PaperlayerView exact revision on arXiv

Source-linked AI summary

Where Do Long Captions Fail? Position-Aware Diagnosis and Reinforcement Learning for Detailed Image Captioning

Yuanze Hu, Zhichao Yang, Junwei Jing, Xin Yang, Wenxuan Liu, Tinghai Zhang, Jia Fu, Hongzhi Zhang, Xiangyu Wu, Ruiming Tang, Tingting Gao, Han Li

arXiv:2610.31111v1cs.CV

TL;DR

Global caption evaluations do not show where long captions fail, even though caption content is ordered and later generation may be less visually grounded. The paper diagnoses middle-caption unsupported specificity and tail generic escape, then trains PAGER with position-specific rewards; on the diagnostic subset, it reduces the middle hallucination effect while improving tail support and lowering tail hallucination.

  • Problem

    Global caption evaluations summarize coverage, factuality, or preference but do not show where unsupported or uninformative content appears in a long caption.

    Source: §1 · p. 1

  • Method

    A position-aware rubric verifies generated captions against atomic claims from dense references, and PAGER uses the resulting position-specific diagnosis to shape GRPO rewards.

    Source: §1 · p. 2

  • Results

    Across six captioners, unsupported specificity peaks in the middle and generic sentences increase at the tail; on the 198-image subset, PAGER reduces Dmid from +.096 to +.036, reaches SQ4 = .436, and has HQ4 = .235.

    Source: §1 · p. 2, §5.2 · p. 8

  • Takeaways & Limitations

    Long-caption failures differ by position, so the paper’s diagnosis and training objective target middle evidence and tail genericity with distinct signals.

    Source: §1 · p. 2

    Verification is reference-supported rather than directly image-grounded, so an incorrect or incomplete reference can create a false contradiction.

    Source: §2 · p. 3

Abstract

from arXiv · show

Detailed image captioning is usually framed as an amount problem: a model should mention more visual facts while avoiding hallucination. However, autoregressive VLMs generate captions as ordered sequences, and later generations may be produced under weaker or noisier visual grounding. We therefore argue that long-form captioning should be studied not only by what facts are mentioned, but also by where they appear. We introduce a position-aware rubric representation that decomposes dense human references into atomic visual claims, verifies generated captions against those claims, and induces a quartile-wise composition of supported, hallucinated, and generic content. This lens reveals two qualitatively different failures in long captions: models tend to produce unsupported specificity in the middle, while the final portion often escapes into vague, non-evidential language rather than simply accumulating more hallucination. A same-reference human control does not exhibit these patterns; additional human audits examine the visual validity of the induced labels. Motivated by this diagnosis, we propose PAGER (Position-Aware Grounded Evidence Reinforcement), a GRPO training objective with separate signals for supported fine-grained middle evidence, evidence-bearing tail continuation, claim coverage, and content-aware length control. The result is a closed loop: the same ordered claim representation identifies where long captions fail and supplies the position-specific reward needed to repair those failures. Experiments on detailed-caption benchmarks show that PAGER improves position-wise grounding and pairwise caption preference while preserving competitive global caption quality.

1 Introduction

Global caption scores capture how many facts are covered or factual, but collapse where a long caption fails. The paper diagnoses position-specific failures with ordered visual claims and uses those findings to motivate PAGER’s separate, position-aware rewards.

1 Introduction

  • Global scores collapse caption order, obscuring where unsupported or uninformative content appears [1] [2] [3] [4].The paper asks where such content occurs as a model writes a long caption.Source: §1 · p. 1
  • A position-aware rubric decomposes dense human references into atomic claims and classifies generated-caption content by position as supported, hallucinated, or generic.Evidence spans induce a quartile-wise composition of these categories.Source: §1 · p. 2
  • Across six captioners, unsupported specificity peaks in the middle, while the final quartile contains more generic sentences than the first.Dense human references do not show these patterns under a same-reference consistency check.Source: §1 · p. 2
  • Tail genericity is not simply increasing hallucination: in three of six models, tail hallucination is lower than head hallucination despite much higher tail genericity.Encouraging more late-caption content uniformly could reduce genericity by inducing fabricated detail.Source: §1 · p. 2
  • PAGER maps the diagnosis to separate GRPO signals for supported middle evidence, generic tail escape, claim coverage, and content-aware length control.Its rubric data are built first from curated dense-caption data, then from a larger PixMo-derived set filtered for image quality, caption length, and pre-inference judge quality.Source: §1 · p. 2

2 A position-aware lens for long captions

The section builds a position-aware representation that links reference claims to evidence spans in generated captions, enabling analysis of caption content by position. It also distinguishes this approach from global rubric scoring by inducing generic content as a residual category and targeting position-specific failures.

2 A position-aware lens for long captions

  • Reference rubrics are decomposed into visually checkable atomic claims, each assigned a dimension and weight.Dimensions are ENTITY, VISUAL_DETAIL, INTERACTION, and SCENE_CONTEXT; their default weights are 1.0, 2.0, 1.5, and 1.0, respectively.Source: §2 · p. 3
  • A verifier evaluates each rubric against the generated caption and returns an evidence span for non-missing judgments.Each rubric judgment inherits the quartile of the sentence containing its evidence span, producing a position-aware claim judgment.Source: §2 · p. 3
  • A sentence is GENERIC when no supported or contradicted reference rubric cites evidence from it.This makes GENERIC a residual category rather than the output of an auxiliary classifier.Source: §2 · p. 3
  • Within each quartile, supported, contradicted, and generic sentence weights are normalized by total sentence weight, while missing rubrics count only toward caption-level coverage.The three within-bin rates sum to one; the cross-model diagnosis uses rubric-dimension weights.Source: §2 · p. 3
  • Positive Dmid denotes elevated middle hallucination, positive Dtail elevated final-quartile generic content, and negative ΔS depleted tail support.Figure 1’s forest plots show positive Dmid and Dtail effects for every evaluated model, with confidence intervals excluding zero; human references do not show these effects.Source: §2 · p. 3, Figure 1 · p. 4
  • Unlike global rubric scoring, this representation ties claims to caption positions, exposes generic content, and targets distinct failures with position-specific rewards.The paper does not claim novelty for using rubrics per se.Source: §2 · p. 4

3 Diagnosing position-dependent failure

The diagnostic tests whether position-dependent failures in long captions are systematic, using six model captioners and a same-reference human control. It finds middle-position hallucination and tail-position genericity, a pattern not reducible to monotonic error growth.

Setup

  • The diagnostic evaluates six captioners on 198 images from a mixed IIW-400 and DOCCI set, with image-level paired-bootstrap confidence intervals.The same pipeline is applied to dense human references using rubrics derived from each reference.Source: §3 · p. 4

Human-reference control

  • Human references show weighted contradiction at or below 0.002 in every quartile, near-zero Dmid, and no tail generic escape in the same-reference control.This is a consistency check, not a test of cross-writer agreement or reference completeness.Source: Table 1 · p. 5, §3 · p. 5

Mid-caption unsupported specificity

  • Across six models, mid-caption hallucination rises in every model, with mean Dmid +0.101 and a 95% confidence interval excluding zero.The middle is where models add finer attributes, smaller objects, and relations, many unsupported by the reference.Source: §3 · p. 5

Tail generic escape

  • Mean Dtail is +0.178 across models, while supported content falls by 0.195 from Q1 to Q4.Generic content rises sharply in the final quartile for every model.Source: §3 · p. 5

Substitution rather than monotonic decay

  • Three of six models have lower tail than head hallucination even as tail genericity rises, indicating substitution rather than monotonic error accumulation.Hallucination-only metrics miss vague tail sentences that are neither supported nor contradicted by a visual claim.Source: §3 · p. 5

4 PAGER: Position-Aware GRPO

PAGER turns position-aware caption diagnostics into a GRPO reward with distinct signals for claim coverage, supported fine-grained middle evidence, generic tail content, and caption length. Its rubric data curriculum and group-relative optimization support training while the inference policy uses a decoding budget.

Reward design

  • PAGER combines claim coverage, a reward for supported fine-grained evidence in Q2/Q3, a penalty for generic final-quartile sentences, and a content-aware length budget.The separate middle and tail terms reflect distinct diagnosed failure modes; the tail term does not directly reward every non-generic sentence.Source: §4 · pp. 6–7

Claim coverage

  • The coverage reward supports verified rubric items and penalizes contradicted items, while missing rubrics are neutral.Neutral missing claims avoid forcing guesses about ungrounded details and keep reward magnitudes comparable across images with different rubric counts.Source: §4 · p. 6

Mid-caption specificity

  • The middle reward is restricted by both position and content type, rewarding reference-supported VISUAL_DETAIL and INTERACTION claims evidenced in Q2/Q3.Rubric metadata provides claim type, while evidence localization provides caption position.Source: §4 · pp. 6–7

Tail evidence density

  • The tail penalty uses a sentence-level genericity judge on final-quartile sentences, while contradiction penalties address fabricated detail.The training judge differs from diagnosis, where genericity is induced as the residual category of itemwise verification.Source: §4 · pp. 6–7

Content-aware length budget

  • The image-specific length budget scales with rubric count, and the implementation measures generated-caption length in characters.Rubric count is unavailable at inference, so reported evaluations use a decoding budget instead.Source: §4 · p. 7

Training data and optimization

  • The training curriculum builds and validates dense rubrics from curated DCI and DOCCI captions, then filters a larger PixMo-derived corpus before GRPO optimization.For each image, GRPO samples a caption group, normalizes rewards into relative advantages, and updates the policy under a KL penalty.Source: §4 · p. 7

5 Experiments

PAGER reduces the mid-caption hallucination effect and improves tail support without increasing tail hallucination, while global metrics show both gains and trade-offs.

5 Experiments

  • Experiments test whether PAGER reduces mid-caption hallucination, sustains evidence-bearing tail content, and avoids replacing generic endings with hallucinations.The setup compares Base, SFT, and CapRL and reports global metrics on IIW-400 and DOCCI alongside position-wise diagnostics.Source: §5 · p. 7, §5.1 · p. 8

5.2 Main results: global quality and positional repair

  • PAGER reduces Dmid from +.096 to +.036, reaches the highest tail supported rate (SQ4 = .436), and the lowest tail hallucination rate (HQ4 = .235).These results are reported on the same 198-image diagnostic subset.Source: §5.2 · p. 8, Table 3 · p. 8
  • PAGER scores .642 on IIW-400 and .610 on DOCCI for CAPTURE, while DCScore-F1 favors CapRL, making the global-quality results a trade-off.PAGER also wins 96.28% of tie-excluded four-model pairwise comparisons on IIW-400.Source: Table 4 · p. 8, §5.2 · p. 9

5.3 Ablations and robustness

  • Removing Rmid raises Dmid to +.103, while removing Rtail lowers SQ4 to .374 and raises GQ4 to .378; a monotone late-position weight instead raises HQ4 to .291.The holistic checklist reward is also weaker than itemwise evidence-span supervision.Source: §5.3 · p. 9, Table 5 · p. 9
  • Verifier judgments agree with human labels at κ = .89/.92, and three-seed 8B/32B transfer runs report Dmid = .041/.043 and HQ4 = .216/.214 at 100 steps.Supplementary image audits, human preferences, and a qualitative caption example are also reported.Source: §5.3 · p. 9

6 Related work

Prior captioning metrics and rubric-based methods assess factuality, coverage, preference, or checklist performance, largely through global scores. This work instead examines where supported, hallucinated, and generic content occurs within captions.

6 Related work

7 Discussion and limitations

The discussion frames long captions as ordered sequences and distinguishes unsupported specificity in the middle from generic escape at the tail. It also describes limitations of reference verification and the construction of atomic visual rubrics.

7 Discussion and limitations

  • PAGER assigns separate reward terms to the middle and tail because unsupported specificity and generic escape are distinct position-specific failures.The paper suggests the ordered-claim diagnosis and aligned training signals may extend beyond captioning to multimodal reasoning, text generation, and tool-use trajectories.Source: §7 · p. 10
  • Reference verification can miss correct details absent from the reference or create false contradictions; human studies used unpaid volunteers from the authors’ personal circle.The post-RL audit is descriptive, diversity comparisons control sentence count despite differing caption lengths, and transfer results cover tested short-budget settings with per-scale learning rates.Source: §7 · p. 10

A.1 Reference rubric construction

  • The reference rubric extractor turns dense captions into atomic claims with dimensions, weights, and keywords, resolving pronouns and preserving compound attributes.The rubric taxonomy includes Entity, Visual_Detail, Interaction, and Scene_Context dimensions with specified weights.Source: §A.1 · p. 14

A.2 Position-aware checklist verification

The appendix describes how itemwise verification spans are mapped to caption sentences and aggregated by quartile. It also specifies the induced labels, bootstrap intervals, and diagnostic subsets.

A.2 Position-aware checklist verification

  • The itemwise judge labels each rubric supported, contradicted, or missing and returns an evidence span, which is matched to sentence positions and quartiles.Exact normalized string matching is tried first; unresolved spans are excluded after fuzzy matching.Source: §A.2 · pp. 15–16
  • A sentence is GENERIC when no supported or contradicted rubric cites it; the separate tail judge flags vague, non-checkable sentences for the tail reward.The tail judge's binary flags are averaged over Q4 with a negative coefficient, while diagnostic G remains the residual verification category.Source: §A.2 · p. 16

A.3 Sentence-level induced labels and aggregation

  • Sentence labels and weights produce within-quartile supported, contradicted, and generic rates that sum to one; missing rubrics affect caption coverage but not quartile denominators.The aggregation defines quartile composition from sentences with evidence spans.Source: §A.3 · p. 17

A.4 Bootstrap confidence intervals

  • Image-level paired bootstrap uses 2,000 resamples, sampling images with replacement before pooling sentences and recalculating quartile counts.This accounts for correlation among sentences from the same image.Source: §A.4 · p. 17

B Additional diagnostic statistics

  • The diagnostic uses 198 images from mixed IIW-400 and DOCCI data; the appendix reports both weighted and unweighted position-wise statistics.The weighted table reports rubric-based rates, while the unweighted table checks that position effects are not artifacts of rubric-dimension weighting.Source: Table 6 · p. 17, §B.1 · p. 18, Table 7 · p. 18

C Verifier reliability

The verifier is evaluated against human-majority rubric judgments, with a human-reference control checking its behavior on captions that generated the rubrics. The judges approach human annotation agreement, although contradiction remains the hardest class.

C Verifier reliability

  • Human annotators agreed at Cohen’s κ = 0.94 across 347 held-out rubric–caption judgments.Human-majority labels served as the reliability target.Source: C Verifier reliability · p. 18
  • All LLM judges were within 0.05 of the human inter-annotator ceiling in Cohen’s κ, while the contradicted class remained hardest.Source: Table 8 · p. 18
  • Reference contradiction stayed at or below 0.002 in every quartile when dense human captions were checked against rubrics derived from them.Source: C Verifier reliability · p. 18

D Training details

The main GRPO policy starts from Qwen3-VL-4B-Instruct and uses filtered PixMo-derived rubric data. The reported setup includes KL regularization and a composite PAGER reward; inference uses a fixed generation budget and specified decoding settings.

D Training details

  • The main GRPO policy is initialized from Qwen3-VL-4B-Instruct and trained on filtered PixMo-derived rubric data.Validation uses the DCI densely-captioned-images test parquet; earlier rubric-construction experiments used DCI and DOCCI training captions.Source: D Training details · p. 18
  • The main run used 4 nodes with 8 A100 GPUs per node and took approximately two days.The run optimizes with GRPO and KL regularization; its reward combines coverage, middle-position, tail-position, and length terms.Source: D Training details · p. 19, Table 10 · p. 19, Table 9 · p. 19

D.1 Inference-time length handling

  • At inference, |R(x)| is unavailable; validation decoding uses temperature 0.6, top-p 0.95, and one sample per prompt.Source: §D.1 · p. 20

E Evaluation protocol details

The evaluation protocol specifies decoding, benchmark measures, and human-labeled verifier judgments. The paper also documents asset-use terms and presents a qualitative case study comparing local caption behavior against a dense human reference.

E Evaluation protocol details

  • Validation decoding uses temperature 0.6, top-p 0.95, one completion per prompt, and maximum prompt and response lengths of 6144 and 8192 tokens.Verifier and reward-judge calls use deterministic decoding at temperature 0.0.Source: E Evaluation protocol details · p. 20
  • CapArena4 compares PAGER directly with displayed baselines on IIW-400 and DOCCI, excluding ties from its reported denominator.CAPTURE, DCScore-F1, facts per 100 tokens, and average length are computed from saved benchmark predictions.Source: E Evaluation protocol details · p. 20
  • The verifier reliability study uses 347 human-labeled rubric judgments; separate appendices describe volunteer sentence audits and caption preference.The paper authors annotated the held-out rubric-level judgments.Source: E Evaluation protocol details · p. 20

F Existing assets, licenses, and terms of use

G Qualitative case study

  • Figure 3 illustrates local caption judgments with colors for supported, incorrect or unsupported specific, and generic or non-evidential spans.The case is illustrative; quantitative claims are supported by aggregate results in Tables 3–5.Source: G Qualitative case study · p. 22

G.1 Dense Human Reference

  • The dense human reference describes the Cruis’n World cabinet, its controls and display, a change machine, and the wooden-plank room.The case-study summary contrasts this reference with generated captions’ local behavior.Source: §G.1 · pp. 22–23, Table 13 · p. 23

G.2 Model Captions with Local Annotations

The local annotations contrast generic or erroneous caption content with descriptions tied to visible evidence. PAGER’s late caption remains more object- and layout-focused, though it still contains errors.

G.2 Model Captions with Local Annotations

  • The base model is mostly correct globally but ends with generic atmosphere, while its captions also include specific errors.Source: §G.2 · p. 26
  • CapRL misreads the title as “Crashin’ World” and speculates about Crash Bandicoot.The caption also describes the displayed game and arcade hardware in detail.Source: §G.2 · pp. 24, 26
  • More of PAGER’s late description refers to visible objects and layout, but its annotated errors remain.Source: §G.2 · p. 26

H Diagnostic limitations

The diagnostic is reference-relative rather than exhaustive image verification, and its supplemental audits and preference study provide descriptive checks of label validity and caption preference. These studies are bounded by volunteer sampling and the absence of real-world or high-stakes deployment evaluation.

H Diagnostic limitations

  • The diagnostic measures agreement with dense references, not exhaustive image entailment, so omitted correct details may remain GENERIC and imperfect references can yield false contradictions.The paper notes that reference omissions, ambiguity, and shared judge biases remain relevant limitations.Source: H Diagnostic limitations · p. 27

I Broader impacts

  • Detailed captioning may aid accessibility and image search, but could also increase over-trust, expose sensitive information, or reproduce verifier and dataset biases.The system was not deployed in real-world applications or evaluated for high-stakes decisionmaking.Source: I Broader impacts · p. 27

J Additional validation and robustness experiments

  • The supplemental studies examine visual validity, caption preference, post-training residual content, judge consistency, and transfer under tested settings.The sampled image recheck motivates confirming contradictions with images, but the main experiments retain the original rule.Source: J Additional validation and robustness… · p. 28, §J.2 · p. 28

J.1 Human evaluation protocol

  • Additional human evaluations used unpaid volunteers recruited from the authors’ personal circle, including friends, so the sample is not representative of users.Two annotators conducted image-based and residual-label audits; five volunteers ranked three anonymized captions per image across 100 images.Source: §J.1 · p. 28

J.2 Image-based audits

  • 142/150 PAGER supported claims (94.7%) were confirmed against images, while 139/150 diagnosed contradicted sentences (92.7%) conflicted with images.Relabeling three false contradictions changed mean Dmid from .101 to .099; the main experiments retain the original reference-supported rule.Source: §J.2 · p. 28

J.3 Blind three-way human preference

  • 389/500 first-place votes (77.8%) favored PAGER, and every rater’s overall ordering was PAGER > CapRL > SFT.Raters cited anchored transitions and fuller coverage for PAGER, while also noting occasional formulaic phrasing.Source: §J.3 · p. 29

J.4 Post-training residual content and lexical diversity

This section presents supplementary evaluations of position-wise gains under independent judges and short-budget transfer, alongside descriptive post-training changes and caveats about study scope, reproducibility, and deployment risks. It also documents asset licensing and human-subjects limitations and identifies the LLMs used in the paper’s core evaluation and training pipeline.

J.4 Post-training residual content and lexical diversity

  • Post-training residual-label differences were descriptive because all intervals included zero, while tail lexical-diversity changes suggested modest stylistic concentration.The diversity comparison used two Q4 sentences per caption from 198 images while retaining generated length; overall captions averaged 257 tokens for Base and 412 for PAGER, and human raters occasionally noted formulaic phrasing.Source: §J.4 · p. 29, Table 17 · p. 29

J.5 Independent-verifier comparison

  • PAGER ranks favorably on Dmid, SQ4, and HQ4 under both judges, with 90% diagnostic-output agreement and κ = .83, though reference-based judges may share blind spots.Residual genericity is slightly lower for CapRL under Qwen3-8B (.322 versus .329) but for PAGER under DeepSeek-V3.2 (.330 versus .336); image audits and human preferences address complementary risks.Source: §J.5 · p. 30, Table 18 · p. 30

J.6 Stronger foundation-model diagnosis

  • The same reference-relative diagnostic was applied without retraining to GPT-5.5 and Gemini-3.1-Pro on the 198-image set; reported original-range tail examples include InternVL3.5-8B at .090.Other examples are Qwen3-VL-8B at .122 [.098, .149] and LLaVA-v1.5-7B at .380.Source: §J.6 · p. 30

J.7 Short-budget transfer and three-seed replication

  • PAGER training lowered Dmid and HQ4 and raised SQ4 relative to pretrained rows, while three-seed means stayed close to single-run values under the tested 100-step settings.The SQ4 intervals overlapped, with greater variability for 8B; model size and learning rate varied together in the transfer comparisons.Source: §J.7 · p. 30, Table 20 · p. 31, Table 21 · p. 31

J.8 Construction details and measured cost

  • Filtering retained 297,839 of 500,000 candidates, and reversing the supported-precedence rule shifted Dmid and Dtail by less than .01.The screen first retained 494,596 candidates; the 187 mixed-evidence sentences were 1.6% of the 11,714-sentence sensitivity sample. The main 4B run used 32 A100 GPUs for approximately two days, while reported token counts cover extraction and itemwise verification rather than all online training costs.Source: §J.8 · p. 31

NeurIPS Paper Checklist

  • The checklist scopes claims to volunteer preferences and short-budget larger-backbone tests, notes limitations and judge dependence, and reports that code, checkpoints, and prediction files are not released.The paper describes implementation-level methods and settings for reproducibility, but compute disclosure covers the main run and component-level measurements rather than total project compute.Source: NeurIPS Paper Checklist · pp. 32–35

10. Broader impacts

  • The paper identifies potential benefits in accessibility and image search as well as risks including over-trust, privacy leakage, surveillance, and verifier bias, with safeguards discussed.Suggested safeguards include privacy filtering, uncertainty communication, human review, and license compliance; the manuscript does not describe releasing a high-risk dataset or model checkpoint.Source: §10 · p. 36
  • The paper answers No on asset-license reporting: Appendix F credits assets and lists identified terms, but does not enumerate all version-specific terms for hosted models.The paper does not release a new public dataset, code, or checkpoint.Source: §10 · p. 37
  • Additional volunteer evaluations lacked the original instruction sheets and interface screenshots; unpaid friends volunteered, and the original 347-item author study is distinguished from these studies.The paper states that participant risks were considered and disclosed and applicable ethics review requirements were followed, with required review or equivalent approval obtained.Source: §10 · pp. 37–38
  • LLMs are central to rubric extraction, itemwise verification, and tail-genericity judging, with Qwen3-8B used for training and main judging and DeepSeek-V3.2 for evaluation outside the GRPO loop.The paper reports the extractor and additional captioning models in Appendix J.Source: §10 · p. 38

Related papers

Loading 2610.31111v1…