Loose ImgCoT: Hybrid Reasoning Framework

• Loose ImgCoT is a hybrid reasoning framework that integrates spatially grounded visual tokens with key textual steps to compress lengthy reasoning chains.
• It optimizes token usage by selectively retaining steps with low model confidence, balancing global structure and fine-grained details.
• Empirical results reveal improved accuracy and generalization across math, science, and logic tasks with significantly reduced token counts.

Loose ImgCoT (“L-ImgCoT”) is a hybrid reasoning framework for LLMs that compresses long chains of thought by combining spatially-grounded visual latent tokens with a curated subset of explicit textual reasoning steps. This method allows efficient encoding of complex procedural reasoning while maintaining the ability to recover critical granular details, thus balancing global abstraction and local precision for downstream reasoning tasks. Loose ImgCoT leverages both spatial and linguistic inductive biases, yielding improved performance, lower token usage, and enhanced generalization compared to purely textual or rigidly structured alternatives (Chen et al., 30 Jan 2026).

1. Conceptual Framework and Motivation

Loose ImgCoT builds on the ImgCoT approach, which replaces the traditional textual chain-of-thought (CoT) input $c = (c_1, \dots, c_m)$ with a visual rendering $I$ of those reasoning steps, subsequently tokenized into a compact sequence $z = (z_1, \dots, z_n)$ of discrete visual tokens by a trained image tokenizer. The motivation for this shift is to supplant the strong linguistic inductive bias—preservation of surface word form and syntax—with a spatial bias, which instead supports retention of the global step layout and logical structure inherent in spatial visualizations. Purely visual latent tokens, however, often obscure fine-grained reasoning details that are critical for mathematical or logical precision.

To overcome this, Loose ImgCoT augments the compressed visual skeleton $z$ with a small selection of key textual steps $c'$ drawn from the original CoT. These steps are chosen automatically: those that the base LLM assigns low token log-likelihood (i.e., high uncertainty) are retained, while others are omitted or collapsed to ellipses. This design enables LLMs to preserve both compactness and necessary explicit information, with significantly fewer tokens than the full chain (Chen et al., 30 Jan 2026).

2. Formalization and Algorithm

Loose ImgCoT constructs the LLM input as a concatenation of the question $q$, visual latent tokens $z$, selected textual steps $c'$, and the target answer $a$: $$x = \bigl[(q),\,\langle\mathrm{BOT}\rangle,\,z_1,\dots,z_n,\,\langle\mathrm{EOT}\rangle,\,(c'),\,(a),\,\langle\mathrm{EOT}\rangle\bigr]$$ where $z_j$ are obtained from the TiTok encoder applied to the visual rendering of $c$.

Key step selection uses the pretrained LLM to compute token-wise log-likelihoods: $$\mathrm{conf}(t,i) = \log\left[\mathrm{softmax}(f_\theta(t_{<i}))_{t_i}\right]$$ A global threshold $y$ is estimated over a reference corpus. Each step $c_i$ with average confidence $\mathrm{conf}(c_i) < y$ is retained in $c'$; consecutive omitted steps are replaced with ellipses. This selection yields a concise textual supplement focused on steps likely to be error-prone for the model.

The composite fine-tuning loss consists of MSE on the visual tokens and cross-entropy (CE) on textual output: $$\mathcal{L} = \sum_{i=s_z}^{e_z}\mathrm{MSE}\bigl(f_\theta(x_{<i}),\,e_{z_i}\bigr) + \sum_{i=s_o}^{e_o}\mathrm{CE}\bigl(f_\theta(x_{<i}),\,x_i\bigr)$$

The summary pseudocode is:

Algorithm Loose_ImgCoT_FineTuning Input: pretrained LLM f_θ, VQVAE encoder 𝓔+codebook C, data {(q,c,a)}, threshold y Output: fine-tuned LLM for each example (q,c,a): 1. Render visual CoT: I ← render_text_with_boxes_and_arrows(c) 2. Encode visual tokens: h = 𝓔(I) for i=1…m: zᵢ = argmin_j ||hᵢ - e_j||² collect z = (z₁…zₙ) 3. Select key textual steps: for each step cᵢ in c: compute conf(cᵢ) = avg_j conf(cᵢ,j) keep those cᵢ with conf(cᵢ) < y, collapse others to “…” call result c′ 4. Tokenize and concatenate: x = [ Tokenize(q), BOT, z, EOT, Tokenize(c′), Tokenize(a), EOT ] 5. Compute loss ℒ as in (7) and backpropagate in f_θ end for

At inference, only the visual tokens and selected textual steps are required as input for the LLM to generate the target answer (Chen et al., 30 Jan 2026).

3. Design Considerations and Hyperparameters

Loose ImgCoT is characterized by strict hyperparameter choices:

• Visual token sequence length $n$ is typically 8.
• Codebook size is 4096; each embedding matches LLM hidden size.
• Threshold $y$ is tuned per model variant, e.g., $y=-1.5886$ for Qwen2.5-0.5B.
• Image rendering uses a 512×512 canvas with arrows to denote reasoning dependencies.
• Consecutive removed steps are represented as a single ellipsis.
• Full-parameter fine-tuning for models ≤1B parameters; LoRA (r=12, α=32) otherwise.

These choices enable the system to operate with a small memory and compute footprint while preserving essential logical information.

4. Empirical Performance and Analysis

Extensive evaluation across GSM8K (math), MATH (competition math), GPQA-Extended (science QA), and ProsQA (logical reasoning) benchmarks demonstrates that Loose ImgCoT achieves high accuracy with dramatically reduced inference token count relative to standard text-based CoT. For instance, on Qwen2.5-0.5B:

Method	MATH Acc	#Tok	GSM Acc	#Tok	GPQA Acc	#Tok	ProsQA Acc	#Tok
Full-CoT	9.2	149.4	16.9	116.3	34.5	150.2	94.6	71.3
ImgCoT	9.8	8.0	9.2	8.0	34.5	8.0	97.4	8.0
L-ImgCoT	10.1	102.8	17.5	64.7	38.1	89.3	98.6	40.9

Consequently, ImgCoT matches or outperforms text CoT using only 8 latent tokens (compared to over 100 for text). Loose ImgCoT further improves accuracy, particularly for cases where step-level granularity is required. Ablations confirm that spatially-grounded visual tokens generalize better out-of-domain than text-compressed latent reasoning alone, particularly for problems requiring non-local logical consistency or reasoning across complex step dependencies (Chen et al., 30 Jan 2026).

5. Comparative Analysis and Extensions

A key property of Loose ImgCoT is the hybridization of spatial and linguistic cues, leveraging the global reasoning skeleton (visual latent tokens) and targeted step-level detail (key textual steps). This stands in contrast to methods that rely exclusively on text (introducing a strong linguistic bias) or exclusively on visuals (potential blurring of essential operations).

The efficacy of explicit spatial layout is evidenced by small but measurable drops in performance when arrows or layout cues are omitted. Degradation as the token count is reduced is gradual for ImgCoT, while text-only baselines experience more rapid performance collapse.

Loose ImgCoT’s approach mirrors broader trends in chain-of-thought guided multi-modal generation. For example, the “loose” CoT strategies in image synthesis (ImageGen-CoT (Liao et al., 25 Mar 2025)) and loosely-coupled T2I instruction decomposition paradigms (DeCoT-inspired (Lin et al., 17 Aug 2025)) all prioritize intermediate, flexible, and multimodal “thought trace” representations over monolithic or rigid end-to-end mappings.

6. Applications and Generalization

Loose ImgCoT is primarily evaluated in the context of LLM-based mathematical reasoning, science QA, and logic tasks where both global structure and stepwise justification are essential. Its methodology enables generalization across task domains and backbones (Qwen2.5, LLaMa3.2), with design principles that allow adaptation to both encoder- and decoder-centric architectures. The spatial bias conferred by visual tokenization confers strong inductive generalization, especially in non-English, non-standard, or out-of-domain settings, with up to 30–80% relative improvement versus text-compressed latent approaches when tested out-of-distribution.

7. Limitations and Prospective Directions

Although Loose ImgCoT achieves a favorable compromise between efficiency and precision, it retains some dependence on rendering quality and the ability of visual tokens to represent arbitrarily complex logical operations. Reasoning steps with highly specific textual computations may still be imperfectly compressible in the visual latent space without occasional reference to preserved text. A plausible implication is that continuous improvements in visual tokenizers or automatic selection of stepwise textual details may further enhance the utility of the approach. The method is modular and readily extensible to broader chain-of-thought workflows in vision-language and T2I reasoning, where dynamic, feedback-driven, or asynchronously composed representations are required (Chen et al., 30 Jan 2026, Liao et al., 25 Mar 2025, Lin et al., 17 Aug 2025).