Critical Reasoning Units (CRUs)

• CRUs are modular constructs that integrate reasoning generation with immediate self-critique for validating intermediate propositions.
• They enable iterative problem-solving by interleaving thought and critique, essential for both textual and multimodal reasoning tasks.
• CRU frameworks improve model performance significantly through structured reinforcement learning and clear error localization.

A Critical Reasoning Unit (CRU) is a modular construct integrating granular reasoning, proposition verification, and self-evaluation, forming the backbone of advanced LLMs and multimodal reasoning systems. CRUs enable iterative, stepwise problem-solving by fusing "think" and "critique" capabilities, provide explicit intermediate proposition structuring in multimodal domains, and constitute a rewardable unit for reinforcement learning. Recent frameworks formalize CRUs as atomic interleaved units containing both generation and immediate self-scrutiny or proposition validation, tightly coupled to the cognitive processes underlying robust and interpretable reasoning (Xu et al., 17 Dec 2025, Wang et al., 16 Dec 2025).

1. Formal Definition of Critical Reasoning Units

In the Stepwise Think-Critique (STC) framework (Xu et al., 17 Dec 2025), a CRU is the minimal unit comprising a pair $(r_t, c_t)$, where $r_t$ denotes an autoregressively generated reasoning step, and $c_t$ is its immediate post hoc critique and binary correctness assessment $s_t \in \{0, 1\}$. The trajectory for a problem $x$ is written as:

$\tau = (r_1, c_1, r_2, c_2, \dots, r_T, c_T)\,,$

with each $(r_t, c_t)$ constituting one CRU.

In multimodal settings such as ViRC (Wang et al., 16 Dec 2025), each CRU spans an intermediate proposition with on-demand visual grounding and a variable-length chunk of textual steps:

$$\mathrm{CRU}^{(i)} \triangleq (v^{(i)}, \{ s^{(i,1)}, s^{(i,2)}, \dots, s^{(i, m_i)} \})\,,$$

where $v^{(i)}$ is the visual context retrieved at step $i$, and $s^{(i,\ell)}$ are consecutive textual sub-steps advancing a coherent proposition.

CRUs unify generation and verification. In STC, this takes the form of natural language critique accompanied by a correctness signal; in ViRC, it manifests as multi-step textual justification, grounded in dynamic or newly acquired visual information.

2. Core Mechanisms and Structural Hierarchies

CRUs impose a two-tiered organization on reasoning processes:

• Intra-Unit Coherence: Each CRU's steps collectively suffice to establish the correctness of a single intermediate proposition, with explicit self-evaluation (STC) or verification against visual evidence (ViRC).
• Inter-Unit Integration: Transition between CRUs corresponds to advancing to the next key logical or perceptual node. In ViRC, selective invocation of visual tools (crop, scale, display) occurs at CRU boundaries, coordinating textual inference and perceptual evidence.

ViRC introduces four reasoning patterns—Planning, Verifying, Backtracking, and Reflecting—encoding strategic behaviors that guide the chunking and completion of CRUs.

3. Processing and Algorithmic Flow

The standard CRU-driven reasoning workflow unfolds as an alternating sequence of generation and critique (STC), or as multimodal proposition-grounded chunking (ViRC). The key procedural elements are summarized below:

Step	STC (LLM context)	ViRC (Multimodal context)
Generation	Produce $r_t$ reasoning tokens	Select next reasoning pattern/tool; update $v^{(i)}$
Immediate Evaluation	Output $c_t$—critique + score $s_t$	Generate $m_i$ textual sub-steps to consolidate $v^{(i)}$
Transition	Advance to next $(r_{t+1}, c_{t+1})$	Move to next CRU; update visual context/tool selection

This alternation ensures that each step of the solution trajectory is explicitly checked before proceeding, with self-evaluation contributing to learning. In ViRC, chunking facilitates explicit, proposition-aligned transitions with selective evidence gathering and validation (Wang et al., 16 Dec 2025).

4. Training Schemes and Reinforcement Objectives

Training with CRUs typically proceeds via multi-stage supervised and reinforcement learning:

• Supervised Finetuning (SFT): For both STC and ViRC, SFT is first applied—either to synthetic trajectories with critique annotation (STC) or to CRU-structured, pattern-tagged samples from the CRUX corpus (ViRC).
• Reinforcement Learning (RL): Both frameworks employ Group Relative Policy Optimization (GRPO), shaping the distribution over CRUs using reward signals:
  • Reasoning (trajectory-level correctness matching target answer).
  • Critique-consistency (STC: binary agreement between critique and ground truth; ViRC: pattern and multimodal alignment).
  • Format rewards (JSON/schema validity or tag structure).
  • Dense rewards for stepwise shaping (process not only end results).

Combining these, the objective aggregates advantages and penalizes policy divergence via KL-regularization against a reference policy:

$$\max_{\theta}\; \mathbb{E} \left[ \frac{1}{G} \sum_{k=1}^G \frac{1}{|\tau_k|} \sum_{t=1}^{|\tau_k|} \mathrm{clipPrimalDual}(\rho_{k,t} \widetilde{\mathcal{A}}_{k,t}) - \beta\, D_{\mathrm{KL}}(\pi_\theta \| \pi_{\rm ref}) \right].$$

(Xu et al., 17 Dec 2025, Wang et al., 16 Dec 2025).

CRU-structured rewards enable granular credit assignment—either for stepwise validity, critique accuracy, pattern alignment, or multimodal coherence.

5. Empirical Results and Measured Gains

Empirical evaluations demonstrate substantial improvements attributable to explicit CRU structuring.

LLM Reasoning (STC):

• The STC-GRPO model outperforms its backbone by up to +9.2% Pass@1 on math reasoning benchmarks (AIME24, AMC23, MATH-500, Minerva, OlympiadBench).
• Critique F1 for final answers exceeds 97% in SFT, with process-level F1 up to ~82% (TNR up to 58–68%) after RL with dense rewards.
• "Best-of-K via critique" scoring approaches Pass@N rates and surpasses majority voting by 3–26% absolute (Xu et al., 17 Dec 2025).

Multimodal Mathematical Reasoning (ViRC):

• ViRC-7B achieves 77.79% average accuracy (+18.8% over Qwen2.5-VL-7B-Instruct baseline) on GeoQA, MMStar-Math, and MathVista-Math; GeoQA improvement alone is +31.6%.
• Ablations reveal that skipping explicit CRU structuring, as well as omitting human-inspired patterns or RL reward components, consistently reduces accuracy—providing strong evidence for the necessity of explicit CRUs (Wang et al., 16 Dec 2025).

6. Interpretability, Error Profiling, and Limitations

CRUs provide human-interpretable, fine-grained traces through explicit labeling of intermediate steps as correct or incorrect, coupled with justifications. This enables error localization and process-level auditability. Qualitative analysis demonstrates early detection of logical errors within the reasoning process.

Documented limitations include:

• Incomplete critique accuracy (over-confident errors persist), especially in process-level judgments.
• Sparse, trajectory-level rewards lead to noisier stepwise judgments unless supplemented by dense shaping terms.
• Training costs restrict scaling to larger LLMs and preclude extensive hyperparameter tuning.
• In ViRC, error tracing across CRUs and cross-alignment with incorrect reasoning paths expose limitations in detector precision and error-correction priors (Wang et al., 16 Dec 2025, Xu et al., 17 Dec 2025).

7. Extensions and Prospective Research

Potential avenues for advancing CRU-based methodologies include:

• Scaling STC to larger LLMs and multimodal models, potentially utilizing joint architectures for decoupled "think" and "critique" towers.
• Richer critique outputs with graded confidence, suggestions for correction, or RLHF targeting critic sub-modules.
• Adaptive step-count learning via halting heads, enabling dynamic determination of CRU sequence length.
• For multimodal contexts, further incorporating cognitive patterns and direct human feedback on CRU quality.
• Enhanced CRU annotation corpora (e.g., CRUX) to improve error tracing, error localization, and sample diversity.

This suggests that Critical Reasoning Units serve as a unifying principle across modalities, supporting both robust solution derivation and explicit intermediate self-correction at scale in contemporary AI systems (Xu et al., 17 Dec 2025, Wang et al., 16 Dec 2025).