AesCoT: Aesthetic Chain-of-Thought

• AesCoT is an aesthetic chain-of-thought pipeline that curates structured, multi-dimensional critiques from image and mean opinion score pairs.
• The pipeline employs a five-stage process—from score masking to multi-dimensional critique generation and rigorous filtering for leakage, misalignment, and factual errors.
• Empirical results show that integrating AesCoT with reinforcement learning significantly improves score calibration and produces expert-level aesthetic analyses.

Searching arXiv for the cited papers to ground the article. arxiv_search(query="(Liu et al., 26 Sep 2025) Aes-R1 AesCoT", max_results=5) arxiv_search(query="(Zeng et al., 9 Feb 2026) Token-Level Adaptive Latent Chain-of-Thought", max_results=5) arxiv_search(query="(Zhang et al., 7 Aug 2025) ASCoT Late-Stage Fragility", max_results=5) AesCoT most specifically denotes Aesthetic Chain-of-Thought, introduced within the Aes-R1 framework as “an aesthetic reasoning data construction pipeline that automatically curates diverse, reliable annotations with multi-dimensional aesthetic rationales” (Liu et al., 26 Sep 2025). In that framework, given an image and its human mean opinion score (MOS) $s_i$, AesCoT produces a structured aesthetic critique $c_i$ together with the ground-truth score $s_i$, and then filters these pairs into a high-quality dataset for cold-start supervision. Its operational role is data-centric rather than algorithmic: it teaches a multimodal LLM (MLLM) to generate structured explanations before scoring, so that subsequent reinforcement learning can optimize aesthetic judgment without collapsing into shallow or reward-hacking explanations (Liu et al., 26 Sep 2025).

1. Definition and formal role

AesCoT stands for Aesthetic Chain-of-Thought. The pipeline begins from an image dataset

$\mathcal{D}=\{(\mathcal{I}_i,s_i), i\in[1,M]\},$

where $\mathcal{I}_i$ is an image and $s_i\in[0,1]$ is its normalized MOS. It constructs a raw chain-of-thought dataset

$$\mathcal{D}_{\text{CoT}}=\{(\mathcal{P},\mathcal{I}_i,c_i,s_i), i\in[1,M]\},$$

then applies a filtering process $\mathcal{F}$ to obtain

$$\mathcal{D}_{\text{AesCoT}}=\mathcal{F}(\mathcal{D}_{\text{CoT}})=\left\{(\mathcal{P},\mathcal{I}_i,c_i,s_i)\;\middle|\; \lVert \mathcal{E}_i\rVert=\mathbf{0}\right\},$$

with

$$\mathcal{E}_i=(e_i^{\text{leak}},e_i^{\text{align}},e_i^{\text{fact}}).$$

Here $c_i$0, $c_i$1, and $c_i$2 encode score leakage, reasoning–score misalignment, and factual errors, respectively (Liu et al., 26 Sep 2025).

Within Aes-R1, AesCoT is explicitly positioned as the pipeline “to construct and filter high-quality aesthetic reasoning data used for cold-start.” Its immediate goals are fourfold: cold-start aesthetic reasoning, high-quality CoT data at low cost through distillation from GPT-4.1, multi-dimensional interpretability through five aesthetic dimensions, and provision of a good RL starting point. The paper contrasts this role with an RL-only initialization, denoted “Aes-R1-Zero,” which can optimize scores but produce shallow explanations (Liu et al., 26 Sep 2025).

A plausible implication is that AesCoT should be understood as the reasoning substrate of Aes-R1 rather than as its reward function. The paper’s own division of labor supports this reading: AesCoT supplies structured supervision, whereas RAPO performs preference optimization over scores.

2. Construction pipeline and representational structure

The AesCoT pipeline proceeds through five stated stages: score masking and expert prompting, CoT generation over five aesthetic dimensions, concatenation with the score, automated and human filtering, and final dataset release as AesCoT-3K and AesCoT-10K (Liu et al., 26 Sep 2025).

For each image–score pair $c_i$3, GPT-4.1 is prompted to write a critique while treating the score as hidden. The construction prompt states in part: “Treat rating score (range 0-1, two-decimal precision) as strictly private information,” and requires a critique that “ends with a judgment of the image that matches the hidden score.” The paper further states that it “first intentionally conceal[s] the continuous score from outputs” and asks the closed-source MLLM to “give aesthetic analysis over five aesthetic dimensions … match the aesthetic levels, which are divided into bad (0–0.4), fair (0.4–0.7) and good (0.7–1.0) to emphasize aesthetic differences” (Liu et al., 26 Sep 2025).

The five aesthetic dimensions are Light and Shadow, Mood and Narrative, Composition, Color, and Exposure. The appendix-level descriptions are detailed and domain-specific. Light and Shadow addresses volume, spatial depth, visual focus, directionality, contrast, atmosphere, and highlight/shadow detail. Mood and Narrative addresses affective tenor, implied story, posture, environmental cues, symbol, metaphor, and gesture. Composition addresses balance, rhythm, hierarchy, eye paths, and negative space. Color addresses hue, value, chroma, harmonies, warm–cool contrasts, and palette. Exposure addresses tonal distribution, high key and low key decisions, and information loss (Liu et al., 26 Sep 2025).

The critique is then paired with the MOS to form a training example, where $c_i$4 is described as “the reasoning trajectory towards $c_i$5.” For scoring-time supervision, the framework uses an explicit tags-based schema: $\mathcal{D}=\{(\mathcal{I}_i,s_i), i\in[1,M]\},$0 The scoring prompt instructs the model to place “all your thinking or reasoning process in only one pair > ... tags” and the rating “inside only one pair of <answer>...</answer> tags” (Liu et al., 26 Sep 2025).

This design gives AesCoT a dual structure. First, it is multi-dimensional at the semantic level because critiques are organized around photography and aesthetic-criticism dimensions. Second, it is serialized at the output level because the model is trained to emit critique first and score second.

3. Filtering, quality control, and dataset configurations

The filtering process $c_i$6 enforces three stated qualities. The critique must show no score leakage, meaning it must not explicitly mention the numeric score or reveal that it was provided; reasoning–score alignment, meaning the tone and evaluation must be consistent with the MOS level; and no factual errors, meaning it should not describe non-existent elements or misinterpret obvious features (Liu et al., 26 Sep 2025).

The mechanism combines automatic checks, human audits, and programmatic filters. Other models are prompted to detect inconsistencies and factual errors, including whether the critique justifies the given score and whether it hallucinates image content. Human audits are used to catch subtle leakage and misalignment. Programmatic filters detect explicit numeric mentions and phrases such as “score 0.85,” “rating 0.7,” or references to hidden variables (Liu et al., 26 Sep 2025).

The pipeline is instantiated on several image aesthetic assessment datasets. The in-domain training sources are AVA, TAD66K, and FLICKR-AES, and the out-of-distribution evaluation datasets are PARA and AADB. Scores are normalized to $c_i$7, and AesCoT is balanced across three aesthetic levels: Bad $c_i$8, Fair $c_i$9, and Good $s_i$0 (Liu et al., 26 Sep 2025).

Two named dataset configurations are reported.

Configuration	Composition
AesCoT-3K	1000 bad, 1000 fair, 1000 good
AesCoT-10K	2928 bad, 3312 fair, 3660 good

The paper states that these are intended to be released publicly (Liu et al., 26 Sep 2025). It also states that subjectivity is handled by using MOS, by discretization into three levels, and by balanced sampling across levels. There is no explicit mathematical calibration of human annotator variance inside AesCoT itself; that calibration enters later through RAPO’s reward design (Liu et al., 26 Sep 2025).

4. Cold-start supervision and relation to RAPO

AesCoT enters Aes-R1 as the supervision source for the cold-start supervised fine-tuning stage. The policy is written as

$s_i$1

where $s_i$2 is the input image, $s_i$3 is the scoring prompt, $s_i$4 is the generated explanation, and $s_i$5 is the predicted numeric aesthetic score. The supervised objective is

$s_i$6

This stage fine-tunes the pretrained MLLM to reproduce both the critique text and the final score conditioned on prompt and image (Liu et al., 26 Sep 2025).

The paper characterizes this cold-start as the stage that instills aesthetic reasoning patterns, stabilizes the <think>/<answer> output format for RL, and yields an initial policy with reasonably calibrated scores. It further reports that 1 epoch of SFT on AesCoT is optimal for the later RL stage in terms of entropy and performance (Liu et al., 26 Sep 2025).

After cold-start, Aes-R1 applies Relative-Absolute Policy Optimization (RAPO). The high-level objective is

$s_i$7

with policy gradient

$s_i$8

and a PPO-like clipped objective with KL regularization. The reward decomposes as

$s_i$9

combining a relative ranking term and an absolute score-regression term (Liu et al., 26 Sep 2025).

The crucial distinction is explicit in the paper. AesCoT is data-centric, focusing on reasoning structure, dimensional coverage, interpretability, and score–critique alignment. RAPO is optimization-centric, focusing on numeric score calibration and pairwise ranking consistency. The paper states that RL reward is computed solely from the predicted score in <answer>, not from the textual explanation directly, while the trajectory as a whole is shaped by SFT on AesCoT (Liu et al., 26 Sep 2025).

This separation explains why AesCoT is presented as necessary for “teaching the model to generate structured explanations before scoring,” whereas RAPO is presented as the stage that improves PLCC and SRCC after that reasoning prior has been installed.

5. Empirical findings, ablations, and explanation quality

The principal quantitative claim is that Aes-R1 (AesCoT + RAPO) improves the backbone Qwen2.5-VL-7B from average PLCC/SRCC 0.4285 / 0.4589 to 0.6337 / 0.6186, corresponding to approximately 47.9% PLCC and 34.8% SRCC improvement over the backbone (Liu et al., 26 Sep 2025).

The SFT/RL initialization study isolates the contribution of AesCoT to reasoning quality. With 0 epochs SFT, RAPO, average PLCC/SRCC is 0.6297 / 0.6102, and the explanations are described as shallow, labeled “Aes-R1-Zero.” With 1 epoch SFT on AesCoT, RAPO, average PLCC/SRCC becomes 0.6337 / 0.6186, and the explanations are described as “expert-level aesthetic analysis.” The study also reports token entropy: 0.921 for SFT 0, 1.609 for SFT 1, 1.417 for SFT 2, and 0.705 for SFT 10. The paper concludes that moderate SFT gives exploration capacity while preserving acquired CoT patterns, whereas excessive SFT reduces entropy and harms RL generalization (Liu et al., 26 Sep 2025).

Reward ablations further distinguish AesCoT from RAPO. The paper compares Error-only, Rank-only, Binary-only, and combined rewards, and reports Error + Rank (RAPO) as the best configuration, with stronger performance and out-of-distribution generalization than its components alone. Since AesCoT is held fixed in these ablations, the reported differences are attributable to reward design rather than to changes in the reasoning corpus (Liu et al., 26 Sep 2025).

A case study reported in the paper illustrates the qualitative role of AesCoT. Compared with GPT-4.1, Qwen2.5-VL-7B, Q-Insight, VisualQuality-R1, and Aes-R1-Zero, the non-AesCoT models are described as over-scoring a sample image at 0.82–0.91 versus a ground truth of 0.613, while Aes-R1 outputs 0.59 and provides a balanced critique including both strengths and weaknesses, such as limited emotional impact and lack of innovation (Liu et al., 26 Sep 2025). This suggests that the AesCoT cold-start is not merely a formatting device; it conditions the model toward critique-aware numerical judgment.

6. Scope, limitations, and terminological ambiguity

The paper identifies several limitations and assumptions. AesCoT depends on existing score datasets such as AVA, TAD66K, and FLICKR-AES; relies on GPT-4.1 as a closed-source expert, which reduces reproducibility and may introduce bias; inherits the subjectivity and cultural contingency of MOS-based aesthetics; and incurs non-trivial filtering cost because human audits and model-based checks remain necessary (Liu et al., 26 Sep 2025). Its assumptions are likewise explicit: aesthetics can be described along five dimensions and captured by language rationales, MOS supplies a stable preference signal for RL, and moderate SFT on distilled reasoning provides a good prior for subsequent optimization (Liu et al., 26 Sep 2025).

The term itself is also potentially ambiguous in adjacent literature. In the cited corpus, AesCoT is explicitly defined as Aesthetic Chain-of-Thought only in Aes-R1 (Liu et al., 26 Sep 2025). Other works discuss related or similarly named ideas without using that exact definition. “Pretraining with Token-Level Adaptive Latent CoT” is presented as an AesCoT-style instance only in the broader sense of Adaptive/Efficient Chain-of-Thought, not as the formal name AesCoT (Zeng et al., 9 Feb 2026). A near-homonymous acronym, ASCoT, denotes Adaptive Self-Correction Chain-of-Thought for late-stage fragility in LLM reasoning (Zhang et al., 7 Aug 2025). Unrelated reinterpretations also appear in hardware security and software protection, where “AesCoT” is used informally for AES core tampering or AES on COTS with concealed secrets (Jain et al., 2020, Hossain et al., 2024), and in educational NLP where AES refers instead to Automated Essay Scoring and attention-derived topical components (Zhang et al., 2020).

Accordingly, in current research usage represented here, AesCoT is best treated as the data-construction and filtering pipeline inside Aes-R1 for multimodal aesthetic reasoning (Liu et al., 26 Sep 2025). Its distinctive contribution is the coupling of image aesthetic assessment to structured chain-of-thought supervision that is multi-dimensional, score-aligned, and explicitly filtered for leakage, misalignment, and hallucination.