• AttrScore uses varied prompts and few-shot demonstrations for LLM-based attribution evaluation, with prompt details provided in Tables 6 and 7.Source: C Prompts for LLMs as AttrScore · p. 15

D Generative Search Engine Examples Annotation Protocol

• The annotation protocol has workers create domain-specific queries and verify whether generated responses are fully supported by their references.Source: D Generative Search Engine Examples… · pp. 15–16
• Examples distinguish attributable support, extrapolatory insufficiency, and contradictory disagreement, including exact numerical conflicts such as 4.31% versus 3.81%.The protocol illustrates these categories with Moon temperatures, U.S. gas prices, and German unemployment.Source: D Generative Search Engine Examples… · pp. 16–17

• 79.4% of ChatGPT’s errors involve overlooking contextual clues and failing to condition judgments on the reference.The analysis is reported in Table 8.Source: E Additional Qualitative Analysis · p. 18, Table 8 · p. 21
• 13.8% of errors reflect insensitivity to fine-grained information such as numerical values, dates, and times.Source: E Additional Qualitative Analysis · p. 18
• 6.8% of errors involve failure to perform symbolic operations, including checking set relationships over the reference.Source: E Additional Qualitative Analysis · p. 18
• The main experiments use the “Attribution” prompt because Table 6 reports it as the best-performing prompt overall.Source: Table 6 · p. 19