acceptodds
Under review as a conference paper at ICLR 2027

When Evidence is Missing: Benchmarking Hallucination and Abstention from Incomplete Observations

Abstract

In real-world settings, document-grounded LLM agents must successfully navigate incomplete corpora, whereas most benchmarks assume all evidence is available. We introduce a scenario-agnostic simulation framework that generates document corpora based on a structured latent world. The framework maintains full provenance for all facts, their inference chain, and in which documents they are surfaced, allowing arbitrary subsets of documents to be masked and exactly recomputing which facts remain observable and inferable when certain evidence is missing. We instantiate our framework on four human-validated enterprise scenarios, each producing approximately 1,000 documents. Based on the fact-to-document provenance, we construct question-answering and structured prediction tasks requiring retrieval, multi-document synthesis, logical inference, and abstention when evidence is missing. Evaluating nine LLMs on our tasks using RAG and CLI agent harnesses, we find that models most accurate at answering when evidence is present differ from those that abstain most reliably when it is missing.

Primary area:
datasets and benchmarks
Keywords:
Synthetic Data Generation, LLM Evaluation, LLM Agent Benchmark, Question-Answering, Structured Prediction, Information Retrieval, Abstention, Hallucination

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Pp5TOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a synthetic benchmark for studying hallucination and abstention under missing document evidence. It builds structured latent worlds, renders facts into documents, masks evidence, and recomputes answerability using fact provenance and inference dependencies. The benchmark covers four enterprise scenarios and evaluates nine LLMs with RAG and CLI agents. The main finding is that strong performance on answerable questions does not imply reliable abstention when evidence is missing.

Critical strengths
  1. The provenance-aware framework can mask documents and automatically recompute which questions remain answerable, enabling systematic study of abstention under incomplete evidence.
  2. The paper evaluates multiple recent models under both RAG and CLI settings and across QA and structured prediction tasks, providing a useful empirical comparison of answering and abstention behavior.
Critical weaknesses
  1. The framework novelty appears limited to me. The ideas of synthetic lated world, multihop inference graphs, evidence masking are already well established. And importantly, the paper does not empirically compare against other factuality/hallucination benchmarks (incomplete-information QA, controlled-evidence-removal). It is unclear whether this benchmark reveals qualitatively new failure modes or model rankings beyond what existing benchmarks already capture.
  2. The paper’s central claim is that provenance and inference dependencies enable exact recomputation of answerability after document masking. However, this exactness is defined relative to the authored symbolic specification and its inference rules. The generated natural-language corpus may contain implicit, redundant, or commonsense evidence that is not represented in the simulator’s dependency graph. Consequently, an item labeled unanswerable by the framework may still be answerable from the visible documents through a valid inference path not encoded in the specification. This concern is acknowledged indirectly by the authors, who note that they do not establish whether humans could answer every item labeled unanswerable from the remaining evidence. This substantially weakens the claim of “exact” evidence control.
  3. The paper validates structural consistency of the specification and simulator, and asks domain experts to review each specification and one generated lane. However, it does not systematically measure whether the generated latent worlds are semantically realistic, whether the fact ontology is sufficiently complete, or whether the authored inference rules capture the valid inferences available in the domain. Since downstream answerability labels depend directly on this symbolic world, the lack of a stronger quality evaluation weakens the benchmark’s validity.
  4. nit: figure 1 is difficult to understand
Questions for authors
  1. How do the authors validate the semantic completeness of the latent world and inference rules, beyond structural checks and limited expert review? How do we measure the quality of the pipeline.
  2. Have the authors tested whether humans can answer questions labeled unanswerable from the remaining visible documents?
  3. How does this benchmark compare empirically with the closest abstention or controlled-evidence-removal benchmarks, and what unique failure modes does it reveal? Also methodology like MiniCheck (Tang et al., 2024)
Reviewer_brNvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a simulation framework for building document-grounded QA benchmarks in which answerability is computed rather than annotated. A spec sheet defines entities, a probabilistic process model, event templates, fact predicates and inference rules. The engine samples a latent world of process lanes, forward-chains inferred facts, and an LLM renders the base facts into emails, letters, reports and spreadsheets while the framework records which document surfaces which fact. When documents are masked, a base fact stays visible if any unmasked document surfaces it and an inferred fact stays visible only if all its prerequisites do; a QA item whose matching facts are all invisible gets "abstain" as its gold answer. Four enterprise scenarios (M&A, employee lifecycle, sales, VC fundraising; about 1,000 documents each, roughly 20% masked) yield 599 QA items (463 answerable, 136 unanswerable) and 375 structured-prediction tasks (Synthesis, Process Reconstruction, Event Reconstruction). Nine models are evaluated under a RAG harness and a shell-based CLI harness with deterministic scoring. Validation consists of programmatic checks, expert review of the spec sheets, and a five-annotator study of whether 200 documents support their assigned facts (77.6% recall, 98.9% distractor rejection, Fleiss' κ = 0.765). The main reported finding is that the models with the highest answerable accuracy (Qwen3.6-27B under CLI, GPT-OSS-120B under RAG) are not the ones that abstain best (Opus 5 under CLI, Sonnet 5 under RAG), together with an oracle-evidence ablation, a relaxed-visibility rescoring, and low F1 on Process and Event Reconstruction for all models.

Critical strengths
  1. The framework makes answerability a computed property of the visible corpus. Pairing a question with a context that lacks its support is not new in itself: MuSiQue-Full (Trivedi et al., 2022) builds an unanswerable twin for every answerable question by making sure one sub-answer appears in no context paragraph, following the contrastive support sufficiency idea of Trivedi et al. (2020), and the paper cites both. What this paper adds is doing it over a corpus of about 1,000 documents in which a fact can be surfaced by several documents and an inferred fact depends on a chain of other facts. Under any mask the gold answer is recomputed from provenance, with no re-annotation. MissDiag (Wang et al., 2026a), the closest recent work the paper cites, removes knowledge-graph support but keeps the question and gold answer fixed. A generator that can produce answerable and unanswerable versions of the same question at corpus scale, for any masking policy, is the part of the paper I expect others to reuse.

  2. The evaluation protocol avoids several easy mistakes. Scoring is deterministic with no LLM judge. Answerable and unanswerable items are reported separately, so a policy that always abstains cannot hide behind an aggregate, and terminal failures are kept apart from abstentions. That last choice matters in practice: Qwen3.6-27B under RAG has the lowest hallucination rate of the three open models only because 56.6% of its unanswerable runs end in a terminal failure (turn limit, context exhaustion or no submission; Table 2, Table 11), and the paper says so instead of counting it as abstention. The oracle-evidence and relaxed-visibility ablations (Tables 5 to 7) and the empty/non-empty split of Synthesis (Table 9) show the authors probing their own numbers, and the human study includes negative-control documents. The paper states that all evaluation outputs are released, so the checks I ask for below can be run on existing logs.

Critical weaknesses
  1. The hallucination metric assumes missing evidence without adequately validating that assumption. The paper explicitly says its human study does not establish whether questions labeled unanswerable can be answered from the remaining corpus (p. 9). Checking that assigned facts appear in their documents does not check that other documents reveal them. The 98.9% distractor-rejection result is per document with an unspecified sampling procedure; it does not cover cross-document leakage or process-based inference. For example, an unmasked later-stage M&A document could imply an earlier NDA despite that fact's absence from the recorded provenance. Relaxing visibility inside the same specification cannot detect evidence outside it. Using one mask per scenario also compares different questions across the answerable/unanswerable groups. The framework could support paired evidence-removal tests, but the current evidence does not isolate missing-evidence hallucination from leakage and question difficulty.

  2. The answer-scoring pipeline is insufficiently validated for the capability conclusions. Opus 5 CLI scores 0.432 on answerable QA without terminal failures; supplying only supporting documents lowers it to 0.421 while increasing false refusal (Tables 5-6). GPT-5.6 Sol CLI reaches 0.540, also without QA terminal failures. Without inspecting incorrect answers, it is unclear how much error comes from reasoning rather than gold labels, answer matching or formatting. Structured prediction has the same issue: gold keys use specification identifiers while prompts expose column headers, so equivalent names may fail alignment. This alone does not explain Opus 5's exactly zero Event Reconstruction precision/recall despite nonzero Process Reconstruction results, and the failure is not analyzed. Until these errors are categorized, the answer-versus-abstention dissociation and uniformly difficult structured-prediction conclusions remain hard to interpret.

Questions for authors
  1. For each model and harness, what fraction of the non-abstaining submissions on unanswerable items match the true latent-world answer? Relatedly, for how many of the 136 unanswerable items does the latent answer string (entity name, value or date) still occur in some visible document? Both can be computed from the released corpora and outputs, and they are the numbers most likely to change my reading of the Hallucination column.

  2. Could you categorise a sample of the answerable items that Opus 5 (CLI) and GPT-5.6 Sol get wrong, say 50 each, into a wrong value, a correct value rejected by the matcher, and a gold label that is not supported by the visible documents? The same for the Opus oracle run would help, since retrieval is no longer the bottleneck there. Do the accepted-answer lists for numeric gold answers include scaled forms such as "$5.3 billion"? Under Appendix B.2, numeric equality is the only lenient rule for numeric gold and such strings do not parse.

  3. Figure 1 sketches one inferred item (time from LOI to close). Could you show two or three more inferred-fact QA items at depth D1 and D2 with their gold answers and the rule behind each? Is the rule that defines the answer (for example what counts as a "superseded bid") stated or implied in the question or the corpus, or must the agent guess the spec's definition?

  4. Were the admissible process names, event template names and categorical values available to the agents in any form? How many tasks are in Process Reconstruction and in Event Reconstruction? The totals (375 SP, 297 Synthesis, 74 lanes) suggest 74 and 4, and the Event examples in Appendix E.2.3 are scenario-wide tables of 326 to 505 rows; if there are four Event tasks, its column in Table 3 is an average of four numbers. What is the key for Process Reconstruction? The Employee example on p. 24 has two performing rows, and B.3 deduplicates predicted rows by key, keeping the first. Finally, what did Opus 5 submit on the Event tasks that produced .000?

  5. Which masking strategy (window or random) and which parameters produced the reported corpora? Which LLM rendered the documents, and is it from one of the evaluated model families?

  6. How were the three distractor statements per document sampled in the human study (same lane and predicate, or arbitrary)? What is the recall under majority vote of the five annotators, and were pairs selected by exactly two annotators audited?

Suggestions that I do not expect in the rebuttal: the framework would support a much more informative analysis than a single leaderboard, for instance hallucination rate as a function of masking fraction, window versus random masking, and paired comparisons of the same question with and without its evidence. With those, the dissociation in Table 2 could become a finding about how models behave as evidence disappears; at present it is a ranking.

Reviewer_Ai3aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a benchmark that evaluates whether document-grounded LLM agents can answer questions from incomplete corpora and abstain when evidence is missing. To build the benchmark, the framework first defines entities, process models, event templates, facts, and inference rules in a specification sheet. A simulation engine then uses the sheet to generate a latent world, and an LLM renders the facts into documents. The framework records which documents surface each fact and which facts each inference depends on. As a result, documents can be masked arbitrarily, and the answerability of every task is recomputed automatically. The benchmark covers four enterprise scenarios with QA and structured prediction tasks. The paper evaluates nine LLMs under RAG and CLI harnesses and finds that the models most accurate on answerable questions are not the ones that abstain most reliably.

Critical strengths
  • The paper addresses an important and underexplored problem, since real-world document agents often face incomplete corpora while most benchmarks assume all evidence is available.
  • Simulating a latent world with full fact-to-document provenance is a novel and sound design. It allows us to construct tasks flexibly and automatically, with controllable difficulty.
  • The evaluation covers nine models and two harnesses with deterministic scoring, and the divergence between answer accuracy and abstention reliability is an informative finding.
Critical weaknesses
  • The quality and difficulty of the benchmark depend largely on the spec sheets, yet the paper does not adequately describe how they are constructed. This raises serious concerns about reproducibility.
  • The validation does not guarantee label correctness. Unanswerable is defined relative to the authored inference rules, and the authors do not check whether humans could answer these questions from the remaining evidence. Moreover, annotators identify only 77.6% of the assigned facts in the documents, suggesting that some "answerable" labels may be incorrect.
  • The results do not exhibit the expected scaling trend, as stronger models do not achieve higher scores on this benchmark. However, the paper does not provide in-depth analysis or explanation for this phenomenon.
Questions for authors

What is the cost of constructing spec sheets? How do the authors ensure diversity and rationality of spec sheets?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.