PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating Explanation-Driven Vision-Language Reasoning via Generation Order Interventions

Siting Liang, Luca Rippe, Omar Adjali, Daniel Sonntag

arXiv:2609.29496v1cs.CL

TL;DR

Prior explanation-driven vision-language work often treats rationales as post-hoc justifications, leaving their connection to model decisions uncertain. This paper tests that connection by intervening on answer–rationale order and evaluating prediction consistency across tasks. It finds that generation order and model behavior vary across settings, while explanation-supervised fine-tuning improves order consistency.

  • Problem

    Prior work has largely overlooked how generating explanations before or after answers affects model behavior, despite explanations being used to expose and assess reasoning.

    Source: §1 · p. 1, §6 · p. 6

  • Method

    The study intervenes on answer–rationale order in a controlled single-response setup and measures task performance and prediction consistency across settings.

    Source: §1 · p. 2, §3.3 · p. 4

  • Results

    Qwen-VL shows stronger reasoning but remains order-sensitive, while LLaVA is weaker and can show positional bias; explanation-supervised fine-tuning substantially improves order consistency.

    Source: §4.2 · pp. 4–5, §6 · p. 6

  • Takeaways & Limitations

    Prediction agreement across generation orders provides a diagnostic of whether explanations are causally grounded in model decisions.

    Source: §1 · p. 2, §3.3 · p. 4

    The human-written e-SNLI-VE explanations are not designed as fine-grained rationales tightly grounded in the decision process.

    Source: §3.2 · p. 3

Abstract

from arXiv · show

Natural language explanation generation serves as a key mechanism for exposing and evaluating vision-language reasoning. Prior work on explanation-driven vision-language models predominantly follows a post-hoc (answer-first) paradigm, implicitly suggesting that supervised rationales can reflect underlying reasoning processes. In contrast, modern large vision-language models increasingly exhibit a rationale-first generation tendency, which more closely aligns with structured, stepwise reasoning. In this work, we systematically evaluate whether explanations are causally tied to model predictions within a single generation step under a controlled experimental setup, explicitly eliminating unnecessary chain-of-thought or other intermediate reasoning processes across knowledge-intensive QA, visual entailment, and compositional grounding benchmarks. We find that larger models emerge as a prerequisite for reliably supporting rationale-first reasoning at scale. However, answer-first generation is less prone to format-related errors in structured output. Overall, explanation ordering, model scale and pre-training knowledge, task-specific fine-tuning, and task structure jointly influence both prediction accuracy and reasoning faithfulness.

1 Introduction

Prior explanation-driven vision-language systems commonly generate answers before post-hoc rationales, which may be detached from the model’s reasoning. This work compares generation orders in joint answer–rationale generation and uses their effects on predictions and output consistency to assess explanation faithfulness.

1 Introduction

  • Prior explanation-supervised vision-language work uses natural-language explanations to jointly train answer prediction and rationale generation, but explanations can remain detached from reasoning.The conventional approach generates the answer first and then provides a post-hoc explanation, often grounded mainly in visual evidence with limited world knowledge.Source: §1 · p. 1, Figure 1 · p. 1
  • Chain-of-thought prompting can articulate intermediate visual-grounding and world-knowledge steps, but may add redundant or task-irrelevant information.The paper positions joint answer–rationale generation as a more task-focused alternative that avoids unnecessarily long or noisy chains.Source: §1 · p. 1
  • Joint answer–rationale generation is evaluated with rationale-first and answer-first ordering across tasks and models, in zero-shot and fine-tuning settings.Rationale-first generation makes the final answer explicitly conditioned on the intermediate rationale within one generation process.Source: §1 · p. 2
  • The generation-order comparison diagnoses faithfulness: stable predictions and output consistency suggest explanations reflect reasoning, while discrepancies indicate they are not causally grounded in decisions.The method evaluates jointly generated answers and explanations without requiring long intermediate reasoning chains.Source: §1 · pp. 1–2

2 Method

The method isolates generation order by changing whether the rationale or answer comes first while holding the task prompt and structured output constant. Fine-tuning uses the same answer-consistent rationale in both target orderings.

2.1 Generation Order Intervention

  • Generation-order intervention compares rationale-first prediction with answer-first post-hoc justification while keeping prompts and structured JSON outputs consistent.The shared instruction requests reasoning grounded in visual evidence and relevant knowledge, isolating ordering as the intervention.Source: §2.1 · p. 2

2.2 Fine-tuning with Explanation Supervision

  • Fine-tuning rationales are conditioned on the ground-truth answer so explanations remain consistent with the correct answer rather than potentially incorrect intermediate reasoning.For each sample, the training procedure constructs a rationale-first sequence and an answer-first sequence using the same rationale.Source: §2.2 · p. 2
  • Training applies autoregressive likelihood to both target orderings, teaching the model to produce consistent answer–rationale content in either sequence.The two likelihood terms correspond to rationale-then-answer and answer-then-rationale supervision formats.Source: §2.2 · p. 2

3 Experiments

Experiments use two 7B vision-language models across knowledge-intensive QA, visual entailment, and compositional image–caption matching. The evaluation combines task-specific performance metrics with agreement between predictions made under different generation orders.

3.1 Models.

  • The experiments compare Qwen-2.5-VL-7B-Instruct and LLaVA-1.5-7B, selected to balance model capacity and deployment efficiency.Qwen-VL was chosen to generate rationales for explanation-augmented fine-tuning based on prior work and preliminary zero-shot results.Source: §3.1 · p. 3

3.2 Datasets.

  • A-OKVQA tests knowledge-intensive visual question answering, where answers require external knowledge grounded in the scene.The experiments use its validation set and leverage human-written rationales for supervised explanation fine-tuning after both models achieved near-perfect accuracy conditioned on those rationales.Source: §3.2 · p. 3
  • For e-SNLI-VE, generated rationales conditioned on ground-truth labels provide controlled, answer-consistent supervision for fine-tuning.Its human-written explanations are not designed as fine-grained rationales tightly grounded in the decision process.Source: §3.2 · p. 3
  • Winoground evaluates compositional reasoning on two images and two contrasting captions, decomposing each sample into caption-to-image and image-to-caption subtasks.The 400 samples are split into 200 training and 200 evaluation samples; rationales are generated for training instances in each subtask.Source: §3.2 · p. 3, Table 1 · p. 3

3.3 Evaluation

  • Evaluation uses accuracy for A-OKVQA and Winoground, three-label F1 for e-SNLI-VE, and cross-order prediction agreement as a consistency score.The combined performance and consistency analysis assesses whether explanations are post-hoc or contribute to reasoning, and whether explanation supervision improves alignment.Source: §3.3 · p. 4

4 Results

Across the three benchmarks, explanation-order effects vary by model and task, while fine-tuning generally improves cross-order consistency or prediction robustness. Structured-output parsing is less reliable with rationale-first generation, and correct answers do not guarantee aligned explanations.

4.1 Performance

  • Qwen-VL outperforms LLaVA across A-OKVQA settings, while harder questions show model-dependent order effects and fine-tuning improves cross-order consistency.On the easier split, RF and AF differ only slightly; explanation supervision primarily improves alignment and robustness rather than task accuracy.Source: §4.1 · p. 4, Table 3 · p. 4
  • On e-SNLI-VE, Qwen-VL is stronger and more balanced; fine-tuning improves per-label F1 and raises contradiction consistency to 93.5% for Qwen-VL and 92.4% for LLaVA.Zero-shot order effects are larger and less predictable for LLaVA; Qwen-VL’s RF/AF gap is small for entailment and neutral but large for contradiction (79.2% vs 52.1%).Source: §4.1 · p. 4, Table 4 · p. 4

4.2 Vision Language Compositional Reasoning

  • On Winoground, Qwen-VL is less position-biased but more response-format-sensitive, while LLaVA shows first-position bias and fine-tuning only partly closes the gaps.Compositional challenges remain for both models; the authors suggest that 200 training samples (800 subtask instances) may be insufficient for robust explanation-driven learning.Source: §4.2 · pp. 4–5, Table 5 · p. 5

4.3 Error Analysis

  • Rationale-first outputs produce more parsing ambiguity than answer-first outputs, while fine-tuning largely removes parsing errors and improves structured-answer extraction.Answer-first can still yield inconsistent explanations, especially with LLaVA.Source: §4.3 · p. 5
  • For the orange-vehicles example, fine-tuned LLaVA gives the correct air-traffic answer in answer-first mode but describes passenger shuttles, showing that answer accuracy need not ensure aligned rationales.Qwen-VL’s prediction in the example is unchanged by response order or fine-tuning.Source: §4.3 · p. 5, Figure 3 · p. 5

5 Related Work

Prior explanation-driven VLR work incorporated explanations into prediction pipelines, but often treated them as auxiliary or post-hoc outputs. Reasoning-oriented prompting and knowledge-grounded datasets expanded the approach, while explanation-order effects remained underexamined.

Prior explanation-driven VLR

  • e-ViL [Kayser et al., 2021] emphasized human-readable explanations alongside answers, while NLX-GPT [Sammani et al., 2022] jointly generated predictions and explanations through explanation-supervision fine-tuning.These approaches could leave explanations loosely aligned with the answer-producing reasoning and post-hoc in nature.Source: §5 · p. 5

Reasoning-oriented methods

  • CoT and compositional reasoning methods add intermediate steps, but reported gains on compositional vision-language tasks remain limited [Mitra et al., 2024].The passage notes that adding intermediate reasoning does not necessarily improve fine-grained alignment or compositional understanding.Source: §5 · p. 5

Knowledge-grounded datasets

  • A-OKVQA [Schwenk et al., 2022] and [Lu et al., 2022] provide rationales grounded in commonsense and external knowledge, but prior work largely emphasized accuracy or explanation quality rather than how explanation order affects reasoning.Source: §5 · p. 6

6 Conclusion

Generation order can significantly alter model predictions, and explanation supervision improves consistency across orders without eliminating sensitivity on compositional tasks. The models also differ: Qwen-VL shows stronger reasoning capability, while LLaVA exhibits weaker reasoning and sometimes strong positional biases.

6 Conclusion

  • Generation order can significantly alter predictions, and explanation supervision improves consistency and robustness across orders, though residual sensitivity remains on compositional tasks.The findings suggest explanations interact directly with reasoning rather than serving only as auxiliary outputs.Source: §6 · p. 6
  • Qwen-VL exhibits stronger reasoning capability but remains order-sensitive, while LLaVA shows weaker reasoning and sometimes strong positional biases.Source: §6 · p. 6
  • Explanation supervision through fine-tuning substantially improves consistency and robustness across generation orders, but does not remove residual sensitivity on compositional tasks.Source: §6 · p. 6
  • The findings support considering generation structure when evaluating explanation faithfulness in vision-language models.Source: §6 · p. 6

Declaration on the Use of AI Assistants

The authors used GitHub Copilot to help develop code, reviewed and validated the generated code, and accept responsibility for the publication’s final content.

Declaration on the Use of AI Assistants

Loading 2609.29496v1…