Toward Faithful and Complete Answer Construction from a Single Document

Modern large language models (LLMs) are powerful generators driven by statistical next-token prediction. While effective at producing fluent text, this design biases models toward high-probability continuations rather than exhaustive and faithful answers grounded in source content. As a result, directly applying LLMs lacks systematic mechanisms to ensure both completeness (avoiding omissions) and faithfulness (avoiding unsupported content), which fundamentally conflicts with core AI safety principles. To address this limitation, we present EVE, a structured framework for document-grounded reasoning. Unlike free-form prompting, EVE constrains generation to a structured, verifiable pipeline that decomposes high-rigor reasoning into extraction, validation, and enumeration. Empirically, this design enables consistent and simultaneous improvements in recall, precision, and F1-score: recall and precision increase by up to 24% and 29%, respectively, with a corresponding 31% gain in F1-score. This effectively breaks the long-standing trade-off between coverage and accuracy typical of single-pass LLM generation, while also mitigating generation truncation caused by length limitations. At the same time, we emphasize that EVE exhibits performance saturation due to the inherent ambiguity of natural language, reflecting fundamental limits of language-based reasoning.

💡 Research Summary

The paper addresses a fundamental shortcoming of modern large language models (LLMs) when they are used for question‑answering over a single, moderately long document: the tendency to produce fluent but incomplete or hallucinated answers. In safety‑critical, legal, or autonomous‑system verification contexts, missing a single relevant fact or inventing unsupported content can have severe real‑world consequences. Existing approaches—free‑form prompting, Chain‑of‑Thought (CoT), Tree‑of‑Thought (ToT), Retrieval‑Augmented Generation (RAG), post‑hoc critics, and program‑aided reasoning—either focus on improving average performance or on reducing hallucinations through external evidence, but they do not provide systematic guarantees of completeness (recall) and faithfulness (precision) simultaneously.

To fill this gap, the authors propose EVE (Extraction‑Validation‑Enumerate), a modality‑agnostic, three‑stage pipeline designed specifically for closed‑world, single‑document settings where the answer must be exhaustively grounded in the source text.

1. Extraction Stage – The system issues (M_e) independent queries, each crafted from a different perspective (definition, example, relational cue, etc.), to the LLM. The union of all responses forms a candidate set (C). By deliberately over‑generating, this stage maximizes recall; the probability that a true element is missed drops exponentially with (M_e) (e.g., with per‑query success (p_{i,e}=0.6) and (M_e=4), the miss probability is ((1-0.6)^4≈0.025)).

2. Validation Stage – For every candidate, the framework sends (M_v) independent validation queries (e.g., “Is this fact present in the document?”). Responses are aggregated by majority voting. Assuming each validation query has an error rate below 50 % (a minimal sanity check for competent LLMs), Chernoff‑type bounds guarantee that the probability of an incorrect majority decision shrinks exponentially with (M_v). This stage filters out false positives, merges aliases, and yields a high‑precision set of verified entities.

3. Enumeration Stage – A higher‑level controller constructs a structured “skeleton” that enumerates the verified entities. For each entity, a focused explanation query (Q_f) is issued, conditioning on both the original document and the entity’s canonical representation. The resulting element‑wise paragraphs are assembled into a final report (R). Because enumeration is driven by an explicit algorithmic order rather than free‑form generation, truncation due to token‑length limits is avoided, and completeness can be formally checked (the skeleton must contain all slots).

The authors provide a theoretical analysis showing that the combined effect of multi‑query extraction and majority‑vote validation yields an overall error probability that decays exponentially with the number of independent queries.

Empirical Evaluation – To demonstrate practical impact, the authors introduce a new dataset for System‑Theoretic Process Analysis (STPA), a safety‑critical methodology that requires exhaustive identification of unsafe control actions. They apply EVE to several state‑of‑the‑art LLMs (including GPT‑4, Claude, and Llama‑2) and compare against standard single‑pass generation. Results show consistent improvements: recall increases up to 24 %, precision up to 29 %, and F1‑score up to 31 % relative to baselines. Moreover, the error‑reduction pattern matches the theoretical predictions, confirming that independent queries and voting indeed compress the error space.

Scope and Limitations – The paper explicitly limits EVE to single‑document, closed‑world tasks where the answer must be fully grounded in the provided text. It is not intended for open‑ended reasoning where a short prompt triggers extensive internal deliberation (the domain of CoT/ToT). The authors also acknowledge a performance ceiling caused by the inherent ambiguity of natural language; even with many queries, some uncertainty remains, suggesting future work should integrate structured metadata, external knowledge graphs, or domain‑specific validators to push beyond the observed saturation.

In summary, EVE offers a principled, theoretically grounded architecture that transforms a high‑variance, single‑pass generation problem into a sequence of low‑variance, verifiable steps. By doing so, it breaks the longstanding trade‑off between coverage and accuracy in document‑grounded QA, providing a concrete pathway toward more reliable, safety‑critical AI systems.