Skip to content

ReflectCAP: Detailed Image Captioning with Reflective Memory

Conference: ECCV 2026
Paper: ECCV Proceedings
Area: Multimodal VLM
Keywords: detailed image captioning / reflective memory / multi-agent distillation / hallucination suppression / detail coverage

TL;DR

ReflectCAP uses an offline multi-agent pipeline to distill the recurring hallucinations and omissions of a target LVLM into two sets of "Structured Reflection Notes" (Avoid / Include), then injects them separately at inference time to produce a grounded base caption and a detail-focused caption that are conservatively merged with the image as an anchor—improving factuality and coverage simultaneously across 8 LVLMs with no training at all, at lower compute than model scaling or existing multi-agent pipelines.

Background & Motivation

Hyper-detailed captions must capture not only salient objects but also attributes, orientations, spatial relations, background context, and subtle visual states, and they have become a key ingredient for downstream multimodal systems—text-to-image and text-to-video generation rely on them as a faithful source of prompts, while compositional and grounded reasoning tasks use them as supporting evidence. LVLMs have long been able to write such long descriptions fluently, but they hallucinate frequently: once generation length grows, language priors progressively dominate over visual evidence, and the model begins to describe what is statistically probable rather than what is actually depicted. Hyper-detailed captioning inherently demands long generation, so the tension between detail and faithfulness has remained unresolved in real deployments.

The most intuitive remedy is supervised fine-tuning on human-authored detailed captions. Section 5.1 of this paper reports a counter-intuitive measurement instead: after LoRA fine-tuning InternVL3.5-4B and Qwen2.5-VL-7B on 9,647 human-authored DOCCI captions, factuality dropped from 64.1 to 58.2 and from 68.3 to 57.1 respectively—human captions are denser than the model's perceptual capacity, so the model never learns to see better, only to say more, and hallucinations are amplified. That route also demands expensive human annotation plus additional training. The alternative is inference-time self-correction: no training, just letting the model iteratively revise its own output. This works for text-only LLMs, but recent work shows that LVLMs without external feedback tend to confirm rather than rectify their own errors, and iterative revision lengthens the context, letting language priors override visual evidence even further. Both routes point to the same conclusion: a single LVLM cannot resolve the detail–faithfulness tension on its own, and external guidance is required.

This paper's angle is to move reflection from online to offline. The authors point out that the long-horizon reflective memory paradigm that succeeded for LLM agents does not transfer to LVLMs: LVLMs cannot reliably extract their own error patterns over long reasoning chains, and as inference steps accumulate the visual evidence is diluted. The fix is an offline, bottom-up distillation—let the model make its mistakes on a small exemplar set first, diagnose those mistakes against the strong signal of "image + human reference caption," and consolidate the image-specific diagnoses into model-level reusable directives that are simply injected into the prompt at inference time, bypassing costly step-by-step refinement. Core idea: a three-agent offline pipeline distills a target LVLM's recurring hallucinations and omissions into two sets of Structured Reflection Notes (Avoid Notes and Include Notes), which are injected separately at inference time to generate a grounded base caption and a detail-focused caption that are then conservatively merged—suppressing hallucinations and recovering missing details at zero training cost.

Method

Overall Architecture

ReflectCAP (Reflective Note-Guided Captioning) is a gradient-free framework: offline it compresses the target LVLM's systematic failure modes into a compact, reusable set of natural-language directives, and online it injects those directives into the generation context as a form of memory. Its input is a new image, its output is a detailed caption that is simultaneously faithful and fine-grained, and the only addition in between is a set of offline-constructed notes.

The pipeline has two phases. In the offline phase, given a small exemplar set \(D_{train}=\{(x_i, y^*_i)\}_{i=1}^{M}\) of images \(x_i\) paired with human-written reference captions \(y^*_i\) (the paper needs only \(M=30\)), three agents hand off in sequence: a Captioning Agent has the target LVLM caption each image zero-shot, a Feedback Agent critiques that caption against both the image itself and the human reference, splitting problems into hallucinations and omissions, and a Note Organizer incrementally consolidates the cross-image critiques, merging and abstracting them into a note set \(N=(N_{avoid}, N_{include})\) capped at \(K=5\) items per category. In the online phase the notes entirely replace the multi-agent pipeline: for a new image \(x\), \(N_{avoid}\) is injected to produce a grounded base caption \(c_{base}\), \(N_{include}\) is injected to produce a detail-focused caption \(c_{detail}\), and finally both captions plus the image are handed to the model to merge into \(c_{final}\)—at most 3 LVLM calls.

The key property is that the notes are model-level rather than image-level: they record what this model habitually gets wrong and habitually overlooks, not what went wrong on one particular image, which is what lets them be reused across images and even across benchmarks (verified in §5.2) while adding almost no token cost at inference time.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["exemplar images + human references"] --> B["offline reflection-note distillation<br/>per-image critique → cross-image abstraction"]
    B --> C["Structured Reflection Notes<br/>Avoid / Include Notes"]
    C -->|incremental batches keep consolidating| B
    C --> D["separate injection<br/>avoid and include run apart"]
    D -->|avoid notes → grounded base caption| E["conservative merge<br/>image-anchored, base caption wins"]
    D -->|include notes → detail-focused caption| E
    E --> F["final detailed caption"]

Key Designs

1. Offline reflection-note distillation: turning "this image was described wrongly" into "this model always errs this way"

Per-image diagnoses have no reuse value, and dumping a pile of image-level feedback into the context would inevitably overflow the window, so the three agents divide the labor. The Captioning Agent deliberately gives no guidance and has the target LVLM caption \(x_i\) zero-shot to obtain \(\hat{y}_i\)—this is intentional: the resulting caption faithfully reflects the model's default behavior, including its characteristic hallucinations and omissions, and thus forms an unbiased basis for the diagnosis that follows. The Feedback Agent then cross-references \(\hat{y}_i\) against the image \(x_i\) and the human reference \(y^*_i\) to produce a structured issue report \(I_i\) with two explicit categories: Hallucinations (details in \(\hat{y}_i\) that are factually wrong or simply not visible in the image, e.g. "the caption states two people are sitting, but the image shows three") and Missing Details (important details present in the human reference but absent from \(\hat{y}_i\), e.g. "the caption does not mention the wooden railing visible in the foreground"). The Feedback Agent sees the image while critiquing, so its judgments are visually grounded rather than text-only speculation.

Generalization is the Note Organizer's job. Critiques collected across images easily exceed the LVLM's context window, so it processes them incrementally: each step consumes one batch of reports, merges semantically similar items, abstracts specific descriptions into broadly applicable rules, and updates a running note set. A hard cap of \(K\) items per category then pushes out one-off mistakes and overly narrow entries, keeping only hallucination patterns that recur frequently and detail types that are commonly omitted. The result is two complementary note sets: Avoid Notes that suppress recurrent hallucination patterns (e.g. "Do not infer object colors when they are ambiguous") and Include Notes that enforce frequently omitted details (e.g. "Describe visible architectural details such as structural supports and railings"). This progression from instance-level diagnosis to cross-instance generalization is precisely what lets the notes capture systematic tendencies rather than one-off errors, and the ablation shows \(M=30\), \(K=5\) already suffices. Notably, a larger exemplar set is not better: with \(N=30\) the notes for GPT-4.1-mini are targeted rules ("Do not add unsupported details to signs, logos, or symbols"), whereas with \(N=100\) they degrade into generic directives ("Avoid subjective or interpretive descriptions not clearly supported by the image or reference"), diluting the corrective signal and slightly hurting performance.

2. Separate injection: running the avoid notes and the include notes in different passes instead of one prompt

Once the notes exist, the online phase must answer whether hallucination suppression and detail recovery can be achieved in a single generation. ReflectCAP's answer is no, so the two note sets drive two independent generations. Step 1 injects \(N_{avoid}\) into the prompt so the model actively suppresses its known hallucination patterns, yielding a grounded base caption \(c_{base}\) that is more reliable than the zero-shot caption while preserving the model's natural descriptive ability; because it needs exactly one forward pass, its cost is nearly identical to zero-shot inference, so it can be used standalone in pipelines where factuality is the primary concern (the paper calls this ReflectCAP-Base). Step 2 switches to \(N_{include}\), directing the model's attention to the content it systematically overlooks—material textures, background elements, spatial arrangements—yielding a detail-focused caption \(c_{detail}\) that supplies exactly the information \(c_{base}\) traded away for factual grounding.

Keeping them separate is supported by measurement (Table 5, 100 images sampled from IIW-400): injecting both note sets into a single prompt is worse for all four tested models, most severely for InternVL3.5-4B, whose F1 drops from 64.4 to 57.3; for GPT-4.1-mini it is 78.3 → 77.1. Instruction overload overwhelms models with weaker instruction-following ability, and hallucination suppression and detail recall are goals pulling in opposite directions—when each is controlled by its own dedicated guidance, the model never has to arbitrate between "say less, be safer" and "say more" within a single generation.

3. Conservative merge: the image is the anchor and the base caption wins

When the two captions are combined, the main risk is that "merging" degenerates into an unconstrained rewrite: the details \(c_{detail}\) surfaced in order to boost recall—details the model may not actually have seen correctly—could be accepted as fact in the third pass and written into the final output. ReflectCAP therefore uses the image as a grounding reference for the merge and imposes an explicitly asymmetric trust relation: \(c_{base}\) was generated under hallucination-suppressing guidance, so it is treated as the primary source of truth, while \(c_{detail}\) is supplementary, and conflicts are resolved in favor of \(c_{base}\). This rule confines "completion" to being purely additive to the existing description rather than allowing it to overwrite existing judgments. The price is that details in \(c_{detail}\) that are real but unmentioned in \(c_{base}\) may be suppressed, and the paper does not quantify that loss separately.

A Worked Example

Take the case study in §5.4 with GPT-4.1-mini: for a photograph of a building, the zero-shot caption is wrong in several fine-grained respects—roof shape, window count and arrangement, and the appearance of signage are all fabricated, with the window count stated as six. With the notes attached, the ReflectCAP-Base stage uses Avoid Notes to correct the window count to five and drops the unverifiable arrangement descriptions, producing a visibly more factual caption. The ReflectCAP-Full stage then uses Include Notes to recover elements the model systematically omitted, including the lion's-head fountain spout and cast shadows, and restores details such as embossed text in the final caption. For the same image, a single zero-shot call yields only a "plausible-looking" version, whereas the three guided calls perform error correction, detail recovery, and merging in turn.

Key Experimental Results

Evaluation runs along three axes: the fine-grained factuality/coverage trade-off (IIW-400), holistic caption quality (pairwise win rate on CapArena-Auto), and inference cost efficiency. The 8 LVLMs span both closed-source and open-source families: GPT-4.1-mini / GPT-4.1-nano, plus InternVL3.5-4B, Qwen2.5-VL-7B, and Qwen3-VL-8B (smaller) and InternVL3.5-38B, Qwen2.5-VL-32B, and Qwen3-VL-32B (larger). Baselines are Zero-shot, Few-shot (3 human exemplars), Self-Correction, and CapMAS (an inference-time multi-agent baseline that decomposes a caption into atomic propositions, verifies each, and rewrites). For fairness, both the offline and online phases of ReflectCAP use only the target model itself, never an external model; the offline exemplar images and references come from the portion of IIW-Eval excluded from IIW-400.

Main Results

Table 1 reports fine-grained results on IIW-400. Factuality (Precision) decomposes a caption into atomic propositions and verifies each against the image and ground truth; Coverage (Recall) answers each image's paired VQA items using only the generated caption; F1 is their harmonic mean \(F1=2PR/(P+R)\).

Model Method P R F1 ΔF1
GPT-4.1-mini Zero-shot 83.6 68.1 75.1 —
GPT-4.1-mini Few-shot 81.9 71.5 76.3 +1.2
GPT-4.1-mini Self-Corr. 82.6 69.3 75.4 +0.3
GPT-4.1-mini CapMAS 84.2 67.7 75.1 0.0
GPT-4.1-mini ReflectCAP 83.8 72.0 77.5 +2.4
GPT-4.1-nano Zero-shot 78.9 62.6 69.8 —
GPT-4.1-nano Few-shot 77.1 67.1 71.8 +2.0
GPT-4.1-nano Self-Corr. 79.2 62.6 70.0 +0.2
GPT-4.1-nano CapMAS 83.1 57.2 67.8 −2.0
GPT-4.1-nano ReflectCAP 78.0 67.2 72.2 +2.4
InternVL3.5-4B Zero-shot 64.1 54.9 59.1 —
InternVL3.5-4B Few-shot 49.4 52.8 51.1 −8.0
InternVL3.5-4B Self-Corr. 62.1 54.1 57.8 −1.3
InternVL3.5-4B CapMAS 72.2 53.3 61.3 +2.2
InternVL3.5-4B ReflectCAP 64.8 61.3 63.0 +3.9
InternVL3.5-38B Zero-shot 72.5 57.8 64.3 —
InternVL3.5-38B Few-shot 63.2 62.1 62.6 −1.7
InternVL3.5-38B Self-Corr. 73.7 58.2 65.0 +0.7
InternVL3.5-38B CapMAS 78.6 57.5 66.4 +2.1
InternVL3.5-38B ReflectCAP 73.7 64.2 68.6 +4.3
Qwen2.5-VL-7B Zero-shot 68.3 57.9 62.7 —
Qwen2.5-VL-7B ReflectCAP 68.8 62.3 65.4 +2.7
Qwen2.5-VL-32B Zero-shot 71.9 64.0 67.7 —
Qwen2.5-VL-32B ReflectCAP 69.9 66.4 68.1 +0.4
Qwen3-VL-8B Zero-shot 76.3 69.2 72.6 —
Qwen3-VL-8B ReflectCAP 77.0 71.2 74.0 +1.4
Qwen3-VL-32B Zero-shot 79.2 72.5 75.7 —
Qwen3-VL-32B ReflectCAP 79.2 73.8 76.4 +0.7

The original table is typeset in two columns; the per-row model assignment here was reconstructed from the cross-references in §4.4 and Table 3 (§4.4 explicitly states the InternVL3.5-38B zero-shot F1 is 64.3 and that Qwen3-VL-8B with ReflectCAP maintains a gap comparable to Qwen3-VL-32B zero-shot), ⚠️ some rows should still be checked against the original. The Few-shot / Self-Corr. / CapMAS columns are omitted for the four Qwen rows for readability; their F1 values are 53.1 / 60.4 / 63.9 (Qwen2.5-VL-7B), 64.7 / 68.1 / 67.9 (Qwen2.5-VL-32B), 71.8 / 73.0 / 73.7 (Qwen3-VL-8B), and 73.9 / 76.0 / 76.2 (Qwen3-VL-32B).

Holistic quality is verified with CapArena-Auto: each caption is compared head-to-head against three fixed reference models (GPT-4o-0806, CogVLM2-llama3-chat-19B, and MiniCPM-V2.6-8B) as judged by GPT-4.1-mini, and the metric is the average win-rate margin over \([-100,100]\).

Model Zero-shot CapMAS ReflectCAP (Ours)
Leaderboard anchors (zero-shot only)
GPT-5.2 70.0 — —
Gemini-1.5-Pro† 62.3 — —
GPT-4o-0806† 44.3 — —
Qwen2.5VL-72B† 39.7 — —
Claude-3.5-Sonnet† 29.7 — —
Evaluated models
GPT-4.1-mini 57.7 53.7 (−4.0) 90.0 (+32.3)
GPT-4.1-nano 21.2 −14.7 (−35.9) 51.3 (+30.1)
InternVL3.5-4B −54.3 −46.7 (+7.6) −18.7 (+35.6)
InternVL3.5-38B −24.0 −15.0 (+9.0) 12.0 (+36.0)
Qwen2.5-VL-7B −33.7 −28.0 (+5.7) −7.3 (+26.4)
Qwen2.5-VL-32B 5.7 8.0 (+2.3) 25.3 (+19.6)
Qwen3-VL-8B 76.0 74.0 (−2.0) 85.3 (+9.3)
Qwen3-VL-32B 87.3 87.0 (−0.3) 91.7 (+4.4)

Rows marked † take their zero-shot reference values from the provided CapArena caption. The abstract reports average gains of +32.2 for the GPT family and +21.9 for open-source models, while §4.3 writes +33.2 for the GPT family (the open-source figure matches the +21.9 average of the six open models in Table 2; the two GPT per-model gains +32.3 / +30.1 average to +31.2), ⚠️ the three figures disagree—refer to the original paper.

Ablation Study

Config Key metric Note
Separate-Merge GPT-4.1-mini 78.3 / nano 73.1 / Qwen3-VL-8B 74.9 / InternVL3.5-4B 64.4 (F1) Both note sets run in separate passes and are merged; this is full ReflectCAP
Combined 77.1 / 72.2 / 72.7 / 57.3 (F1) Both note sets crammed into one prompt; worse for all four models, −7.1 on InternVL3.5-4B
Zero-shot (InternVL3.5-4B) 59.1 F1 (64.1 / 54.9) Baseline
SFT w/ human captions (InternVL3.5-4B / Qwen2.5-VL-7B) 57.4 / 57.6 F1 LoRA on 9,647 DOCCI human captions; factuality falls to 58.2 / 57.1, below zero-shot
SFT w/ ReflectCAP-generated captions (same) 63.9 / 66.5 F1 Same 9,647 DOCCI images with labels replaced by ReflectCAP outputs; both factuality and coverage beat zero-shot
Frozen notes across benchmarks (GPT-4.1-mini / Qwen3-VL-8B) CaptionQA 74.5→81.8 / 74.9→78.0; CAPability Avg 68.6→71.0 / 69.5→71.9 Notes are not re-mined; transferred directly to CaptionQA (50 images per domain, 4 domains) and CAPability (100 images per dimension, 9 dimensions)

Key Findings

  • No baseline improves both axes at once, and ReflectCAP attains the highest F1 on all 8 models. Few-shot raises coverage by imitating human demonstrations but pushes the model beyond its perceptual boundary, so factuality drops (on InternVL3.5-4B, F1 falls from 59.1 to 51.1). Self-Correction yields only marginal gains of +0.2 to +0.7 regardless of scale, showing that models can neither find nor fix their own errors through revision. CapMAS lifts factuality by deleting unverifiable content (GPT-4.1-nano's P reaches 83.1) but necessarily sacrifices coverage (R only 57.2), ending below zero-shot on F1.
  • Coverage guidance inherently risks degrading factuality, and factuality guidance separately contains that degradation. This is exactly why the two note sets are injected separately: each controls one objective, so the model never has to satisfy two opposing requirements in a single generation. The ablation in Table 5 confirms this directly—combined injection is consistently worse across all four models.
  • The notes can close model-scale gaps. On CapArena-Auto, GPT-4.1-mini with ReflectCAP reaches 90.0, surpassing zero-shot GPT-5.2 (70.0); meanwhile CapMAS degrades models that are already strong zero-shot (Qwen3-VL-32B 87.3→87.0, GPT-4.1-mini 57.7→53.7), whereas ReflectCAP gains on all 8.
  • On cost efficiency it beats both model scaling and inference-time multi-agent pipelines. The paper approximates inference cost as \(C \approx 2NT\) (\(N\) = non-embedding parameters, \(T\) = total tokens; with multiple calls per image, image tokens are counted only once via KV caching). InternVL3.5-4B with ReflectCAP reaches an F1 of 63.0, approaching InternVL3.5-38B zero-shot at 64.3, at 27.5 vs 213.7 TFLOPs (about 7.8× fewer; the abstract rounds this to roughly 8×). Qwen3-VL-8B with ReflectCAP maintains a comparable gap to Qwen3-VL-32B zero-shot at roughly 3.3× lower compute. Against CapMAS, ReflectCAP uses 21%–36% fewer TFLOPs on InternVL3.5-4B, Qwen2.5-VL-7B, and Qwen3-VL-8B while scoring higher F1—because CapMAS runs its whole multi-agent pipeline at inference time, whereas ReflectCAP runs it once offline.
  • The gains from the Grounded Base Caption scale with instruction-following ability. Injecting Avoid Notes alone improves factuality on nearly all models, but the magnitude varies sharply: strong instruction followers such as GPT-4.1-mini / nano suppress hallucination patterns cleanly, while the Qwen2.5-VL family improves only marginally or even degrades. The authors conclude that as LVLM instruction following keeps improving, the framework will gain more without any modification.
  • Both note-construction hyper-parameters are "small but sufficient." For the number of exemplar images \(N\), all models gain substantially going from zero-shot to 10, largely plateau around 30, and slightly decline at 100 (notes become generic). For the per-category cap \(K\), even \(K=1\) already gives a substantial gain (the model aggregates several corrective signals into one composite instruction); \(K=5\) is better, and while some models keep improving up to \(K=10\), others peak around 5 and decline, so the optimum is model-dependent.
  • Frozen notes transfer across benchmarks (Table 4): without any re-mining, notes learned on IIW consistently beat zero-shot and CapMAS on CaptionQA and CAPability, indicating that they capture model-level error tendencies rather than benchmark-specific artifacts.

Highlights & Insights

  • Moving reflection from online trajectories to offline distillation is the paper's key trade-off. Reflective memory for LLM agents relies on long-horizon trajectory tracking, but for LVLMs the visual evidence is diluted as inference steps accumulate and their own error patterns become harder to extract; switching to an offline setting where the model first makes its mistakes and is then diagnosed against the strong signal of "image + human reference" sidesteps that failure mode, and once the notes are learned they replace the entire multi-agent pipeline at almost no inference overhead.
  • What the memory stores is neither visual features nor partial generated captions, but natural-language directives about the model's own failure tendencies. That choice determines the granularity: the memory is inherently model-level and reusable across images and even benchmarks (verified by the frozen-note transfer in §5.2), so it is far lighter than per-image reflection and needs no retrieval mechanism—the whole set is simply prepended to the prompt.
  • Generating for the two objectives in separate passes beats cramming both constraint sets into one prompt. The combined-injection ablation provides an unusually clean comparison: same directives, same model, only the organization differs, and the gap reaches 7.1 F1 points. This transfers directly to any multi-constraint prompting scenario—more constraints are not better, and instruction overload is itself a performance loss.
  • Using the method's own outputs as SFT labels is safer than using human labels. Human detailed captions exceed the model's perceptual boundary and amplify hallucinations, whereas ReflectCAP-generated captions stay within the model's capability, so training on them improves both factuality and coverage—a scalable path to auto-labeled detailed-caption data.
  • The cost accounting is reusable. The paper compares inference cost through \(C\approx 2NT\) and notes that with multiple calls per image the image tokens are counted only once via KV caching—an easily overlooked detail when comparing multi-turn or multi-agent methods, and one that determines that "3 calls" is not 3× the cost under this accounting.

Limitations & Future Work

  • The offline phase still depends on a small human-annotated exemplar set. \(M=30\) images with human references is cheap but not free, and while the paper shows notes transfer frozen to new benchmarks, it does not show they transfer across models—as constructed, the notes are model-specific, so a new target model requires re-running the whole offline pipeline, and the optimal \(K\) must be re-tuned per model.
  • The online phase takes up to 3 LVLM calls, still a 2–3× overhead for latency-sensitive settings. Using only the Grounded Base Caption costs nearly the same as zero-shot but forfeits the coverage gain, and the paper offers no middle ground such as "one call with both note sets."
  • The conservative merge is a lossy rule. Always resolving conflicts in favor of \(c_{base}\) prevents \(c_{detail}\)'s spurious details from contaminating the output, but it may also suppress details that are real in \(c_{detail}\) yet unmentioned in \(c_{base}\); the paper neither quantifies this loss nor experiments with a learned merge.
  • Evaluation scale is limited. Fine-grained evaluation is mainly on IIW-400 (400 images), CapArena-Auto uses 600 images, and part of the ablation evidence (Table 5, the \(N\)/\(K\) curves) comes from 100-image samples. Moreover, "hallucination" is measured as proposition-level factuality (Precision) rather than by a dedicated hallucination detector, and the two are not equivalent.
  • Self-identified issues: the two-column typesetting of Table 1 means the per-row model assignment has to be reconstructed from the prose cross-references (this note checked it against §4.4 and Table 3), and the CapArena average gain disagrees across the abstract (+32.2), the body (+33.2), and the per-model values in Table 2 (+32.3 / +30.1, averaging +31.2). Both should be resolved against the original paper.
  • Improvement directions: (1) make reflection notes shareable at the model-family level so smaller models inherit notes mined by larger siblings, removing one offline mining run per model; (2) automate the choice of \(K\) (early stopping on a small validation set) instead of tuning by hand; (3) introduce a light learned or verified decision into the merge stage—for example, verifying each increment from \(c_{detail}\) against the image before accepting it—so its recall can be fully exploited without giving up factuality.
  • vs CapMAS: Both target factuality in detailed captioning and both use multi-agent machinery, but in completely different places. CapMAS runs at inference time, decomposing a caption into atomic propositions, verifying each against the image, and rewriting to drop unverifiable content; it therefore only suppresses hallucinations, necessarily sacrificing coverage, and it is more expensive (ReflectCAP uses 21%–36% fewer TFLOPs). ReflectCAP moves the multi-agent work offline, leaving only prompt-level steering online, and handles the two error types with two dedicated note sets.
  • vs Self-Correction / Self-Refine: These methods have the model re-examine the image and revise its own output at inference time. This paper's measurements show they are almost ineffective here (+0.2 to +0.7 F1 across models), because an LVLM without external feedback confirms rather than rectifies its own errors, and iteration lengthens the context, amplifying language priors. ReflectCAP's critique signal comes from external supervision—the image plus a human reference—and is used only offline.
  • vs SFT on human detailed captions (DOCCI / IIW-style data): Fine-tuning directly on human captions amplifies hallucinations (InternVL3.5-4B factuality 64.1→58.2, Qwen2.5-VL-7B 68.3→57.1) because annotation density exceeds perceptual capability; ReflectCAP is a training-free route, and the captions it produces are in turn better training data (63.9 / 66.5 F1).
  • vs reflective memory for LLM agents (Reflexion / ReasoningBank and similar): Those works perform online, trajectory-level reflection and experience reuse on long-horizon text tasks. This paper explicitly argues that the paradigm does not transfer to LVLMs—visual evidence is diluted over reasoning steps and error patterns cannot be reliably extracted—and instead performs offline, bottom-up cross-sample distillation, reducing the "memory" from trajectory-level to a static set of model-level directives.

Rating

  • Novelty: ⭐⭐⭐⭐ Reflective memory and multi-agent pipelines are neither new on their own, but the argument that LVLMs are unsuited to online long-horizon reflection plus the offline distillation of model-level error directives with dual-path injection and conservative merging is new; methodologically it leans on prompt engineering rather than a mechanical innovation.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ 8 LVLMs × 4 baselines across fine-grained, holistic, and cost evaluations, plus an SFT comparison, frozen-note cross-benchmark transfer, and ablations on \(N\) and \(K\)—broad coverage for this class of work.
  • Writing Quality: ⭐⭐⭐⭐ The motivation chain is clear and Figures 3–5 tie closely to the claims, but Table 1's two-column layout is hard to read and the CapArena average gain disagrees between abstract and body, forcing readers to verify it themselves.
  • Value: ⭐⭐⭐⭐⭐ Training-free, near-zero-shot inference overhead, and cheaper than scaling model size; it applies directly to existing LVLMs and doubles as a low-cost pipeline for producing detailed-caption training data, so its practical value is clear.