Anchored Chain-of-Thought Reasoning

• Anchored Chain-of-Thought (ACoT) is a structured reasoning method that ensures every step is directly grounded in observable visual evidence and verified through causal logic.
• It uses a triadic chain model linking an initiating event, a mediating step, and a logical outcome, effectively mitigating hallucinations in multimodal outputs.
• ACoT is benchmarked by MM-CoT, which employs distractor chains to systematically diagnose and enhance visual grounding and logical consistency in AI models.

Anchored Chain-of-Thought (ACoT) reasoning refers to a formal approach for ensuring that intermediate steps in a multimodal model's reasoning chain are both directly grounded in observable evidence and logically coherent. Unlike traditional Chain-of-Thought (CoT) methods that prioritize plausibility and generative fluency, ACoT emphasizes verification: every step in a reasoning sequence must be simultaneously anchored to input data (e.g., an image or video) and must adhere to causal and commonsense constraints. MM-CoT is the benchmark introduced to systematically evaluate this phenomenon, offering a discriminative alternative to generative CoT by requiring the selection of the singular sequence meeting both visual and logical criteria (Zhang et al., 9 Dec 2025).

1. Formal Underpinnings of Anchored Chain-of-Thought

Anchored CoT is operationalized using a triadic structure for reasoning chains:

$c = (A \rightarrow B \rightarrow C)$

Here, $A$ is an initiating condition (visible change/action within the scene), $B$ is a mediating event (visually supported intermediate step), and $C$ is the outcome (logical consequence under physical or commonsense laws).

The anchoring of a chain $c$ to visual input $V$ is specified by two orthogonal predicates:

• $\Phi_{\text{vis}}(c, V) = 1$ iff every event $E_i \in \{A, B, C\}$ is directly observable in $V$, precluding hallucinated content.
• $\Phi_{\text{log}}(c) = 1$ iff all transitions $E_i \rightarrow E_{i+1}$ respect causal and commonsense logic.

A reasoning chain $c^*$ qualifies as a valid, anchored CoT precisely when

$$\mathcal{V}(c^*|V) = \Phi_{\text{vis}}(c^*, V) \land \Phi_{\text{log}}(c^*) = 1 \quad \text{(Equation 1)}$$

This structure enables rigorous disentanglement of perceptual and inferential failure modes in model outputs (Zhang et al., 9 Dec 2025).

2. MM-CoT Benchmark Design and Distractor Construction

MM-CoT implements ACoT as a discriminative verification task, not a generative rationale. For each visual sample $V$, the benchmark provides one valid chain $c^*$ and multiple adversarial chains, each violating exactly one of the anchoring constraints:

• Visual inconsistency distractors ($\neg \Phi_{\text{vis}}$): plausible-seeming sequences that reference objects, attributes, or relations absent from the input.
• Logical incoherence distractors ($\neg \Phi_{\text{log}}$): factually grounded but causally invalid (e.g., temporally reversed, physically impossible, or spurious sequences).

Benchmark instances are systematically created in three stages, as follows:

Stage	Process Description	Automated/Manual
Valid-Chain Generation	GPT-4 o generates three grounded chains per sample	Automated
Distractor Generation	Counterfactual (visual) and causal perturbation (logic) strategies produce distractors	Automated
Verification	BERTScore filtering + two-stage expert validation	Automated + Manual

The dataset covers 5,615 images (Flickr30k) and 2,100 videos (ShareGPT4Video), stratified by reasoning difficulty from Easy to Extreme, with tiering based on inference step depth and, for video, both temporal span and motion complexity (Zhang et al., 9 Dec 2025).

3. Evaluation Protocols and Diagnostic Metrics

MM-CoT quantifies anchored reasoning fidelity using several metrics:

• End-to-End Chain Selection Accuracy:

$$\text{Acc} = \frac{1}{N} \sum_{i=1}^{N} 1[\hat{c}_i = c^*_i]$$

where the model must pick the sole chain simultaneously satisfying $\Phi_{\text{vis}}$ and $\Phi_{\text{log}}$ (Equation 2).

• Diagnostic Axes:
  • Visual-Grounding Verification: Fraction of unique rejection of visual-inconsistent distractors.
  • Logical-Coherence Verification: Fraction of unique rejection of logically flawed distractors.

Each distractor is labeled by the type of failure induced, directly supporting error diagnosis. Ablation studies demonstrate severe degradation in performance (−30% to −70%) when the image/video input is omitted, evidencing that MM-CoT cannot be resolved through language-alone shortcuts (Zhang et al., 9 Dec 2025).

4. Empirical Model Performance and Failure Patterns

Top vision–LLMs were tested across three reasoning paradigms: Direct Answer, Standard CoT, and Reflective Reasoning. Models evaluated include proprietary systems (GPT-5, Gemini-2.5-Pro, Claude-Sonnet-4, Grok-2-Vision-1212) and leading open-source models (Qwen2.5-VL-72B, LLaMA-3.2-90B, GLM-4.5V, LLaVA-1.5-7B, Idefics2-8B, InternVL3-8B/3.5, Ovis-2.5).

Empirically:

• The strongest system (Gemini-Pro) achieves approximately 61.8% accuracy on single-step image chains, declining to 43.9% on multi-step chains.
• Video tasks exhibit greater difficulty: on Medium-Hard splits, leading models display 40–60% accuracy (easy tier), 20% (hard), and near-zero on Extreme.
• Reflective CoT reasoning—where models critique and revise their own output—yields improvements of 5–12% absolute, particularly for long or complex video chains (e.g., GPT-5 rises from 7.5% to 13.8% on Extreme) (Zhang et al., 9 Dec 2025).

Qualitative error analysis reveals:

• For images, failures are dominated by semantic redundancy (42–54%), distraction by irrelevant details (∼20%), and over-reliance on drivable text.
• For video, frequent issues include non-causal attribute selection (34–44%), omission of direct causes (20–22%), and counterfactual inference errors.

5. Advancements, Insights, and Methodological Recommendations

A principal insight from MM-CoT is that generative fluency (producing plausible rationales) does not equate to genuine visual grounding or logical fidelity. The discriminative verification setting enables fine-grained diagnosis of model failures, separating perceptual hallucination from incorrect causal inference.

Reflective reasoning paradigms, where models are encouraged to iteratively critique and revise their own reasoning chains, systematically enhance ACoT performance, particularly on more challenging, long-horizon tasks.

Recommendations for advancing ACoT include:

• Structured Distractor Rejection: Integrating binary heads for $\Phi_{\text{vis}}$ and $\Phi_{\text{log}}$ verification into model architectures.
• Multi-Agent Debate: Employing adversarial or consensus-driven cross-chain reasoning, similar to textual self-critique, to surface subtle inconsistencies.
• Counterfactual Training: Augmenting with explicit adversarial perturbations (both visual and causal) to increase both fidelity and robustness.

MM-CoT establishes a concrete research program: reasoning chains must be tied to ground-truth evidence and enforced through causal-structural constraints, validated by discriminative selection rather than likelihood-based generation (Zhang et al., 9 Dec 2025).

6. Significance and Future Trajectories

MM-CoT demonstrates that current vision–LLMs exhibit marked gaps between surface-level generativity and deeply anchored reasoning. The low correlation between MM-CoT and prior benchmarks confirms that ACoT measures a unique joint capability—simultaneous visual and logical grounding.

A plausible implication is that further progress in multimodal AI requires approaches emphasizing not just plausible storytelling, but veridical, evidence-attached reasoning. Prospective ACoT systems are likely to benefit from modular verification layers, causal structure enforcement, and explicitly adversarial data augmentations. This suggests an ongoing transition for the field, away from purely generative evaluation toward hybrid generative–discriminative techniques to achieve faithful and coherent reasoning within the visual world (Zhang et al., 9 Dec 2025).