PaperlayerView exact revision on arXiv

Source-linked AI summary

ReCEval: Evaluating Reasoning Chains via Correctness and Informativeness

Archiki Prasad, Swarnadeep Saha, Xiang Zhou, Mohit Bansal

arXiv:2304.10703v2cs.CLcs.AIcs.LG

TL;DR

Answer-focused evaluation can miss whether a reasoning chain is valid and informative, while reference-based evaluation has practical constraints. ReCEval scores chains as informal proofs using correctness and informativeness metrics, and reports improved error detection and downstream performance across multiple datasets.

  • Problem

    Answer-only evaluation may confound reasoning quality with unfaithful or spurious shortcuts, while reference-based evaluation depends on human-written chains that can be costly and incomplete.

    Source: §1 · pp. 1–2

  • Method

    ReCEval evaluates reasoning chains as informal proofs using reference-free correctness and informativeness metrics built from NLI models and V-information.

    Source: §1 · p. 2

  • Results

    ReCEval has the highest correlation for 5 of 6 Entailment Bank error types and improves overall-quality correlation on GSM-8K and DROP.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Selecting chains using both correctness and informativeness improves QA accuracy by 3.2% over greedy decoding, compared with up to 1.7% for ROSCOE metric combinations.

    Source: §6.3 · p. 9

    The metrics may be less suitable when correctness depends on substantial implicit knowledge not present in the input or intermediate steps.

    Source: Limitations · p. 10

Abstract

from arXiv · show

Multi-step reasoning ability is fundamental to many natural language tasks, yet it is unclear what constitutes a good reasoning chain and how to evaluate them. Most existing methods focus solely on whether the reasoning chain leads to the correct conclusion, but this answer-oriented view may confound reasoning quality with other spurious shortcuts to predict the answer. To bridge this gap, we evaluate reasoning chains by viewing them as informal proofs that derive the final answer. Specifically, we propose ReCEval (Reasoning Chain Evaluation), a framework that evaluates reasoning chains via two key properties: (1) correctness, i.e., each step makes a valid inference based on information contained within the step, preceding steps, and input context, and (2) informativeness, i.e., each step provides new information that is helpful towards deriving the generated answer. We evaluate these properties by developing metrics using natural language inference models and V-Information. On multiple datasets, we show that ReCEval effectively identifies various error types and yields notable improvements compared to prior methods. We analyze the impact of step boundaries, and previous steps on evaluating correctness and demonstrate that our informativeness metric captures the expected flow of information in high-quality reasoning chains. Finally, we show that scoring reasoning chains based on ReCEval improves downstream task performance. Our code is publicly available at: https://github.com/archiki/ReCEval

1 Introduction

Answer-only evaluation can miss whether a reasoning chain is valid and useful, while reference-based evaluation depends on costly, potentially incomplete gold chains. ReCEval evaluates chains as informal proofs and reports stronger error detection and downstream performance.

1 Introduction

  • Answer-only scores may conflate reasoning ability with unfaithful or spurious shortcuts, motivating intrinsic evaluation of reasoning-chain quality.Source: §1 · p. 1
  • Reference-based evaluation can be costly and depend on incomplete or non-unique gold chains; ReCEval extends reference-free evaluation beyond generic error metrics.Source: §1 · pp. 1–2
  • ReCEval treats chains as informal proofs and defines quality through valid within-step and context-based inference, plus steps that add helpful information toward the answer.Source: §1 · p. 2
  • Its reference-free metrics use NLI models and V-information, and the paper reports improved downstream task performance from selecting chains by ReCEval scores.Source: §1 · p. 2
  • On Entailment Bank, ReCEval has the highest correlation for 5 of 6 error types, raising hallucination correlation from 0.62 to 0.89; it also improves overall-quality correlation on GSM-8K and DROP.On GSM-8K the reported correlation rises from 0.28 to 0.36, and on DROP from 0.19 to 0.22; ReCEval excels on 5 of 7 error types across those datasets.Source: §1 · p. 2

2 Reasoning Chains: Preliminaries

The paper represents a reasoning chain as steps made up of fine-grained claims, then uses V-information as an information-theoretic measure. These representations support detailed analysis of reasoning and information flow.

2 Reasoning Chains: Preliminaries

  • A reasoning chain is a multi-step rationale used to arrive at a predicted answer, given an input context.Source: §2 · p. 2
  • Reasoning Content Units (RCUs) are the claims within a step; a step can be divided into premise RCUs and a conclusion RCU.The paper describes the units as supporting fine-grained analysis and verification of reasoning abilities.Source: §2 · pp. 2–3
  • Pointwise V-information estimates usable information in an individual input-output pair, and conditional PVI measures it relative to an instance.Source: §2 · p. 3

3 Properties of Good Reasoning Chains

The paper defines a good reasoning chain by two complementary properties: each step must be correct in relation to its premises and prior context, and informative toward the answer.

3 Properties of Good Reasoning Chains

  • Correctness requires each step's conclusion to be correct, both from the step's premise units and in light of the input and earlier steps.The first aspect captures consistency within a step; the second captures global consistency.Source: §3 · p. 3
  • Informativeness measures how much a step contributes toward the final answer, since plausible inferences can differ in relevance.Source: §3 · p. 3
  • A step can be correct yet redundant: the illustrated third step adds no progress beyond the second, making informativeness useful for detecting repetition.Source: §3 · p. 3

4 RECEVAL: Evaluation Metrics

ReCeval operationalizes reasoning-chain quality through step-level correctness and informativeness metrics, then aggregates those scores across the chain. Its metrics test local inference, consistency with prior information, and added information toward the answer.

4.1 Evaluation of Intra-Step Correctness

  • Intra-step correctness scores whether a step's conclusion is entailed by its premise RCUs; a relaxed PVI-based alternative measures how easily the conclusion follows.Strict entailment assigns low probability to neutral conclusions, accounting for unsupported claims; the relaxed criterion allows informal steps with omitted premises.Source: §4.1 · p. 4

4.2 Evaluation of Inter-Step Correctness

  • Inter-step correctness checks whether a new conclusion contradicts the input or earlier conclusions, using an NLI model to estimate contradiction probability.The method excludes earlier premise RCUs because they overlap with the input context; the authors report that excluding them does not affect performance.Source: §4.2 · p. 4

4.3 Evaluation of Informativeness

  • Informativeness uses conditional PVI to measure the information gain from adding a step, with low gain indicating that a step may be redundant.The score conditions on preceding steps and measures information toward the predicted answer.Source: §4.3 · p. 4

4 RECEVAL: Evaluation Metrics

  • The framework combines step-level metrics into chain-level scores intended to help identify different error types.Source: §4.4 · pp. 4–5

4.4 RECEVAL: Overall Algorithm

  • The overall algorithm identifies premise and conclusion RCUs from semantic-role frames, then evaluates their correctness and informativeness.It uses the minimum step score for each metric, reflecting the assumption that a chain is only as good as its least correct or informative step.Source: §4.4 · pp. 4–5

5 Meta-Evaluation Setup

The meta-evaluation tests whether metric scores track reasoning-chain errors across three datasets, comparing ReCEval with text-generation, model-based, and ROSCOE baselines. It measures association with error status using Somer’s-D correlation.

5 Meta-Evaluation Setup

  • The evaluation detects errors by correlating metric scores with ground-truth annotations for each error category.When multiple metric variants are available, the highest correlation is reported.Source: §5 · p. 5

5.1 Meta-Evaluation: Datasets

  • Entailment Bank, GSM-8K, and DROP provide the evaluation datasets, with EB perturbations and human judgments covering different reasoning errors.EB includes regular and challenge perturbations; GSM-8K and DROP chains can contain multiple errors.Source: §5.1 · p. 5

5.2 Meta-Evaluation: Baselines

  • Baselines include text-generation and model-based metrics, plus ROSCOE metrics for semantic similarity, alignment, logical inference, and redundancy.The ROSCOE comparisons are grouped by correctness and informativeness properties.Source: §5.2 · p. 5

5.3 Meta-Evaluation: Correlation Measure

  • Somer’s-D correlation measures association between a metric score and binary error status, using Kendall’s τ; the highest variant is reported when available.The paper defines D_SE = τ(E, S)/τ(E, E).Source: §5.3 · p. 5

6 Results and Discussion

The section presents meta-evaluation results on Entailment Bank, GSM-8K, and DROP. For EB-challenge, it reports ReCEval outperforming text-generation baselines.

6.1 Effectiveness of RECEVAL

  • The main meta-evaluation results cover Entailment Bank, GSM-8K, and DROP.Source: §6.1 · p. 5
  • Table 1 presents EB-challenge results across different error types.Source: §6.1 · p. 5
  • On EB-challenge, ReCEval metrics outperform text-generation baselines on both reported dimensions.Source: §6.1 · p. 5

Metric HALL NEG SWAP

Across benchmark evaluations and analyses, ReCEval’s metrics identify reasoning-chain errors, while design choices about reasoning units, prior context, and step boundaries affect correctness evaluation.

Metric HALL NEG SWAP

  • 0.62→0.89 on hallucination versus ROSCOE on EB; ReCEval also beats text-generation metrics on GSM-8K and DROP across many annotated errors.On GSM-8K, ReCEval has higher correlations across all error types than text-generation metrics; on DROP, it performs similarly or better than ROSCOE on overall quality and coherence.Source: Table 1 · p. 6, Metric HALL NEG SWAP · p. 6, Table 2 · p. 6

6.2 Analysis of RECEVAL Metrics

  • RCU decomposition enables accurate identification of hallucinations and swap errors; gold RCUs improve correctness correlations by up to 0.20.Identified RCUs also strengthen correctness evaluation, though a gap remains relative to gold RCUs.Source: §6.2 · p. 7, Table 4 · p. 7
  • Using only the immediately preceding one or two steps decreases hallucination and negate-error correlations by up to 0.11, supporting evaluation against all prior steps.The comparison varies the number of preceding steps included in inter-step correctness evaluation.Source: §6.2 · p. 7, Table 5 · p. 7
  • Sentence-level step boundaries work effectively in practice; treating each RCU or the entire chain as a step lowers correlations on some or all errors.Whole-chain boundaries lower correlations across all errors, while RCU-level boundaries lower them for hallucination and swap errors.Source: §6.2 · p. 7, Table 6 · p. 7

How does informativeness vary across steps?

  • 72% of gold chains have positive information gain at every step, versus 38% of uninformative chains; API also finds positive gains across two or three consecutive steps in 87% and 92% of gold chains.API measures whether each group of k consecutive steps is more informative than the preceding step.Source: How does informativeness vary across… · p. 8, Table 7 · p. 8
  • Trained PVI models achieve correlations at least 0.05 higher than pretrained GPT-2 XL, while a larger pretrained model such as LLaMA-7B achieves the highest correlation in the comparison.The paper proposes pretrained-LM log-likelihood as an alternative when gold reasoning chains for training are unavailable.Source: How does informativeness vary across… · p. 8, Table 8 · p. 8

6.3 Utilizing RECEVAL for Evaluating and Improving Downstream Tasks

  • A small study on 50 Entailment Bank chains shows that prompted GPT-3.5-turbo can provide scores for correctness and informativeness in ReCEval.The study prompts the model for entailment and answer-generation probabilities.Source: §6.3 · p. 8, Table 9 · p. 8
  • 3.2% higher QA accuracy than greedy decoding results from selecting chains using both correctness and informativeness; ROSCOE metric combinations improve accuracy by up to 1.7%.Using correctness alone improves accuracy by 2.3%, and informativeness alone by 1.4%.Source: §6.3 · p. 9, Table 10 · p. 9

7 Related Work

Prior work uses text-generation metrics, gold reasoning chains, and reference-free reasoning-chain metrics, each with limitations for evaluating coherent, multi-step reasoning. ReCEval formalizes desired chain properties and proposes reference-free metrics that outperform ROSCOE across datasets.

  • Traditional text-generation metrics are inadequate for evaluating coherent reasoning chains and are unsuitable for reference-free evaluation when they rely on references.Source: §7 · p. 9
  • Dataset-specific proof or reasoning-tree metrics can be less generalizable, while gold-chain approaches depend on reference coverage.Examples include symbolic proofs evaluated against gold proofs and comparisons to gold reasoning trees.Source: §7 · p. 9
  • ReCEval builds on ROSCOE by formally defining correctness and informativeness and proposing reference-free RCU- and V-information-based metrics that outperform ROSCOE across datasets.Source: §7 · p. 9

8 Conclusion

ReCEval evaluates reasoning chains through correctness and informativeness, using reference-free metrics based on entailment, PVI, and granular claims. The conclusion reports stronger meta-evaluation than prior baselines and downstream task-performance improvements.

  • ReCEval considerably outperforms previous baseline metrics in meta-evaluation across multiple datasets and leads to downstream task-performance improvement.Source: §8 · p. 9
  • The framework evaluates reasoning chains for correctness and informativeness using reference-free metrics based on entailment, PVI, and Reasoning Content Units.Source: §8 · p. 9
  • The conclusion reports detailed analyses of the metrics and their effectiveness across varied evaluation settings.Source: §8 · p. 9

Limitations

The supplied appendix passages describe implementation and dataset details alongside limitations of ReCEval. They note assumptions about explicit knowledge and variability in informativeness scores across problems.

Limitations

  • ReCeval may be poorly suited to correctness judgments requiring substantial implicit knowledge, and generated chains may not faithfully reflect a model’s internal reasoning.The approach assumes needed knowledge is present in the input or prior steps; whether chains faithfully represent internal reasoning remains open.Source: Limitations · p. 10

A RECEVAL: Background and Details

  • RCUs are extracted with semantic-role frames, selected as nonoverlapping units, then classified as premise or conclusion using sentence position and conjunction rules.For “because” or “since,” the RCU following the conjunction is treated as the premise.Source: A RECEVAL: Background and Details · p. 14
  • PVI models are fine-tuned on gold chains, while info-gain models predict the answer from partial chains and preceding steps.T5-large is used as the predictive model family; the GSM-8K answer target combines the question and predicted answer.Source: A RECEVAL: Background and Details · pp. 14–15
  • PVI-based metrics can be positive, negative, or zero, while info-gainPVI values may vary substantially across different reasoning problems.The authors identify score normalization and cross-problem variability as future-work directions.Source: A RECEVAL: Background and Details · p. 15

B.1 Entailment Bank

  • Entailment Bank perturbations target reasoning content absent from the input, reducing shortcuts from overlap between the input and gold chains.Hallucinations are drawn from distractors, and swap and repetition errors are introduced in intermediate reasoning-tree nodes.Source: §B.1 · p. 16

B.2 GSM-8K

  • GSM-8K uses human-annotated chains, where unlike EB-challenge, multiple errors may co-occur in a chain.The cited collection of GSM-8K chains comes from Golovneva et al. (2023).Source: §B.2 · p. 17

C Additional RECEVAL Meta-Evaluation

The additional meta-evaluation reports that ReCEval outperforms baselines on correctness errors, while results vary by perturbation setting and error type. ROSCOE performs better on REP errors, which are rare in the evaluated data.

C Additional RECEVAL Meta-Evaluation

  • On EB-Regular perturbations, ROSCOE’s HALL, NEG, and SWAP correlations are much higher than when shortcut overlap is removed.Text-generation metrics’ HALL correlations also decrease when spurious shortcuts are removed.Source: C Additional RECEVAL Meta-Evaluation · p. 17
  • ReCEval outperforms baselines on correctness errors, excluding grammar and missing-step errors from this comparison.Missing steps are excluded because they involve a confounder and are hard to evaluate reference-free.Source: C Additional RECEVAL Meta-Evaluation · p. 17
  • Grammar issues that do not change correctness can be measured with grammar-checking metrics such as ROSCOE-LC.Source: C Additional RECEVAL Meta-Evaluation · p. 18
  • ROSCOE outperforms ReCEval on REP errors in GSM-8K, but REP is rare and label imbalance can spuriously associate it with ROSCOE-LI’s overall coherency score.The passage attributes the spurious correlation to label imbalance, not to a causal effect of REP errors.Source: C Additional RECEVAL Meta-Evaluation · p. 18

D RECEVAL Correctness Metrics

The correctness analyses compare metric variants and views of correctness, while the informativeness analysis examines information gain across steps. The reported findings show that the best correctness view depends on error type and that using two preceding steps improves information-gain correlations.

D RECEVAL Correctness Metrics

  • Concatenation and pair-wise inter-correct variants perform comparably across error types, with similar results when premise-RCUs are included.The variants are compared on the EB-challenge development split.Source: D RECEVAL Correctness Metrics · p. 19, Table 17 · p. 19
  • The best correctness view depends on error type: intra-correctentail and intra-correctPVI complement one another, while inter-correctPVI favors hallucinations and swaps and no-contradiction favors negations.Inter-correctentail correlates well across error types at different values of k.Source: D RECEVAL Correctness Metrics · pp. 19–20

E Informativeness and Approximately Positive Information Gain (API)

Related papers

Loading 2304.10703v2…