English-Centered Latent Reasoning Pathway

• English-centered latent reasoning pathway is a mechanism in which large language models internally process information following patterns learned from English chain-of-thought data.
• This approach is evaluated using metrics like cosine similarity and latent accuracy scores, highlighting robust performance in high-resource languages.
• Reliance on an English trajectory introduces vulnerabilities in low-resource languages due to translation errors and diminished native reasoning capabilities.

An English-centered latent reasoning pathway refers to the property of large reasoning models (LRMs) and LLMs whereby internal, nonverbal computations (“latent reasoning”) are structurally and functionally anchored to the English language, regardless of the input or output language. Empirical evidence demonstrates that, when tackled in any of dozens of languages, both predictive dynamics and latent state evolution within such models proceed along a trajectory that is strongly aligned with the pathway used for English chain-of-thought (CoT) reasoning. This phenomenon yields robust latent reasoning for resource-rich and English-adjacent languages, but manifests as diminished reasoning capability and increased brittleness for low-resource and structurally distant languages. The English-centered pathway underpins a spectrum of architectures, from vanilla transformers to models augmented with adapters, RL objectives, or latent-space reasoning modules, and constrains the development of genuinely multilingual latent reasoners.

1. Conceptual Framework and Formalization

An English-centered latent reasoning pathway is characterized by a model’s tendency to converge, at the latent-variable and hidden-state level, onto a canonical sequence of transformations originally acquired through English training data and English CoT supervision. Formally, for a question $x$ in language $l$ with a CoT trace $(t_1, \dots, t_T)$ and hidden states $h_{l,k}$, empirical work demonstrates that the stepwise latent accuracy $A_{l,k} = \Pr(\hat y = y \mid h_{l,k})$ and hidden-state trajectories $\{h_{l,k}\}_{k=1}^T$ closely match those for English inputs and traces, $h_{\mathrm{en},k}$, across layers and tokens (Liu et al., 6 Jan 2026).

Layer-wise logit-lens and cosine similarity analyses reveal that for each $k$, the cosine similarity $\mathrm{CosSim}(h_{l,k}, h_{\mathrm{en},k})$ is consistently high ($0.90$–$0.98$ for high-resource languages), indicating that multilingual inputs are mapped into an English-aligned internal solution trajectory prior to answer extraction (Liu et al., 6 Jan 2026). This induces a functional unification: even CoT written in, e.g., French or Chinese, is internally “translated” to an English-aligned code before further reasoning.

2. Experimental Evidence and Metrics

The existence and extent of English-centered latent reasoning are validated through quantitative metrics and probing protocols:

• Truncation-based latent prediction: By truncating CoT traces at various ratios $r$ and decoding answers directly from the current hidden state $h_{l, \lfloor rT \rfloor}$, one measures the area under accuracy curves ($\mathrm{AUTC}_l$) and the latent reasoning score $\mathrm{LRS}_l = \int_{0}^{1}A_l(r)[1-g_l(r)]dr$ (Liu et al., 6 Jan 2026). High $\mathrm{LRS}_l$ at low $r$ indicates early latent answer formation.
• Hidden-state alignment: Cosine similarity trajectories across layers and steps reveal strong alignment to the English pathway for high- and mid-resource languages, and weaker but still substantial alignment for low-resource languages (Liu et al., 6 Jan 2026).
• Cognitive attribute prevalence: Reasoning traces in English demonstrate higher frequencies of sub-goal setting, verification, backtracking, and backward chaining compared to native-language traces. Quantitatively, cognitive behavior rates such as $\mu_{BT}^{English} \approx 0.8$ versus $\mu_{BT}^{native} \approx 0.2$ further support the centrality of the English pathway (Saji et al., 23 Oct 2025).

These observations hold across diverse LLM families (e.g., Qwen, DeepSeek-R1-Distill-Llama) and reasoning benchmarks (MGSM, AIME, GPQA Diamond, PolyMath) (Liu et al., 6 Jan 2026, Saji et al., 23 Oct 2025, Zhang et al., 8 Oct 2025).

3. Model Architectures and Training Protocols

The English-centered pathway arises across multiple architectural paradigms:

• Plain, monolingual-finetuned transformers: Models pretrained predominantly on English (or using English-dense web corpora) and then fine-tuned with English CoT data naturally form the canonical latent pathway (Liu et al., 6 Jan 2026).
• RL policies with cross-lingual CoT alignment: Methods such as M-Thinker impose a cross-lingual thinking alignment (CTA) reward, which explicitly compares the non-English CoT to the model’s own English CoT for the same question, transferring reasoning capability from English to other languages without changing the underlying latent pathway (Zhang et al., 8 Oct 2025).
• Latent-space and shortcut-based architectures: Adaptive shortcut models (e.g., System-1.5 Reasoning) distill explicit English CoTs into latent hidden-state representations and train routers/adapters to dynamically allocate computation along the English-rooted pathway, irrespective of source language (Wang et al., 25 May 2025).
• Pivoted CoT training: For low-resource languages, English-Pivoted CoT Training enforces English intermediate reasoning (CoT) with the answer output in the target language, maximizing latent alignment and mitigating performance collapse seen with native-language CoT (Tran et al., 2 Apr 2025).

No additional transformer layers or explicit architectural changes are strictly necessary to induce the English-centered pathway; instead, it emerges as an inductive artifact of data and objective imbalance.

4. Performance Trends and Resource Dependence

The alignment to an English-centered latent pathway is strongest in high-resource languages (e.g., English, French, Chinese, German, Spanish, Russian), as measured by:

Language Category	CosSim(h, h_en)	AUTC (latent accuracy)	Effects of English CoT
High-resource	0.90–0.98	High	Strong latent reasoning
Mid-resource (BN, JA)	0.85–0.92	Moderate	Conditional alignment
Low-resource (SW, TE)	0.75–0.85	Low	Weaker, delayed latent

On simple tasks (MGSM), high-resource languages yield non-trivial pass@1 even at $r=0\%$ (e.g., $A_{\mathrm{en}}(0) \approx 0.2$). For complex benchmarks (Multilingual AIME), $A_l(r)$ remains near zero for all but the final steps, and LRS vanishes (Liu et al., 6 Jan 2026). For low-resource languages, nearly the entire explicit CoT is required to reach comparable latent accuracy.

Empirically, English-pivoted approaches yield absolute improvements of up to +28.33% on low-resource benchmarks (e.g., Irish AIME2024), demonstrating the direct utility of anchoring to the English pathway (Tran et al., 2 Apr 2025). However, for high-resource languages already well-represented in training, enforcing additional English pivoting can introduce mild interference.

5. Mechanistic Interpretations and Causes

Three broad factors explain the emergence of English-centered latent reasoning:

• Pretraining data bias: The overwhelming prevalence of English in pretraining corpora leads models to optimize parameters primarily for English-centric patterns (Liu et al., 6 Jan 2026, Saji et al., 23 Oct 2025).
• Supervised CoT imbalance: Most fine-tuning datasets for chain-of-thought reasoning are constructed in English, causing latent reasoning modules to adapt to English trajectories and representations (Liu et al., 6 Jan 2026).
• Architectural/functional consolidation: Transformer models tend to learn a single high-capacity latent pathway for reasoning, attaching minimal language-specific adapters around a “core” reasoning module. Thus, other languages are mapped into an English-resembling intermediate state before the actual latent computation proceeds (Liu et al., 6 Jan 2026, Zhang et al., 8 Oct 2025).

A plausible implication is that multilingual reasoning, as currently implemented, is not truly parallel or language-native but is instead mediated through a dominant English-centric “latent code.”

6. Implications, Limitations, and Future Directions

The English-centered latent pathway enhances coherence, planning, and verification in reasoning for certain languages, but introduces systematic risks:

• Vulnerability to translation failures: Translation steps introduce the “Lost in Translation” failure mode, where errors are caused by misalignment between source and English representations. On MGSM, the Lost-in-Translation fraction for errors decreases from ≈0.77 (low-resource) to ≈0.30 (high-resource) (Saji et al., 23 Oct 2025).
• Fragility for low-resource languages: Reliance on an English pivot impairs performance and limits model robustness when handling underrepresented languages, as semantic nuances and reasoning patterns may be collapsed or distorted (Liu et al., 6 Jan 2026, Tran et al., 2 Apr 2025).
• Limits on genuine multilinguality: Even with cross-lingual alignment objectives, models often transfer English-centric reasoning rather than building native-language latent trajectories (Zhang et al., 8 Oct 2025).

Future technical strategies include expanding CoT supervision directly in target languages, investigating architectural mechanisms for parallel latent pathways, and developing cross-lingual adapters to facilitate language-specific latent computation (Liu et al., 6 Jan 2026, Zhang et al., 8 Oct 2025, Tran et al., 2 Apr 2025). Measuring mutual information $I(h_t; y)$ across languages and reasoning steps is suggested as a route to more granular diagnosis of answer emergence in latent space (Liu et al., 6 Jan 2026).

7. Summary Table: Core Metrics for English-Centered Latent Reasoning

Metric	Description	Reference
$A_{l,k}$	Stepwise latent accuracy at token $k$ in language $l$	(Liu et al., 6 Jan 2026)
$\mathrm{CosSim}(h_l, h_{en})$	Cosine similarity of latent states to English	(Liu et al., 6 Jan 2026)
$\mathrm{AUTC}_l$	Area under truncation latent accuracy curve	(Liu et al., 6 Jan 2026)
Latent Reasoning Score (LRS)	Correct predictions before explicit answer present	(Liu et al., 6 Jan 2026)
Cognitive attribute rates ($\mu_X$)	Frequency of reasoning behaviors (e.g., BT, V, SG, BC)	(Saji et al., 23 Oct 2025)
Lost-in-Translation rate	Fraction of errors caused by translation-induced failure	(Saji et al., 23 Oct 2025)

High values of these metrics for English and English-aligned traces, and their consistent propagation to other languages conditioned on explicit English CoT, provide robust evidence of the English-centered latent reasoning pathway. Recognition and mitigation of this phenomenon are central to advancing truly multilingual reasoning models.