acceptodds
Under review as a conference paper at ICLR 2027

Beyond Flat Attribution: Recovering Boolean Evidence Formulas for Black-Box Language Models

Abstract

Post-hoc context attribution for closed LLMs asks which spans of a prompt support a claim the target asserts. Most black-box methods return flat per-span scores, which cannot express that either of two spans may satisfy the same evidence requirement. We formulate attribution as the recovery of a Boolean evidence formula , a monotone AND-of-OR representation for which flat span sets have a clause-averaged F1 ceiling of . LEAF builds a fixed set of candidate formulas from LLM-derived span roles and uses counterfactual queries to select among them. Under a registry-only answering instruction, we evaluate on synthetic Easy/Hard, two long-context extensions, and a curated 500-example fact-verification benchmark from HoVer, FEVER, and SciFact. Under the primary binary-feedback protocol, LEAF achieves Formula F1 of 0.83–0.99. Its margin over the strongest reported baseline ranges from +0.10 on Natural to +0.49 on the 100-span setting. The Hard result extends across four closed targets and Llama-3.3-70B. Held-out behaviour and weak-cartographer tests probe the roles of the metric and LLM prior.

Primary area:
interpretability and explainable AI
Keywords:
context attribution, interpretability, large language models, black-box methods, active learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_41zyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that flat, per-span context attribution cannot tell evidence that is jointly necessary from evidence that is interchangeable. It proposes representing attribution as a monotone CNF (“AND-of-OR”) evidence formula over prompt spans. Under a clause-matched Formula F1, a flat span set read as an AND of singletons is bounded by 1m∑r2/(1+wr)\frac{1}{m}\sum_r 2/(1+w_r), which is 0.83 for R∧(D1∨D2)R\land(D_1\lor D_2). The evaluation uses synthetic Easy (four phenomena), Hard (25 spans), and Long-50/100, plus a 500-example Natural set. The Natural set was curated from HoVer/FEVER/SciFact by keeping examples whose GPT-4o-mini behaviour matches BRIDGE ∧ (EV$_1$ ∨ EV$_2$). All runs use a registry-only answering instruction.

Critical strengths
  • Flat scores really cannot express that two spans are substitutes rather than joint requirements. The monotone CNF representation fits this naturally: each clause is a requirement, and its members are interchangeable alternatives. It also needs only text-API access. The flat-output ceiling is simple, but it states exactly why flat outputs lose under a structure-aware metric. This framing is the main reason I think the direction is worth pursuing.

  • Unusually candid diagnostic reporting. The paper includes: a seed-beam oracle audit, an arithmetic split of each gap around the flat ceiling, a negative cross-target result, etc.

Critical weaknesses
  • The benchmarks match the method’s built-in structure, so the gains do not show the value of query-based formula recovery. Every benchmark except Easy uses one gold template: R∧(D1∨D2)R\land(D_1\lor D_2) for Hard and Long-50/100, or its relabelling BRIDGE ∧ (EV$_1$ ∨ EV$_2$) for Natural. LEAF is built around exactly this shape. The cartographer’s role vocabulary names it (REDUNDANT_DATA is defined as “an alternative wording of the same fact”), and the first beam element is R1∧(D1∨X1)R_1\land(D_1\lor X_1). Three results suggest the hand-written beam and the LLM prior carry most of the gain, not the query loop: (a) The anchor is decisive. Without it, Easy-Redundant falls to ≈0.67 (Table 7 note). That is below BM25 (0.83) and Direct LLM (0.81), even though the beam still holds other OR candidates. No-anchor results on Hard and Natural are not reported. (b) The stress test stays inside the beam. Table 4 is described as using shapes the anchor does not match. But all three shapes are INITBEAM templates in Appendix G: R1∧(D1∨X1∨D2)R_1\land(D_1\lor X_1\lor D_2), (R1∨R2)∧(D1∨X1)(R_1\lor R_2)\land(D_1\lor X_1), and R1∧(D1∨X1)∧D2R_1\land(D_1\lor X_1)\land D_2. No experiment tests a gold structure outside the template list. (c) The Long-100 margin is baseline weakness. LEAF scores exactly the flat ceiling there (0.83; the above-ceiling margin in Table 2 is 0.00). So the whole +0.49 is headroom left by flat baselines that cannot ablate each of 100 spans even once at B=16B=16, while LEAF’s cartographer reads the full context. Disagreement acquisition also has no measurable effect on Easy (Table 7). As it stands, the evidence supports a narrower claim. An LLM prompted with this role vocabulary, plus a matched template beam, recovers this template. That falls short of a general method for recovering Boolean evidence structure.

  • The empirical claims cannot be checked, and several are internally inconsistent. The released code is a re-implementation written from the paper, and the authors state it did not produce the tables. Also withheld: (a) the curated Natural set; (b) the response cache and the judge canonicalisation rules; (c) per-example results, the held-out mask manifest, and the per-budget sweep. So none of the headline numbers can be verified. That would matter less if the reported results were consistent, but several are not: (a) Table 13. Direct LLM never queries the target, so on the Llama-refiltered set it must score the same in the GPT-4o-mini-evaluated and Llama-evaluated rows. Yet best-baseline reads 0.73 vs. 0.62. That implies Direct LLM ≤ 0.62 there and a target-querying flat baseline at 0.73, far above the ≤0.58 flat baselines reach on the original set. Also, with filter and evaluation target matched (Llama/Llama), LEAF scores 0.73. The paper’s explanation of Table 12 predicts a recovery toward 0.84, so this result undercuts it. (b) Appendix E. At T=0.3T=0.3 over three seeds, LEAF has exactly zero SD at every budget, while Direct LLM (same planner model) varies by ±0.02. With content-hash caching, this suggests cached responses may have been replayed across seeds. (c) Table 6. It is not a strict held-out test, since the mask manifest was not kept. In addition, §5.4 compares LEAF with Flat MAB (0.72) on Hard B=8B=8, but BM25 (0.75) is the best baseline in that column. (d) Point estimates only. Table 5, the shape test, and the continuous-feedback rerun have no intervals.

Questions for authors
  1. Zero-query control. What Formula F1 does the MAP-from-prior formula (no target queries, the actual GPT-4o-mini cartographer) reach on Easy, Hard, Long-100 and Natural? If it is close to LEAF, the contribution is the cartographer plus templates rather than the validation loop.

  2. Out-of-beam structures. Please report the no-anchor ablation on Hard and Natural. Please also test gold shapes that are not among the INITBEAM templates, such as three disjoint OR clauses or overlapping clauses.

  3. Table 13. Please give per-method scores for both Llama-refiltered rows. How can best-baseline differ between them if Direct LLM is target-independent? Why does Llama/Llama not recover toward 0.84?

  4. Cache sharing. Were the seven filtering-stage target responses and LEAF’s evaluation queries served from the same content-hash cache? If so, overlapping masks (full prompt, empty prompt, leave-one-out) mean LEAF observes the very responses that defined gold. Does the cache key include temperature and seed?

  5. Selection rule. With binary yy and σ0=0.15\sigma_0=0.15, one mismatch costs ≈22 nats, far more than τ=0.5\tau=0.5. Is LEAF therefore effectively version-space elimination over 16 candidates plus an Occam rule? How are ties between equal-clause formulas broken? If by beam order, the OR-anchor wins whenever observations do not discriminate.

  6. Will you release the code that produced the tables, the Natural benchmark with gold, and the response cache? Why was a separate reference implementation written?

Reviewer_8qyCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses a meaningful limitation of conventional context attribution: flat importance scores cannot distinguish evidence that is jointly necessary from evidence that is interchangeable. The proposed LEAF framework represents evidence dependencies using monotone Boolean formulas and selects among candidate formulas through black-box model queries.

Critical strengths

The distinction between flat attribution and structured evidence dependencies is meaningful. Conventional span scores cannot directly express whether two passages are redundant alternatives or jointly necessary.

Critical weaknesses

LEAF's candidate formulas are generated before any target-model validation, and the beam remains fixed throughout the algorithm. Counterfactual queries only change candidate weights. This creates a fundamental question: How much does the target model actually contribute to the recovered evidence structure?

Benchmark construction may favor the proposed hypothesis class and candidate initialization

Questions for authors

What is LEAF's Formula F1 when selecting the MAP candidate from the cartographer prior without any target queries? How much improvement does posterior-disagreement acquisition provide over random and fixed-mask selection when all methods use the same candidate beam?

Reviewer_YgHnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper reformulates black-box context attribution as recovering a Boolean evidence formula rather than assigning independent importance scores to context spans. The proposed monotone AND-of-OR representation distinguishes jointly required evidence from substitutable alternatives. LEAF uses an LLM cartographer to assign span roles and construct a fixed set of 16 candidate formulas, then selects among them through masked-context queries, likelihood-based reweighting, and disagreement-based query acquisition.

Experiments cover synthetic examples, longer contexts, a curated 500-example fact-verification benchmark, and multiple target models. The paper reports strong Formula F1 results, supplemented by span-level evaluation, behavioural agreement, candidate-oracle analysis, and ablations.

Critical strengths

Useful problem formulation. Flat importance scores cannot explicitly distinguish evidence that is jointly required from evidence that can substitute for another span. The Boolean representation captures this distinction clearly.

Practical black-box interface. LEAF requires only text responses rather than gradients, hidden states, or token probabilities. Separating a formula proposer from a target-model validator is a reasonable design for closed-model attribution.

Critical weaknesses
  1. The contribution of target querying is not isolated from the candidate-construction prior

LEAF constructs its candidate formulas before any target queries and cannot introduce new formulas afterward. Table 9 shows that the initial beam already contains the exact gold formula for 99% of Easy, 94% of Hard, and 75% of Natural examples. This establishes that candidate construction is strong, but does not show how much improvement comes from observing the target.

The most important missing control is the formula selected from the same cartographer output and candidate beam without target validation. The manuscript explicitly acknowledges that this result is unreported. Direct LLM is a useful structured baseline, but it produces one formula through a different prompting procedure; it is not equivalent to selecting the highest-prior formula from LEAF’s initialized beam.

Please compare the same candidate set under prior-only selection, random query acquisition, simple greedy verification, and the proposed disagreement strategy, holding the target-query budget fixed. The existing acquisition ablations provide partial evidence, but do not replace the prior-only control.

This comparison is particularly important because many existing baselines are converted to conjunctions of singleton clauses and consequently face the paper’s metric-specific ceiling. The continuous-feedback rerun is helpful, but a matched structured comparator is needed to establish the algorithmic benefit beyond the output representation and initialization.

  1. The reported empirical results are not independently reproducible from the supplied materials

The reproducibility statement makes an important distinction: the supplementary implementation was written from the paper description and is not the code that produced the reported tables. The curated Natural benchmark, response cache, natural-judge canonicalization rules, and per-example results are also not included.

Providing a reference implementation is useful, but it cannot establish reproducibility of the reported measurements, confidence intervals, or behavioural evaluation. This is particularly consequential because the natural benchmark is newly curated and the evaluation depends on historical API responses and implementation-specific judgment rules.

Please provide an anonymized evaluation package containing the benchmark source identifiers, span boundaries, gold formulas, intervention masks, scoring rules, per-example predictions, and scripts reproducing the main tables. Where redistribution of full responses or source text is restricted, derived evaluation records and a reconstruction procedure would still materially improve verifiability. The manuscript should clearly distinguish reference-pipeline reproducibility from reproduction of the reported experiments.

Questions for authors

What is the performance of selecting a formula from LEAF’s initialized beam without any target queries?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.