Paper deep dive
Native Multilingual Chain-of-Thought Reasoning in Low-Resource Southeast Asian Languages
Sean Gip Lim, William Chandra Tjhi, Hai Leong Chieu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/4/2026, 4:36:30 AM
Summary
The paper introduces Onramp-Sequence Cross-Distillation (OSCD), a post-training algorithm designed to enable native multilingual Chain-of-Thought (CoT) reasoning in low-resource Southeast Asian languages. OSCD addresses 'cross-lingual collapse' where models revert to English during complex reasoning by projecting high-resource reasoning trajectories into low-resource vocabulary subspaces via an integrated translator agentic loop. It employs joint-embedding semantic alignment to bridge cross-lingual representational gaps, preventing catastrophic forgetting. Evaluations on AIME25 and HMMT25 benchmarks show OSCD yields up to 3.2x improvements in native reasoning and significant linguistic debiasing compared to translation-only baselines.
Entities (11)
Relation Signals (8)
OSCD → addresses → Cross-lingual collapse
confidence 95% · To address these challenges, we introduce the Onramp-Sequence Cross-Distillation (OSCD)... ensuring the stable and efficient translation... bridging the pairwise cross-lingual representational gaps.
OSCD → evaluatedon → HMMT25
confidence 95% · Comprehensive evaluations using the AIME25 and HMMT25 benchmarks demonstrate that OSCD yields up to 3.2 times overall improvements
OSCD → evaluatedon → AIME'25
confidence 95% · Comprehensive evaluations using the AIME25 and HMMT25 benchmarks demonstrate that OSCD yields up to 3.2 times overall improvements
OSCD → usescomponent → L JEPA
confidence 93% · This is coupled with joint-embedding semantic alignment... we introduce a secondary objective L JEPA
OSCD → usescomponent → L CE
confidence 93% · Full-Sequence Cross-Entropy To drive native multilingual acquisition, we apply full-sequence log-likelihood (L CE )
OSCD → improves → Qwen3-VL-8B-Thinking
confidence 90% · Ours-Qwen3-VL-8B... 73.8±11.0... compared to Qwen3-VL-8B-Thinking... 83.3±4.7 (Any-CoT) but significantly higher Target-CoT scores indicating native reasoning improvement.
L JEPA → reduces → Cross-lingual collapse
confidence 90% · joint-embedding semantic alignment... bridging the pairwise cross-lingual representational gaps... contributes up to 6.4% improvements in linguistic debiasing
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large Language Models have achieved substantial progress in reasoning capabilities. Yet in low-resource native settings, many suffer from cross-lingual collapse, reverting to English during intermediate steps that require complex logical reasoning. This presents a cold-start bottleneck for policy optimization, whereas standard fine-tuning risks catastrophic forgetting due to cross-lingual representation drift. To address these challenges, we introduce the Onramp-Sequence Cross-Distillation (OSCD), a post-training algorithm that projects high-resource reasoning trajectories into low-resource vocabulary subspaces during generative training rollouts via an integrated translator agentic loop, ensuring the stable and efficient translation of dynamically generated reference samples for fine-tuning. This is coupled with joint-embedding semantic alignment of both reference and target-language reasoning traces, thereby bridging the pairwise cross-lingual representational gaps. Comprehensive evaluations using the AIME25 and HMMT25 benchmarks demonstrate that OSCD yields up to 3.2 times overall improvements in native Southeast Asian languages for mathematical reasoning, of which the joint-embedding semantic alignment component contributes up to 6.4% improvements in linguistic debiasing over translation-only baselines.
Tags
Links
- Source: https://arxiv.org/abs/2608.00533v1
- Canonical: https://arxiv.org/abs/2608.00533v1
Trouble viewing inline? Open PDF directly →
Full Text
72,240 characters extracted from source content.
Expand or collapse full text
Native Multilingual Chain-of-Thought Reasoning in Low-Resource Southeast Asian Languages Sean Gip Lim 1,2,3∗ , William Chandra Tjhi 2 , Hai Leong Chieu 3 1 Nanyang Technological University 2 AI Singapore 3 DSO National Laboratories li0002ip@e.ntu.edu.sg, wtjhi@aisingapore.org, chaileon@dso.org.sg Abstract Large Language Models have achieved substantial progress in reasoning capabilities. Yet in low-resource native settings, many suffer from cross-lingual collapse, reverting to English during intermediate steps that require complex logical reason- ing. This presents a cold-start bottleneck for policy optimiza- tion, whereas standard fine-tuning risks catastrophic forgetting due to cross-lingual representation drift. To address these chal- lenges, we introduce the Onramp-Sequence Cross-Distillation (OSCD), a post-training algorithm that projects high-resource reasoning trajectories into low-resource vocabulary subspaces during generative training rollouts via an integrated translator agentic loop, ensuring the stable and efficient translation of dy- namically generated reference samples for fine-tuning. This is coupled with joint-embedding semantic alignment of both ref- erence and target-language reasoning traces, thereby bridging the pairwise cross-lingual representational gaps. Comprehen- sive evaluations using the AIME25 and HMMT25 benchmarks demonstrate that OSCD yields up to 3.2 times overall improve- ments in native Southeast Asian languages for mathematical reasoning, of which the joint-embedding semantic alignment component contributes up to 6.4% improvements in linguistic debiasing over translation-only baselines. Code, Dataset, Model — https://github.com/SG-Lim/OSCD Introduction Large Language Models (LLMs) have achieved remark- able reasoning capabilities, largely driven by advancements in Chain-of-Thought (CoT) prompting and Reinforcement Learning (RL) (Wei et al. 2022; Shao et al. 2024). Never- theless, these advancements remain overwhelmingly condi- tioned on high-resource typographic regimes, predominantly English (Tran, O’Sullivan, and Nguyen 2026; Schut, Gal, and Farquhar 2025; Barua et al. 2026). Consequently, linguistic regions characterized by severe data deficits are left behind. Southeast Asia (SEA), for instance, represents a population of 671 million but suffers from a lack of training data for regional development of native frontier capabilities (Lovenia et al. 2024). To compensate for this resource asymmetry, mul- tilingual models inherently exhibit an English-centric bias, leveraging high-resource languages as a structural anchor ∗ Corresponding author. Chinese English Filipino Indonesian Tamil Thai Vietnamese Target Language 0 20 40 60 80 100 Pass@5 Score SmolLM3-3B Qwen3-4B-Thinking-2507 Qwen3-VL-8B-Thinking Qwen-SEA-LION-v4-8B-VL Ours-Qwen3-VL-8B Figure 1:Pass@5 evaluation of large language models on the AIME25 benchmark, with linguistic alignment verifications across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). for low-resource problem-solving (Schut, Gal, and Farquhar 2025). This creates a fundamental barrier to user accessibil- ity, wherein the ability to interpret the step-by-step reasoning trace remains paramount to critical domains such as educa- tion and research. Because the intermediate process of CoT generation reverts systemically to English following complex logical transitions (Park et al. 2026; Zhao et al. 2026a; Kang et al. 2026), non-English native speakers are placed at a se- vere disadvantage, undermining the true value and equity of artificial intelligence. Because many thinking-enabled models inherently strug- gle to generate low-resource native trajectories on complex, multi-step reasoning tasks (Figure 1), we identify a critical bottleneck when applying modern RL approaches, particu- larly policy optimization algorithms relying on pre-existing knowledge space and reasoning trajectories to optimize sam- pling efficiency (Yue et al. 2026; Kim et al. 2025). This presents a cold-start training problem, wherein absent of arXiv:2608.00533v1 [cs.CL] 1 Aug 2026 Figure 2: Graphical representation of the Onramp Sequence Cross-Distillation (OSCD) algorithm. successful native samples to trigger sufficient rewards, the model cannot optimize its policy away from high-resource English pathways. On the other hand, standard Supervised Fine-Tuning (SFT) or Continued Pre-Training (CPT) on cu- rated multilingual datasets introduces separate challenges, most notably the risk of catastrophic forgetting (Alexandrov et al. 2024; Aggarwal et al. 2024b; Liu and Niehues 2025). This is largely attributed to the per-language misalignment of model hidden state representations, causing the mapping of parallel texts with linguistic differences to disjoint em- bedding subspaces despite semantic similarities (Aggarwal et al. 2024a; Lim, Aji, and Cohn 2025; Li et al. 2025). As a result, naive training without bridging this gap exacerbates representation drift (Gurgurov et al. 2026), risking instead the established geometry of high-resource embeddings. To address both the representational bottleneck and cold- start RL dilemma, we propose a training framework that opti- mizes for semantic equivalence of generative rollouts across different languages, while expanding the knowledge search space for low-resource multilingual reasoning (Figure 2). Our summary of key contributions are as follows: • We introduce OSCD, a novel training framework de- signed to enable native multilingual CoT reasoning in low-resource regional languages, specifically SEA lan- guages. This involves the dynamic, cross-policy localiza- tion of generative rollouts for fine-tuning, coupled with joint-embedding semantic alignment bridging the repre- sentational differences of parallel texts across different languages. • Extensive experiments on open-ended, mathematical rea- soning tasks demonstrate the efficiency and robustness of our training approach. We prove that the reasoning capabilities of English-dominant models can be effec- tively transferred to low-resource settings using syntheti- cally localized data, therefore expanding the multilingual knowledge search space required for subsequent RL. • Comprehensive evaluations on theAIME25 andHMMT25 benchmarks across language variants show that models post-trained with OSCD outperform existing multilingual SEA model of a comparable scale. Our training frame- work preserves model intelligence in high-resource set- tings, while successfully addressing systemic linguistic biases that induce unwanted language fallbacks. Related Work Multilingual Reasoning Multilingual reasoning is a ma- jor NLP challenge, particularly for low-resource languages (Tran, O’Sullivan, and Nguyen 2026). Consequently, models often map decisions into an English-adjacent latent space be- fore generating target-language outputs (Schut, Gal, and Far- quhar 2025). This English-centric bias causes cross-lingual understanding failures, uneven reasoning quality, and ’cross- lingual collapse’ where intermediate CoT reverts to English under increased difficulty (Hwang et al. 2025; Park et al. 2026; Zhao et al. 2026a; Kang et al. 2026). To bridge this gap, several methods leverage English as an intermediary via translation, code-switching, cross- lingual distillation, or parallel fine-tuning (Chen et al. 2024; Kang et al. 2026; Chai et al. 2025; Barua et al. 2026; Zheng et al. 2026). Native reasoning approaches include mapping problems into language-agnostic symbolic spaces (Ranaldi and Pucci 2025), causal interventions to subtract language-specific hidden states (Zhao et al. 2026b), and us- ing reinforcement learning for language-consistency rewards (Hwang et al. 2025). Nevertheless, native target-language reasoning systematically lags English-pivoted approaches, suffering from language-specific generation errors and con- ceptual misunderstandings (Barua et al. 2026). Joint-Embedding Predictive Architecture The mapping of varying semantic contexts into a shared latent space aligns with the foundational principles of Joint-Embedding Predictive Architecture (JEPA) (Garrido et al. 2024; Maes Figure 3: Graphical representation of the agentic loop system for translation of long-context reasoning traces. et al. 2026). Huang, LeCun, and Balestriero (2026a) intro- duced LLM-JEPA, combining standard autoregression with an embedding-space prediction loss to align different seman- tic views of the same knowledge. Subsequent frameworks leverage this objective to bound hidden-state trajectories via geometric regularizers (Huang, LeCun, and Balestriero 2026b; Yuan 2026), decouple latent reasoning from token generation (Liu, Chen, and Woodruff 2026), and project query and document embeddings into shared spaces (Chen et al. 2026). Lim, Aji, and Cohn (2025) demonstrated the importance of aligning semantic properties in low-resource multilin- gual contexts, without which models typically default to disjoint, less accurate representations. Along the topic, con- trastive learning applied to monolingual English data have been shown to project cross-lingual representations into a shared, language-invariant space (Wang, Wu, and Neu- big 2022), whereas minimizing cross-lingual divergence be- tween probabilistic latent variables guides models to cap- ture structured, language-agnostic semantic representations (Sherborne, Hosking, and Lapata 2023). Methodology The OSCD framework comprises two components: (1) Roll- out with Sequence Mutation, of which dynamically syn- thesizes native-language reasoning trajectories to populate low-resource vocabulary subspaces, and (2) Multi-Objective Cross-Distillation, which enforces sequence-level semantic alignment while fine-tuning is carried out on completions with localized reasoning traces. Rollout with Sequence Mutation Given a prompt p targeting a low-resource language l ∈ L, the pipeline initializes by sampling a completion sequence c ∼ π φ (· | p) from a high-resource reference model π φ . To isolate its intermediate CoT, we decode the token se- quence into a text string s =D(c), whereD(·) is the decod- ing function. Using a predefined set of structural delimiters ∆ =δ open ,δ close (e.g., opening and closing think tags), we partition s into a reasoning string s r and an answer string s a : (s r ,s a ) = Split D(c), ∆ (1) To localize dynamically the reasoning process into the low-resource vocabulary subspaceV L , an external translation functionT is then applied to the generated reasoning string, yielding a low-resource native trace conditioned on l: s ′ r =T (s r ,l)(2) Thereafter, the text segments are projected back into the discrete token space while preserving topological boundaries of the original completion. Let E ∗ (·) denote the controlled encoding function. The mutated completion sequence c ′ is reconstructed via an ordered concatenation (⊕) of the re- encoded segments and their structural delimiters: c ′ = δ open ⊕E ∗ (s ′ r )⊕ δ close ⊕E ∗ (s a )(3) We constrain the translation functionT entirely to the rea- soning trace, preserving the teacher’s original answer string s a verbatim. This aims to establish a stable anchor, thus guar- anteeing that the underlying inferential trajectory remains bound to a known, valid outcome. Multi-Objective Cross-Distillation Full-Sequence Cross-Entropy To drive native multilin- gual acquisition, we apply full-sequence log-likelihood (L CE ) across the entirety of the student completion, initializing the student model π θ as its reference teacher π φ . This ensures π θ constructs its intermediate reasoning steps over the target mediumV L , while remaining anchored to a stable answer in close proximity to its original distribution: L CE =− 1 |c ′ | |c ′ |−1 X t=0 logπ θ (c ′ t | p,c ′ <t )(4) Joint-Embedding Semantic Alignment While L CE in- duces a language shift, naively fine-tuning on translated text risks further isolation and drifting of model representations into language-specific subspaces (Lim, Aji, and Cohn 2025). To avert this concern, we introduce a secondary objective L JEPA , referencing and adapting from Huang, LeCun, and Balestriero (2026a) to fulfil our goal of bridging representa- tional differences through latent-space alignment of specific pairwise tokens. Let h T k and h S k ′ denote the last-layer hidden states corre- sponding to the closing think tag (δ close ) at token indices k and k ′ , respectively, conditioned on native reasoning trace r ′ for the student model π θ and r for teacher model π φ . We minimize the cosine distance between these corresponding hidden states as follows: L JEPA = 1− h S k ′ · h T k ∥h S k ′ ∥h T k ∥ (5) We specifically target δ close , because it serves as an infor- mational bottleneck encapsulating the aggregated semantics of reasoning process prior to answer generation, while pre- serving token-level flexibility across different languages so the model is not overconstrained when generatively navigat- ing its internal representations. Consequently, the total loss is formalized as a multi- objective function: L total = λ CE L CE + λ JEPA L JEPA (6) where λ CE and λ JEPA are hyperparameter weights corre- sponding to vocabulary acquisition and cross-lingual seman- tic alignment, respectively. For post-training optimization, we train models for 1 epoch in BF16 precision using the AdamW optimizer with a cosine scheduled learning rate of 2× 10 −5 , a 0.1 warmup fraction, and a 1.0 gradient clip norm. Training sequences are capped at 4096 prompt tokens and 8192 completion tokens. We also keep inference parameters for generative rollouts and reason- ing benchmarks consistent with the default recommendations of the respective model developers, using a maximum com- pletion length of 81920 tokens for evaluations. Both λ CE and λ JEPA are set to 1.0 for simplicity of this study. Experimentation The primary objective of this study is to empirically validate our proposed OSCD training framework. To this end, we design our experiments to address the following core research questions: 1. Native Reasoning Accuracy: Can the framework demon- strate improved benchmark performance on open-ended, deterministic mathematical reasoning tasks for models with no prior native Southeast Asian CoT capabilities, conditioned on native language alignment? 2. Efficiency of Training: Does the framework provide im- proved training efficiency in terms of its required com- pute or dataset size? How does it compare to conventional supervised fine-tuning approaches that optimize multilin- gual performance on a massive SEA corpora? 3. Robustness and Generalizability: Is the framework ag- nostic to model size and model family? Furthermore, does the acquisition of native multilingual CoT reasoning capa- bilities introduce catastrophic forgetting in high-resource base languages (e.g., English)? Experimental Setup We evaluate our proposed approach using the base reasoning models SmolLM3-3B, Qwen3-4B-Thinking-2507, and Qwen3-VL-8B-Thinking. For our training dataset, we extract a lean subset of 70,000 question-only samples from the OpenMathReasoning-tir corpus, discarding the accompanied reasoning and answer traces. Each question is then paired with a user prompt presented natively in one of 7 target languages (ZH,EN,Fi,IN,TA,TH,VI), explic- itly instructing the model to generate its intermediate CoT in that matching language. To translate dynamically the long-context reasoning traces into specific target languages during training, we serve a lo- cal instance of Gemma-SEA-LION-v4-27B-IT, coupled with an agentic system T to optimize throughput and con- sistency (Figure 3). This offers a functional alternative to rejection sampling on dynamic datasets. The temperature for translation initializes at 0.0, incrementing by 0.1 with each failed attempt (capped at 0.4) for up to 10 tries. Temperatures for scoring and language classification were fixed at 0.0 to maintain deterministic evaluations. We compare our post-trained models against the base mod- els SmolLM3-3B, Qwen3-4B-Thinking-2507, and Qwen3-VL-8B-Thinking, as well as a multilingual model Qwen-SEA-LION-v4-8B-VL of comparable size supervised fine-tuned on 9 million samples across different SEA languages. We report the pass@5 and mean@5 scores to proxy reasoning intelligence and consistency, respectively, with standard deviations where applicable. Post-training experiments were conducted on an Ubuntu 22.04.5 LTS platform comprising four NVIDIA H200 GPUs (141,GB HBM3e VRAM) interconnected via NVLink. The software stack consists of PyTorch 2.9.0, DeepSpeed 0.18.4, HuggingFace Transformers 4.57.1, and vLLM 0.13.0. Total GPU runtime accumulated to 384 hours for SmolLM3-3B, 440 hours for Qwen3-4B-Thinking-2507, and 625 hours for Qwen3-VL-8B-Thinking, each inclusive of model training alongside the dedicated translator server. Evaluation Metrics To evaluate the native reasoning capabilities of LLMs across low-resource Southeast Asian languages, we introduce a comprehensive evaluation framework using theAIME25 and HMMT25 mathematical reasoning benchmarks (Zhang and Math-AI 2025; Dekoninck et al. 2026), each translated into 7 different languages (ZH,EN,Fi,IN,TA,TH,VI) to acco- modate a lack of multilingual benchmarks comprising all tar- get variants. The framework comprises two scoring modes, to evaluate both a model’s linguistic biases and instruction- following capabilities. Both modes evaluate on the same ex- act generated completions: • Any-CoT: In this mode, the model is scored solely on the correctness of its final parsed answer, regardless if its reasoning CoT exhibits linguistic drift to English or non-target languages. • Target-CoT: In this mode, the model is scored on the correctness of its final parsed answer, as well as its ad- herence to the target languages during CoT reasoning per user instructions. To analyze linguistic bias, we also define Linguistic Re- tention Index (LRI) as the probability a model successfully AIME25HMMT25 Any-CoTTarget-CoTAny-CoTTarget-CoT ModelPass@5 Mean@5 Pass@5 Mean@5 LRI Pass@5 Mean@5 Pass@5 Mean@5 LRI Qwen-SEA-LION-v4-8B-VL55.6±6.937.3±3.147.8±13.928.4±9.775.2 37.8±6.924.2±3.732.2±8.417.8±4.874.2 Qwen3-VL-8B-Thinking82.2±7.769.6±8.750.0±44.841.6±38.056.1 52.2±7.744.4±6.032.2±29.126.0±25.155.6 Ours-Qwen3-8B-L CE 56.7±14.536.9±16.037.8±31.023.3±24.957.6 42.2±15.427.3±10.531.1±22.719.2±16.756.4 Ours-Qwen3-8B-L CE+JEPA 58.9±13.936.0±16.240.0±26.024.2±26.759.4 35.6±10.221.1±9.427.8±17.115.0±15.457.0 Ours-Qwen3-8B-L CE (Agentic) 80.0±3.362.7±8.774.4±6.952.8±13.785.4 50.0±6.737.6±5.447.8±8.428.3±10.981.1 Ours-Qwen3-8B-L CE+JEPA (Agentic) 80.0±3.364.4±5.778.9±3.854.2±11.687.9 51.1±5.139.3±5.348.9±5.131.9±9.186.3 Table 1: Overall ablation performance comparison across AIME25, HMMT25 benchmarks evaluating Any-CoT vs Target-CoT setups, averaged across 3 target languages (ZH,EN,IN) comparing different training loss configurations. 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 6 × 10 1 (a) Total Backpropagated Loss CE CE+JEPA CE (Agentic) CE+JEPA (Agentic) 0.00.20.40.60.81.0 Epoch 10 0 (b) Gradient Norm CE CE+JEPA CE (Agentic) CE+JEPA (Agentic) 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 5 × 10 1 (c) Cross-Entropy Metric CE CE+JEPA CE (Agentic) CE+JEPA (Agentic) 0.00.20.40.60.81.0 Epoch 10 2 10 1 (d) Cosine Distance Metric CE CE+JEPA CE (Agentic) CE+JEPA (Agentic) Figure 4: Training dynamics comparing different training loss configurations using Qwen3-VL-8B-Thinking. maintains its logical derivation process natively using the target language, rather than collapsing into a high-resource fallback. Given a transition matrix M ∈R N×N , where the diagonal element M i,i denotes the probability that a model with target prompt language ℓ i maintains its reasoning na- tively within ℓ i , the LRI across a set of target languagesG is computed as: LRI = 100 |G| X ℓ i ∈G M i,i (7) Thereafter, a high LRI for a given|G| languages bounded between 0 and 100 (inclusive) reflects native reasoning capa- bilities robust from biased linguistic preferences, whereas a low LRI signals systemic fallback behaviour into non-target languages. For linguistic alignment verification of intermediate rea- soning steps, we implement LLM-as-a-judge using the model Gemma-SEA-LION-v4-27B-IT. We also truncate the first 70% and tail 10% of sequence lengths, leaving only a 20% critical snippet of reasoning traces for classification. This prevents false positives induced by data leakage, often caused by model regurgitation of the native prompts near the beginning and end of its reasoning completion. If language- mixing or switching were detected, the classification defaults strictly to high-resource languages, prioritizing English. Ablation Studies We conduct ablation studies using the base reasoning model Qwen3-VL-8B-Thinking, along with 4 distinct training configurations to isolate the core contributions of individual components within the OSCD framework, specifically the secondary cosine distance loss L JEPA as well as the agen- tic translator systemT . TheL CE configuration accounts for the non-agentic deterministic translation of fine-tuned rea- soning traces as a baseline referencing Chen et al. 2024 and Barua et al. 2026, less rejection sampling due to the dynamic nature of data localization involved. For efficiency of exper- imentation, we use a representative, downscaled variant of the primary training dataset. We construct a balanced sub- set of 7,500 samples, interleaved and uniformly distributed across 3 target languages (ZH,EN,IN) to enable a stream- lined analysis of performance trade-offs. Accordingly, the L CE+JEPA (Agentic) configuration outper- forms its other ablation variants, yielding the highest Target- CoT and LRI scores across both benchmarks (Table 1). With- out the agentic translator to suppress zero-shot translation ar- tifacts and variance, we observe increased noise in its respec- tive gradient norms that led to severe performance degrada- tions (Figure 4b), in addition to slower inference speeds by a factor of two. Similarly, without a secondary loss to mitigate cross-linguistic representational drifts, models consistently suffer from the divergence of its cosine distance metric (Fig- ure 4d), causing a wider spread in reasoning accuracies and lower LRI scores despite a small set of 3 languages. This validates both our initial hypothesis and the robustness of our training framework. Low-Resource Native Reasoning The robustness of model native reasoning performance in low-resource SEA languages improves substantially with OSCD post-training. As shown in Table 2, we observe an AIME25HMMT25 Any-CoTTarget-CoTAny-CoTTarget-CoT ModelPass@5 Mean@5 Pass@5 Mean@5 LRI Pass@5 Mean@5 Pass@5 Mean@5 LRI SmolLM3-3B41.9±15.623.2±10.412.9±23.66.8±13.027.4 22.9±7.612.0±5.57.1±13.13.8±7.827.3 Ours-SmolLM3-3B31.4±14.518.6±8.429.5±14.814.4±8.889.8 13.3±9.27.0±5.512.4±9.26.1±4.791.8 Qwen3-4B-Thinking-250783.8±2.369.0±8.022.9±39.218.9±33.126.9 52.9±5.241.0±5.015.2±26.110.9±19.125.8 Ours-Qwen3-4B85.2±2.669.0±6.268.6±13.742.5±16.759.7 54.8±2.641.7±4.341.0±13.823.4±12.955.5 Qwen-SEA-LION-v4-8B-VL 49.5±10.132.2±8.138.6±14.621.0±11.265.8 31.0±10.717.4±8.122.9±12.510.6±8.065.5 Qwen3-VL-8B-Thinking83.3±4.768.6±5.821.4±37.217.8±31.224.0 52.4±5.342.4±4.213.8±24.111.1±20.123.8 Ours-Qwen3-VL-8B73.8±11.060.9±13.668.6±14.047.5±17.986.0 43.8±12.233.2±10.541.4±13.725.6±11.784.9 Table 2: Overall main performance comparison across AIME25, HMMT25 benchmarks evaluating Any-CoT vs Target-CoT setups, averaged across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI) comparing base reasoning models and OSCD post- trained models. Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.920.080.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 SmolLM3-3B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.830.170.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.220.780.000.000.000.00 0.000.040.000.960.000.000.00 0.000.060.000.000.940.000.00 0.000.070.000.000.000.930.00 0.000.070.000.000.000.000.93 Ours-SmolLM3-3B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.840.160.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 Qwen3-4B-Thinking-2507 Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.490.510.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.660.340.000.000.000.00 0.000.180.000.820.000.000.00 0.010.650.000.000.340.000.00 0.000.640.000.000.000.360.00 0.000.320.000.000.000.000.68 Ours-Qwen3-4B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.500.500.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.800.200.000.000.000.00 0.000.250.000.750.000.000.00 0.020.110.000.000.870.000.00 0.030.290.000.000.000.680.00 0.000.400.000.000.000.000.60 Qwen-SEA-LION-v4-8B-VL Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.670.330.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 Qwen3-VL-8B-Thinking Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.770.230.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.500.500.000.000.000.00 0.000.060.000.940.000.000.00 0.000.100.000.000.900.000.00 0.000.080.000.000.000.920.00 0.000.060.000.000.000.000.94 Ours-Qwen3-VL-8B 0.0 0.2 0.4 0.6 0.8 1.0 Probability of Language Fallback Figure 5: Overall main language fallback matrices comparing base reasoning models and OSCD post-trained models across AIME25, HMMT25 benchmarks and 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). average 2–3 times overall improvement in both Pass@5 and Mean@5 Target-CoT scores across both AIME25 and HMMT25 mathematical benchmarks, reflecting the success- ful cross-lingual transfer of high-resource base capabilities into low-resource target languages (Table 2). Importantly, the baseline reasoning models exhibit an English sinking state as default fallback language, with a minor exception for Chi- nese language due to abundance in its pre-training web cor- pus (Figure 5). Our models, however, demonstrate consistent mitigation of this bias, achieving a 2.2–3.6× improvement in LRI relative to model baselines. With Any-CoT, a mild reduction in overall scores for the Ours-SmolLM3-3B and Ours-Qwen3-VL-8B models were observed, attributed to the alteration of its natural fall- back language. Isolated for high-resource languages, how- ever, the performance drops remain marginal compared to baseline (Figure 1). This suggests that foundational rea- soning capabilities remain well-preserved, with noticeable Any-CoT reductions largely driven by a linguistic shift to- ward a more unbiased, multilingual distribution. In contrast, Ours-Qwen3-4B demonstrated modest improvements un- der Any-CoT settings, accompanied by a 40% probability of non-target fallback behavior; indicating a continued ten- dency to optimize reasoning trajectories for accuracy via English, while concurrently benefiting from knowledge ex- pansion across low-resource settings. Training Efficiency and Generalizability A primary advantage of the OSCD training framework is sample efficiency, which leverages on stable and dy- namically localized synthetic completions to mitigate the scarcity of quality data in low-resource target lan- guages. To evaluate this, we compare our models against Qwen-SEA-LION-v4-8B-VL, a competitive baseline trained on 9 million high-quality instruction-text pairs span- ning English and SEA languages. On the other hand, our models utilize a significantly smaller dataset of only 70,000 samples, distributed uniformly across 7 languages. Despite a raw dataset size reduction of over 90%, our framework yields substantial improvements in LRI met- −60 0 60 −80 0 80 −100 −50 0 50 SmolLM3-3B Ours-SmolLM3-3B −60 −30 0 30 −120 −60 0 60 −50 0 50 100 Qwen3-4B-Thinking-2507 Ours-Qwen3-4B −200 −100 0 −100 −50 0 50 −50 0 50 Qwen3-VL-8B-Thinking Ours-Qwen3-VL-8B Figure 6: 3D t-SNE plots of layer-wise hidden activations comparing base reasoning models and OSCD post-trained models, using the AIME25 dataset across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI) as input corpus. The layers 1 to N are indicated by expanding marker sizes. −150 0 150 −150 0 150 −200 −100 0 100 SmolLM3-3B Ours-SmolLM3-3B −40 0 40 80 −50 0 50 −40 0 40 Qwen3-4B-Thinking-2507 Ours-Qwen3-4B −50 0 50 −60 0 60 −60 0 60 Qwen3-VL-8B-Thinking Ours-Qwen3-VL-8B Figure 7: 3D t-SNE plots of layer-wise hidden activations comparing base reasoning models and OSCD post-trained models, using the HMMT25 dataset across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI) as input corpus. The layers 1 to N are indicated by expanding marker sizes. rics (Table 2). Ours-Qwen3-VL-8B, in particular, sur- passesQwen-SEA-LION-v4-8B-VL in LRI by 30%, thus demonstrating the robustness of dynamic OSCD over static SFT pipelines for multilingual acquisition. To evaluate the generalizability of OSCD, we experiment across three parameter scales (i.e., 3B, 4B, and 8B), spanning two distinct model families (i.e., Qwen3 and SmolLM3) as well as both text-only and multimodal dense architectures. Although our compute constraints prevented scaling beyond 8B parameters, we observed consistent improvements across different models. Specifically, an average 2–3 times overall improvement in Target-CoT reasoning performance across different benchmarks, as well as an increase in LRI up to 3.6 times corresponding to the successfully neutralization of high-resource English bias present in baseline models. This provides a working foundation for subsequent post-training, particularly with open-ended downstream applications that involve regional and cultural contexts. Scaling Behaviour and Hidden Activations Our scaling analysis reveals a non-linear behavior, driven by long-context reasoning capabilities relative to model weight sizes. Ours-SmolLM3-3B, for instance, lacks the capacity for extended generations which limits its exposure duration to non-target linguistic fallbacks. A substantially higher LRI is hence achieved, though its Target-CoT improvements remain modest. Ours-Qwen3-4B, on the other hand, features ex- tended reasoning capabilities but remains bottlenecked by its compact size, therefore compromising on LRI to sustain the accuracy of its answers. Among all, Ours-Qwen3-VL-8B provides the optimal foundation, supporting the multilingual expansion of knowledge space while simultaneously resisting non-target linguistic fallbacks over prolonged generations. Figures 6 and 7 illustrate the layer-wise hidden-state acti- vations comparing base reasoning models and OSCD post- trained models, using the AIME25 and HMMT25 datasets across 7 target languages as input corpus, respectively. In gen- eral, it is observed the initial layers remain clustered and well- aligned between models before and after post-training, while geometric shifts occur predominantly within the middle-to- later layers. The observation reflects a greater dependency on generative steering along deeper layers for downstream benchmark improvements, as opposed to multilingual com- prehension across problems in low-resource languages, par- ticularly with baseline models of larger sizes and stronger default capabilities. With smaller models, however, substan- tial shifts can emerge along the early representational layers, indicating a fundamental alteration to how challenging prob- lems are semantically comprehended as new knowledge gets ingested. Conclusion To conclude, we present the OSCD algorithm which enables 2–3 times substantial improvements in Target-CoT reason- ing performance, conditioned on linguistic alignment verifi- cations across low-resource Southeast Asian languages. The enabling of native multilingual reasoning capabilities proved largely additive, with limited observations of catastrophic forgetting observed in high-resource base capabilities. Cru- cially, we extend the upper-bound intelligence potential for native CoT reasoning in low-resource languages, offering the open community an additional layer of training foundation for subsequent reinforcement learning. This empowers non- native English speakers the ability to leverage advanced AI capabilities, thereby improving system usability and broad accessibility. Future research will focus on integrating rein- forcement learning and cultural fine-tuning for downstream applications beyond deterministic reasoning. This includes scaling the approach to open-ended tasks that heavily rely on regional and cultural contexts. Acknowledgments This work was supported by the DSO-AISG Incentive Award. The views expressed are solely those of the authors and do not represent the opinions of DSO National Laboratories or AI Singapore. References Aggarwal, D.; Damle, S.; Goyal, N.; Lokam, S.; and Sitaram, S. 2024a. Towards Exploring Continual Fine-Tuning for En- hancing Language Ability in Large Language Model. In NeurIPS 2024 Workshop on Fine-Tuning in Modern Machine Learning: Principles and Scalability. Aggarwal, D.; Sathe, A.; Watts, I.; and Sitaram, S. 2024b. MAPLE: Multilingual Evaluation of Parameter Efficient Finetuning of Large Language Models. In Ku, L.-W.; Mar- tins, A.; and Srikumar, V., eds., Findings of the Association for Computational Linguistics: ACL 2024, 14824–14867. Bangkok, Thailand: Association for Computational Linguis- tics. Alexandrov, A.; Raychev, V.; Müller, M. N.; Zhang, C.; Vechev, M.; and Toutanova, K. 2024. Mitigating Catastrophic Forgetting in Language Transfer via Model Merging. In Al- Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Findings of the Association for Computational Linguistics: EMNLP 2024, 17167–17186. Miami, Florida, USA: Association for Computational Linguistics. Barua, J.; Eisape, S.; Yin, K.; and Suhr, A. 2026. Long Chain-of-Thought Reasoning Across Languages. arXiv:2508.14828. Chai, L.; Yang, J.; Sun, T.; Guo, H.; Liu, J.; Wang, B.; Liang, X.; Bai, J.; Li, T.; Peng, Q.; and Li, Z. 2025. xCoT: cross- lingual instruction timing for cross-lingual chain-of-thought reasoning. In Proceedings of the Thirty-Ninth AAAI Con- ference on Artificial Intelligence and Thirty-Seventh Con- ference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artifi- cial Intelligence, AAAI’25/IAAI’25/EAAI’25. AAAI Press. ISBN 978-1-57735-897-8. Chen, N.; Zheng, Z.; Wu, N.; Gong, M.; Zhang, D.; and Li, J. 2024. Breaking Language Barriers in Multilingual Mathematical Reasoning: Insights and Observations. In Al- Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Findings of the Association for Computational Linguistics: EMNLP 2024, 7001–7016. Miami, Florida, USA: Association for Computational Linguistics. Chen, T.; Xu, S.; Guo, F.; Wang, X.; Gu, Q.; Li, H.; and Ji, L. 2026. Efficient Rationale-based Retrieval: On-policy Distillation from Generative Rerankers based on JEPA. In Proceedings of the 2026 International Conference on Multi- media Retrieval, ICMR ’26, 1721–1729. ACM. Dekoninck, J.; Jovanović, N.; Gehrunger, T.; Rögnvaldsson, K.; Petrov, I.; Sun, C.; and Vechev, M. 2026. Beyond Bench- marks: MathArena as an Evaluation Platform for Mathemat- ics with LLMs. Garrido, Q.; Assran, M.; Ballas, N.; Bardes, A.; Najman, L.; and LeCun, Y. 2024. Learning and Leveraging World Models in Visual Representation Learning. arXiv:2403.00504. Gurgurov, D.; Baeumel, T.; van Genabith, J.; and Os- termann, S. 2026. Sparse Subnetwork Enhancement for Underrepresented Languages in Large Language Models. arXiv:2510.13580. Huang, H.; LeCun, Y.; and Balestriero, R. 2026a. LLM- JEPA: Large Language Models Meet Joint Embedding Pre- dictive Architectures. In The Fourteenth International Con- ference on Learning Representations. Huang, H.; LeCun, Y.; and Balestriero, R. 2026b. Semantic Tube Prediction: Beating LLM Data Efficiency with JEPA. In The 1st Workshop on Scaling Post-training for LLMs. Hwang, J.; Tanmay, K.; Lee, S.-J.; Agrawal, A.; Palangi, H.; Ayush, K.; Fiete, I.; and Liang, P. P. 2025. Learn Globally, Speak Locally: Bridging the Gaps in Multilingual Reasoning. arXiv:2507.05418. Kang, D.; Hwang, S.; Kim, D.; Kim, H.; and Lee, G. G. 2026. Why Do Multilingual Reasoning Gaps Emerge in Reasoning Language Models? arXiv:2510.27269. Kim, M.; Shrestha, A.; Shrestha, S.; Nepal, A.; and Ross, K. W. 2025. RLVR vs. Distillation: Understanding Accuracy and Capability in LLM Mathematical Reasoning. In The 5th Workshop on Mathematical Reasoning and AI at NeurIPS 2025. Li, Z.; Shi, Y.; Liu, Z.; Yang, F.; Payani, A.; Liu, N.; and Du, M. 2025. Language ranker: a metric for quantifying LLM performance across high and low-resource languages. In Proceedings of the Thirty-Ninth AAAI Conference on Ar- tificial Intelligence and Thirty-Seventh Conference on Inno- vative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelli- gence, AAAI’25/IAAI’25/EAAI’25. AAAI Press. ISBN 978-1-57735-897-8. Lim, Z. W.; Aji, A. F.; and Cohn, T. 2025. Language- Specific Latent Process Hinders Cross-Lingual Performance. arXiv:2505.13141. Liu, B. K.; Chen, Z. P.; and Woodruff, D. P. 2026. JEPA- Reasoner: Decoupling Latent Reasoning from Token Gener- ation. arXiv:2512.19171. Liu, D.; and Niehues, J. 2025. Conditions for Catastrophic Forgetting in Multilingual Translation. In Adelani, D. I.; Arnett, C.; Ataman, D.; Chang, T. A.; Gonen, H.; Raja, R.; Schmidt, F.; Stap, D.; and Wang, J., eds., Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025), 347–359. Suzhuo, China: Association for Com- putational Linguistics. ISBN 979-8-89176-345-6. Lovenia, H.; Mahendra, R.; Akbar, S. M.; Miranda, L. J. V.; Santoso, J.; Aco, E.; Fadhilah, A.; Mansurov, J.; Imperial, J. M.; Kampman, O. P.; Moniz, J. R. A.; Habibi, M. R. S.; Hudi, F.; Montalan, R.; Ignatius, R.; Lopo, J. A.; Nixon, W.; Karlsson, B. F.; Jaya, J.; Diandaru, R.; Gao, Y.; Amadeus, P.; Wang, B.; Cruz, J. C. B.; Whitehouse, C.; Parmonangan, I. H.; Khelli, M.; Zhang, W.; Susanto, L.; Ryanda, R. A.; Hermawan, S. L.; Velasco, D. J.; Kautsar, M. D. A.; Hendria, W. F.; Moslem, Y.; Flynn, N.; Adilazuarda, M. F.; Li, H.; Lee, J.; Damanhuri, R.; Sun, S.; Qorib, M. R.; Djanibekov, A.; Leong, W. Q.; Do, Q. V.; Muennighoff, N.; Pansuwan, T.; Putra, I. F.; Xu, Y.; Chia, T. N.; Purwarianti, A.; Ruder, S.; Tjhi, W.; Limkonchotiwat, P.; Aji, A. F.; Keh, S.; Winata, G. I.; Zhang, R.; Koto, F.; Yong, Z.-X.; and Cahyawijaya, S. 2024. SEACrowd: A Multilingual Multimodal Data Hub and Benchmark Suite for Southeast Asian Languages. In Al- Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 5155–5203. Miami, Florida, USA: Association for Computational Linguistics. Maes, L.; Lidec, Q. L.; Scieur, D.; LeCun, Y.; and Balestriero, R. 2026. LeWorldModel: Stable End-to- End Joint-Embedding Predictive Architecture from Pixels. arXiv:2603.19312. Park, C.; Kim, J.; Lee, J.; Bae, S.; Choo, J.; and Yoo, K. M. 2026. Cross-lingual Collapse: How Language-Centric Foun- dation Models Shape Reasoning in Large Language Models. arXiv:2506.05850. Ranaldi, L.; and Pucci, G. 2025. Multilingual Reasoning via Self-training. In Chiruzzo, L.; Ritter, A.; and Wang, L., eds., Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computa- tional Linguistics: Human Language Technologies (Volume 1: Long Papers), 11566–11582. Albuquerque, New Mexico: Association for Computational Linguistics. ISBN 979-8- 89176-189-6. Schut, L.; Gal, Y.; and Farquhar, S. 2025. Do Multilingual LLMs Think In English? In ICLR 2025 Workshop on Build- ing Trust in Language Models and Applications. Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Zhang, M.; Li, Y. K.; Wu, Y.; and Guo, D. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. CoRR, abs/2402.03300. Sherborne, T.; Hosking, T.; and Lapata, M. 2023. Optimal Transport Posterior Alignment for Cross-lingual Semantic Parsing. Transactions of the Association for Computational Linguistics, 11: 1432–1450. Tran, K.-T.; O’Sullivan, B.; and Nguyen, H. D. 2026. Reason- ing transfer for an extremely low-resource and endangered language: Bridging languages through sample-efficient lan- guage understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 33277–33286. Wang, Y.; Wu, A.; and Neubig, G. 2022. English Contrastive Learning Can Learn Universal Cross-lingual Sentence Em- beddings. In Goldberg, Y.; Kozareva, Z.; and Zhang, Y., eds., Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 9122–9133. Abu Dhabi, United Arab Emirates: Association for Computational Lin- guistics. Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; Chi, E. H.; Le, Q. V.; and Zhou, D. 2022. Chain- of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Con- ference on Neural Information Processing Systems, NIPS ’22. Red Hook, NY, USA: Curran Associates Inc. ISBN 9781713871088. Yuan, Y. 2026. Semantic Step Prediction: Multi-Step Latent Forecasting in LLM Reasoning Trajectories via Step Sam- pling. arXiv:2604.18464. Yue, Y.; Chen, Z.; Lu, R.; Zhao, A.; Wang, Z.; Yue, Y.; Song, S.; and Huang, G. 2026. Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? In The Thirty-ninth Annual Conference on Neural Information Processing Systems. Zhang, Y.; and Math-AI, T. 2025. American Invitational Mathematics Examination (AIME) 2025. Zhao, R.; Liu, Y.; Schuetze, H.; and Hedderich, M. A. 2026a. A Comprehensive Evaluation of Multilingual Chain- of-Thought Reasoning: Performance, Consistency, and Faith- fulness Across Languages. In Demberg, V.; Inui, K.; and Marquez, L., eds., Findings of the Association for Computa- tional Linguistics: EACL 2026, 5223–5247. Rabat, Morocco: Association for Computational Linguistics. ISBN 979-8- 89176-386-9. Zhao, W.; Guo, J.; Deng, Y.; Wu, T.; Zhang, W.; Hu, Y.; Sui, X.; Zhao, Y.; Che, W.; Qin, B.; Chua, T.-S.; and Liu, T. 2026b. When Less Language is More: Language-Reasoning Disen- tanglement Makes LLMs Better Multilingual Reasoners. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. Zheng, W.; Huang, X.; Liu, Z.; Vangani, T. K.; Zou, B.; Tao, X.; Wu, Y.; Aw, A.; Chen, N. F.; and Lee, R. K.-W. 2026. Adamcot: Rethinking cross-lingual factual reasoning through adaptive multilingual chain-of-thought. In Proceed- ings of the AAAI Conference on Artificial Intelligence, vol- ume 40, 33863–33871. Training Details Algorithm Pseudocode The details of training process are provided in Algorithm 1. Hyperparameters Table 3 provides the hyperparameter settings for post-training experiments. Inference parameters for generative rollouts and reasoning benchmarks were consistent with the default rec- ommendations of the respective model developers. HyperparameterValue Optimization: OptimizerAdamW Learning rate2× 10 −5 Training: Training epochs1 PrecisionBF16 Gradient clip norm1.0 Learning Rate Schedule: Scheduler typeCosine Warmup fraction0.1 Loss Weights: Cross-Entropy (λ CE )1.0 JEPA (λ JEPA )1.0 Sequence Lengths: Max prompt tokens (Training)4096 Max completion tokens (Training)8192 Max completion tokens (Benchmark)81920 Table 3: Hyperparameter settings. Compute Resources • Hardware: 4× NVIDIA H200 GPUs (141 GB HBM3e VRAM each) interconnected via NVLink. • Software: PyTorch 2.9.0, DeepSpeed 0.18.4, Hugging- Face Transformers 4.57.1, and vLLM 0.13.0. • GPU Runtime: 384 hours for SmolLM3-3B, 440 hours for Qwen3-4B-Thinking-2507, and 625 hours for Qwen3-VL-8B-Thinking. (Note: Total runtime accounts for both the target model training and local vLLM server for dynamic translations.) Prompt Templates This appendix provides example prompt templates and scaf- folds used within this study. Specifically, Table 4 provides the language classification prompt, Table 5 the zero-shot transla- tion template, and Table 6 the quality evaluation framework. All scaffolds were integrated into our agentic translator dur- ing training. For benchmarks involving language classifica- tion, we rely exclusively on Table 4. Supplementary Results Detailed Benchmarks This appendix provides benchmark results for both our ab- lation studies and main experiments. Tables 7–8 provide the ablation results for AIME25 and HMMT25, respectively, with Figures 8–9 illustrating language fallback matrices compar- ing different model training configurations across the 3 tar- get languages (ZH,EN,IN). Tables 9–12 provide the main results for AIME25 and HMMT25, respectively, with Fig- ures 10–11 illustrating language fallback matrices compar- ing different model training configurations across the 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). Training Dynamics Figure 12 plots the total backpropagated loss, gradient norm, cross-entropy, and cosine distance metrics for our main train- ing experiments. Hidden Representations We provide the hidden-state representations of models before and after OSCD post-training using t-distributed Stochastic Neighbor Embedding (t-SNE) visualizations. Figures 13–16 illustrate the 2D t-SNE projections across sequential layers, using 300 parallel texts spanning the 7 target SEA languages, subsampled from the FLORES-Plus devtest split. Algorithm 1: Onramp Sequence Cross-Distillation Require: Student Model π θ ; Reference Model π φ ; Translator SystemT Require: High-Resource Base LanguagesM =EN, ZH; Escape Tokens ∆ =t <think> ,t </think> Require: Low-Resource Prompt-Language PairsP =(p i ,l i ) B i=1 , whereL =l i B i=1 andL∩M =∅ Require: Loss Weights: λ CE = 1.0,λ JEPA = 1.0 1: for each training step do 2: 1. Rollout and Modify Reasoning 3: Sample reference completion IDs c i ∼ π φ (·| p i ) for i∈ [1,B] 4: for i = 1 to B do 5:if l i ∈M then continue 6:str_r i , str_a i ← split π θ .decode(c i ), delimiters =decode(t)| t∈ ∆ 7:str_r ′ i ←T .modify(str_r i , target_lang = l i )▷ Translate reasoning trace 8:tok r ′ i ,tok a i ← π θ .encode(str_r ′ i , str_a i , add_special_tokens=False) 9:c ′ i ← (t <think> if t <think> ∈ c i else∅)⊕tok r ′ i ⊕ t </think> ⊕tok a i ▷ Reconstruct completion IDs 10: end for 11: 2. Multi-Objective Loss Computation 12: Using TrainEngine, compute: 13:Logprobs P T and hidden states H T for c i given π φ , where k = pos(t </think> ∈ c i ) 14:Logprobs P S and hidden states H S for c ′ i given π θ , where k ′ = pos(t </think> ∈ c ′ i ) 15: for i = 1 to B do 16:Cross-Entropy 17: L (i) CE ←− 1 |c ′ i | P |c ′ i |−1 t=0 logP S (c ′ i,t | p i ,c ′ i,<t )▷ Completion w/ modified reasoning 18:Cosine Distance 19:h S i ,h T i ← H S i=k ′ ,H T i=k 20: L (i) JEPA ← 1− cos_sim(h S i ,h T i )▷ Let cos_sim(u,v) = u·v ∥u∥v∥ 21: end for 22: 3. Gradient Computation and Update 23: L total ← 1 B P B i=1 λ CE L (i) CE + λ JEPA L (i) JEPA 24: Update parameters: θ ← θ− η∇ θ L total 25: end for Prompt Template for Linguistic Classification [System] You are an impartial linguistic judge. Evaluate the provided text and determine the primary language utilized for its core communicative framework. Evaluation Guidelines: 1. Structural Focus: Focus exclusively on syntax, grammatical connectors, and structural framework. Ignore target object bias such as the specific data being analyzed, translated, or discussed. 2. Candidate Labels: Classify the language strictly into one of the following permissible categories: English, Chinese, Filipino, Indonesian, Tamil, Thai, Vietnamese. 3. Cross-Tier Resolution: If language-switching or multi-lingual mixing occurs between Tier 1 (English, Chinese) and Tier 2 (other languages), instantly disqualify Tier 2 and default to the Tier 1 language. 4. Intra-Tier Resolution: If both Tier 1 languages (English and Chinese) co-occur anywhere within the text, default strictly to English as the final structural fallback category. Begin the evaluation with a structured linguistic breakdown separating framework grammar from referenced data. If mixing occurred, explicitly justify the tier-resolution application. Conclude the response STRICTLY in this format: “Result: Language”. [Input Text] REASONING_TRACE [Evaluation] Table 4: Prompt template for LLM-as-a-judge classification of REASONING_TRACE language. Prompt Template for Zero-Shot Translation [System] You are a professional translator. Your primary job is to translate the provided text from its original source language into the designated target language: TARGET_LANG. Translation Guidelines: 1. Accuracy & Tone: Preserve the exact meaning, context, and original tone (formal, casual, technical, etc.). Maintain all semantic nuances without omission. 2. Localization: Adapt cultural references, idioms, numerical formats, dates, and currencies to native target language conventions seamlessly. 3. Fluency & Formatting: Ensure a natural, coherent flow. Maintain original formatting (bold, italics) and preserve proper nouns unless standard translations exist. 4. Incomplete Input: If the text is truncated or ends mid-sentence, use the preceding context to logically and structurally complete the translation. Return ONLY the finished translation. STRICTLY DO NOT include any meta-comments, placeholders, introductory text, or explanations. [Original Text] SOURCE_TEXT [Translated Text] Table 5: Prompt template for zero-shot translation of SOURCE_TEXT from source language to TARGET_LANG. Prompt Template for Parallel Text Evaluation [System] You are an impartial translation judge. Your primary job is to evaluate the quality of the provided translation based on accuracy, fluency, and adherence to structural constraints. Evaluation Guidelines: 1. Scoring Metric: Rate the translation on an integer scale from 1 to 10 using a standard rubric. Assess semantic accuracy, nuance preservation, and intent alignment. 2. Language Homogeneity: Ensure the translation remains completely homogeneous. Technical jargon or specific domain terminology may remain untranslated only if core logical connectors maintain the target framework. 3. Leakage Penalty: If the translation contains language leakage (e.g., untranslated prose, mixed-language sentences) violating homo- geneity, automatically award a strict final score of 1. 4. Format Compliance: Verify that the output strictly adheres to the requested markdown conventions, structural alignment, and tag parameters without introducing peripheral text. Begin the evaluation with a concise explanation detailing the specific strengths and weaknesses observed based on the rubric criteria. Conclude the response STRICTLY in this format: “Result: Score”. [Original Text] ORIGINAL_REASONING_TRACE [Translated Text] TRANSLATED_REASONING_TRACE [Evaluation] Table 6: Prompt template for LLM-as-a-judge quality evaluation of parallel text REASONING_TRACEs. Any-CoTTarget-CoT ModelChinese English IndonesianOverall Chinese English IndonesianOverall Pass@5 (%)↑ Qwen-SEA-LION-v4-8B-VL53.363.350.055.6±6.936.763.343.347.8±13.9 Qwen3-VL-8B-Thinking73.386.786.782.2±7.763.386.70.050.0±44.8 Ours-Qwen3-8B-L CE 46.773.350.056.7±14.523.373.316.737.8±31.0 Ours-Qwen3-8B-L CE+JEPA 43.370.063.358.9±13.926.770.023.340.0±26.0 Ours-Qwen3-8B-L CE (Agentic) 76.783.380.080.0±3.366.780.076.774.4±6.9 Ours-Qwen3-8B-L CE+JEPA (Agentic) 76.783.380.080.0±3.376.783.376.778.9±3.8 Mean@5 (%)± std↑ Qwen-SEA-LION-v4-8B-VL36.7±4.240.7±3.934.7±3.437.3±3.119.3±6.538.7±5.427.3±2.528.4±9.7 Qwen3-VL-8B-Thinking61.3±4.578.7±3.468.7±5.469.6±8.750.0±2.174.7±3.40.0±0.041.6±38.0 Ours-Qwen3-8B-L CE 26.7±8.255.3±4.528.7±5.836.9±16.013.3±4.751.7±3.75.0±3.723.3±24.9 Ours-Qwen3-8B-L CE+JEPA 25.3±5.054.7±6.228.0±7.236.0±16.29.2±2.855.0±3.78.3±3.724.2±26.7 Ours-Qwen3-8B-L CE (Agentic) 58.0±6.272.7±4.457.3±10.262.7±8.742.5±6.468.3±5.547.5±10.152.8±13.7 Ours-Qwen3-8B-L CE+JEPA (Agentic) 59.3±6.570.7±2.563.3±6.064.4±5.748.3±5.067.5±4.346.7±2.454.2±11.6 Table 7: Ablation performance on the AIME25 benchmark comparing different training loss configurations, evaluating Any-CoT vs Target-CoT setups across 3 target languages (ZH,EN,IN). Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.530.470.00 0.010.990.00 0.000.270.73 Qwen-SEA-LION-v4-8B-VL Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.680.320.00 0.001.000.00 0.001.000.00 Qwen3-VL-8B-Thinking Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.470.530.00 0.001.000.00 0.000.750.25 Ours-Qwen3-8B (L CE ) Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.500.500.00 0.001.000.00 0.000.720.28 Ours-Qwen3-8B (L CE+JEPA ) Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.690.310.00 0.001.000.00 0.000.130.87 Ours-Qwen3-8B (L CE (Agentic) ) Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.790.210.00 0.001.000.00 0.000.150.85 Ours-Qwen3-8B (L CE+JEPA (Agentic) ) 0.0 0.2 0.4 0.6 0.8 1.0 Probability of Language Fallback Figure 8: Ablation language fallback matrices on the AIME25 benchmark comparing different training loss configurations across 3 target languages (ZH,EN,IN). Any-CoTTarget-CoT ModelChinese English IndonesianOverall Chinese English IndonesianOverall Pass@5 (%)↑ Qwen-SEA-LION-v4-8B-VL43.340.030.037.8±6.933.340.023.332.2±8.4 Qwen3-VL-8B-Thinking43.356.756.752.2±7.740.056.70.032.2±29.1 Ours-Qwen3-8B-L CE 33.360.033.342.2±15.423.356.713.331.1±22.7 Ours-Qwen3-8B-L CE+JEPA 26.746.733.335.6±10.223.346.713.327.8±17.1 Ours-Qwen3-8B-L CE (Agentic) 50.056.743.350.0±6.746.756.740.047.8±8.4 Ours-Qwen3-8B-L CE+JEPA (Agentic) 50.056.746.751.1±5.150.053.343.348.9±5.1 Mean@5 (%)± std↑ Qwen-SEA-LION-v4-8B-VL26.0±4.926.7±4.220.0±2.124.2±3.714.7±3.423.3±5.615.3±1.617.8±4.8 Qwen3-VL-8B-Thinking38.0±4.050.0±4.745.3±3.444.4±6.028.0±4.550.0±4.70.0±0.026.0±25.1 Ours-Qwen3-8B-L CE 20.0±6.339.3±2.522.7±2.527.3±10.511.7±5.038.3±1.77.5±1.419.2±16.7 Ours-Qwen3-8B-L CE+JEPA 15.3±4.032.0±2.716.0±4.921.1±9.49.2±4.932.5±2.83.3±2.415.0±15.4 Ours-Qwen3-8B-L CE (Agentic) 40.7±3.340.7±1.331.3±1.637.6±5.423.3±8.540.8±1.420.8±6.028.3±10.9 Ours-Qwen3-8B-L CE+JEPA (Agentic) 37.3±3.345.3±3.435.3±4.039.3±5.326.7±7.842.5±2.826.7±2.431.9±9.1 Table 8: Ablation performance on the HMMT25 benchmark comparing different training loss configurations, evaluating Any-CoT vs Target-CoT setups across 3 target languages (ZH,EN,IN). Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.470.530.00 0.001.000.00 0.000.240.76 Qwen-SEA-LION-v4-8B-VL Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.670.330.00 0.001.000.00 0.001.000.00 Qwen3-VL-8B-Thinking Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.430.570.00 0.001.000.00 0.000.740.26 Ours-Qwen3-8B (L CE ) Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.420.580.00 0.001.000.00 0.000.710.29 Ours-Qwen3-8B (L CE+JEPA ) Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.660.340.00 0.010.990.00 0.000.220.78 Ours-Qwen3-8B (L CE (Agentic) ) Chinese English Indonesian Reasoning Language Chinese English Indonesian Question Language 0.710.290.00 0.001.000.00 0.000.120.88 Ours-Qwen3-8B (L CE+JEPA (Agentic) ) 0.0 0.2 0.4 0.6 0.8 1.0 Probability of Language Fallback Figure 9: Ablation language fallback matrices on the HMMT25 benchmark comparing different training loss configurations across 3 target languages (ZH,EN,IN). Any-CoT ModelChinese English Filipino IndonesianTamilThaiVietnameseOverall Pass@5 (%)↑ SmolLM3-3B30.060.043.346.713.346.753.3 41.9±15.6 Ours-SmolLM3-3B40.053.333.336.710.016.730.031.4±14.5 Qwen3-4B-Thinking-250780.086.783.386.783.383.383.383.8±2.3 Ours-Qwen3-4B86.786.780.086.783.386.786.785.2±2.6 Qwen-SEA-LION-v4-8B-VL53.363.350.050.030.046.753.349.5±10.1 Qwen3-VL-8B-Thinking73.386.783.386.783.383.386.783.3±4.7 Ours-Qwen3-VL-8B80.083.376.776.750.073.376.773.8±11.0 Mean@5 (%)± std↑ SmolLM3-3B14.7±2.738.0±2.727.3±5.726.7±4.25.3±1.626.0±6.824.7±5.423.2±10.4 Ours-SmolLM3-3B23.3±3.732.0±6.920.7±3.319.3±2.55.3±2.712.0±1.617.3±2.518.6±8.4 Qwen3-4B-Thinking-250760.0±3.080.7±2.569.3±4.473.3±3.057.3±5.371.3±5.470.7±2.569.0±8.0 Ours-Qwen3-4B73.3±4.772.7±3.372.0±4.066.0±4.456.0±3.372.0±1.671.3±4.569.0±6.2 Qwen-SEA-LION-v4-8B-VL 36.7±4.240.7±3.934.0±5.734.7±3.415.3±3.430.0±5.634.0±4.932.2±8.1 Qwen3-VL-8B-Thinking61.3±4.578.7±3.471.3±2.768.7±5.463.3±7.066.0±6.570.7±9.868.6±5.8 Ours-Qwen3-VL-8B64.7±3.474.7±4.064.0±5.766.7±6.032.0±5.058.7±2.765.3±2.760.9±13.6 Table 9: Main performance on the AIME25 benchmark evaluating Any-CoT, comparing base reasoning models and OSCD post-trained models across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). Any-CoT ModelChinese English Filipino IndonesianTamilThaiVietnameseOverall Pass@5 (%)↑ SmolLM3-3B16.733.323.326.710.026.723.322.9±7.6 Ours-SmolLM3-3B16.730.013.313.30.06.713.313.3±9.2 Qwen3-4B-Thinking-250750.056.750.056.743.356.756.752.9±5.2 Ours-Qwen3-4B56.756.756.753.353.350.056.754.8±2.6 Qwen-SEA-LION-v4-8B-VL43.340.030.030.010.030.033.331.0±10.7 Qwen3-VL-8B-Thinking43.356.753.356.746.753.356.752.4±5.3 Ours-Qwen3-VL-8B43.360.046.746.720.040.050.043.8±12.2 Mean@5 (%)± std↑ SmolLM3-3B6.7 ±2.120.7±1.310.7±4.914.7±2.74.0±1.313.3±3.714.0±2.512.0±5.5 Ours-SmolLM3-3B8.0±1.618.0±4.55.3±3.46.7±3.70.0±0.04.7±2.76.7±0.07.0±5.5 Qwen3-4B-Thinking-250739.3±4.948.7±3.434.0±6.144.0±4.436.0±3.943.3±3.042.0±7.541.0±5.0 Ours-Qwen3-4B46.0±4.946.0±3.940.0±2.144.7±4.534.0±7.140.0±3.741.3±4.041.7±4.3 Qwen-SEA-LION-v4-8B-VL 26.0±4.926.7±4.216.7±2.120.0±2.12.7±2.515.3±5.014.7±5.417.4±8.1 Qwen3-VL-8B-Thinking38.0±4.050.0±4.741.3±4.045.3±3.438.0±4.541.3±3.442.7±2.542.4±4.2 Ours-Qwen3-VL-8B40.7±2.544.0±3.935.3±3.438.0±3.412.7±1.327.3±2.534.7±6.233.2±10.5 Table 10: Main performance on the HMMT25 benchmark evaluating Any-CoT, comparing base reasoning models and OSCD post-trained models across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). Target-CoT ModelChineseEnglish Filipino IndonesianTamilThaiVietnameseOverall Pass@5 (%)↑ SmolLM3-3B30.060.00.00.00.00.00.0 12.9±23.6 Ours-SmolLM3-3B40.053.323.336.710.016.726.729.5±14.8 Qwen3-4B-Thinking-250773.386.70.00.00.00.00.022.9±39.2 Ours-Qwen3-4B73.386.760.080.046.760.073.368.6±13.7 Qwen-SEA-LION-v4-8B-VL36.763.316.743.326.743.340.038.6±14.6 Qwen3-VL-8B-Thinking63.386.70.00.00.00.00.021.4±37.2 Ours-Qwen3-VL-8B80.083.356.773.343.370.073.368.6±14.0 Mean@5 (%)± std↑ SmolLM3-3B13.3 ±2.134.0±3.30.0±0.00.0±0.00.0±0.00.0±0.00.0±0.06.8±13.0 Ours-SmolLM3-3B21.3±4.029.3±6.56.0±5.717.3±2.54.7±1.610.7±3.911.3±3.414.4±8.8 Qwen3-4B-Thinking-250752.7±2.579.3±3.90.0±0.00.0±0.00.0±0.00.0±0.00.0±0.018.9±33.1 Ours-Qwen3-4B42.7±10.270.7±3.927.3±5.754.0±3.322.7±2.532.7±5.747.3±3.942.5±16.7 Qwen-SEA-LION-v4-8B-VL 19.3±6.538.7±5.44.7±2.727.3±2.510.0±4.723.3±7.623.3±2.121.0±11.2 Qwen3-VL-8B-Thinking50.0±2.174.7±3.40.0±0.00.0±0.00.0±0.00.0±0.00.0±0.017.8±31.2 Ours-Qwen3-VL-8B50.0±4.774.0±4.925.3±5.458.0±5.423.3±4.753.3±3.048.7±5.047.5±17.9 Table 11: Main performance on the AIME25 benchmark evaluating Target-CoT, comparing base reasoning models and OSCD post-trained models across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.920.080.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 SmolLM3-3B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.850.150.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.310.690.000.000.000.00 0.000.040.000.960.000.000.00 0.000.060.000.000.940.000.00 0.000.080.000.000.000.920.00 0.000.070.000.000.000.000.93 Ours-SmolLM3-3B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.880.120.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 Qwen3-4B-Thinking-2507 Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.540.460.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.590.410.000.000.000.00 0.000.200.000.800.000.000.00 0.000.630.000.000.370.000.00 0.000.630.000.000.000.370.00 0.000.300.000.000.000.000.70 Ours-Qwen3-4B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.530.470.000.000.000.000.00 0.010.990.000.000.000.000.00 0.000.820.180.000.000.000.00 0.000.270.000.730.000.000.00 0.010.130.000.000.860.000.00 0.030.290.000.000.000.680.00 0.000.360.000.000.000.000.64 Qwen-SEA-LION-v4-8B-VL Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.680.320.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 Qwen3-VL-8B-Thinking Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.760.240.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.430.570.000.000.000.00 0.000.040.000.960.000.000.00 0.000.120.000.000.880.000.00 0.000.090.000.000.000.910.00 0.000.060.000.000.000.000.94 Ours-Qwen3-VL-8B 0.0 0.2 0.4 0.6 0.8 1.0 Probability of Language Fallback Figure 10: Main language fallback matrices comparing base reasoning models and OSCD post-trained models across the AIME25 benchmark and 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). Target-CoT ModelChinese English Filipino IndonesianTamilThaiVietnameseOverall Pass@5 (%)↑ SmolLM3-3B16.733.30.00.00.00.00.0 7.1±13.1 Ours-SmolLM3-3B13.330.010.013.30.06.713.312.4±9.2 Qwen3-4B-Thinking-250750.056.70.00.00.00.00.015.2±26.1 Ours-Qwen3-4B46.756.726.753.323.330.050.041.0±13.8 Qwen-SEA-LION-v4-8B-VL33.340.06.723.36.726.723.322.9±12.5 Qwen3-VL-8B-Thinking40.056.70.00.00.00.00.013.8±24.1 Ours-Qwen3-VL-8B43.360.033.346.716.740.050.041.4±13.7 Mean@5 (%)± std↑ SmolLM3-3B6.0 ±3.320.7±1.30.0±0.00.0±0.00.0±0.00.0±0.00.0±0.03.8±7.8 Ours-SmolLM3-3B6.7±2.115.3±5.44.0±3.96.7±3.70.0±0.04.7±2.75.3±1.66.1±4.7 Qwen3-4B-Thinking-250730.0±3.046.0±2.50.0±0.00.0±0.00.0±0.00.0±0.00.0±0.010.9±19.1 Ours-Qwen3-4B20.0±6.044.0±3.911.3±6.236.0±3.910.0±3.715.3±3.427.3±5.723.4±12.9 Qwen-SEA-LION-v4-8B-VL 14.7±3.423.3±5.61.3±1.615.3±1.61.3±1.610.7±6.87.3±3.910.6±8.0 Qwen3-VL-8B-Thinking28.0±4.550.0±4.70.0±0.00.0±0.00.0±0.00.0±0.00.0±0.011.1±20.1 Ours-Qwen3-VL-8B30.0±3.744.0±3.911.3±4.531.3±3.410.7±2.525.3±3.426.7±4.725.6±11.7 Table 12: Main performance on the HMMT25 benchmark evaluating Target-CoT, comparing base reasoning models and OSCD post-trained models across 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.910.090.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 SmolLM3-3B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.810.190.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.140.860.000.000.000.00 0.000.040.000.960.000.000.00 0.000.070.000.000.930.000.00 0.000.060.000.000.000.940.00 0.010.070.000.000.000.000.92 Ours-SmolLM3-3B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.800.200.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 Qwen3-4B-Thinking-2507 Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.450.550.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.730.270.000.000.000.00 0.000.160.000.840.000.000.00 0.010.670.000.000.310.000.00 0.000.650.000.000.000.350.00 0.000.350.000.000.000.000.65 Ours-Qwen3-4B Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.470.530.000.000.000.000.00 0.001.000.000.000.000.000.00 0.010.780.210.000.000.000.00 0.000.240.000.760.000.000.00 0.030.090.000.000.880.000.00 0.020.290.000.000.000.680.00 0.000.430.000.000.000.000.57 Qwen-SEA-LION-v4-8B-VL Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.670.330.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 0.001.000.000.000.000.000.00 Qwen3-VL-8B-Thinking Chinese English Filipino Indonesian Tamil Thai Vietnamese Reasoning Language Chinese English Filipino Indonesian Tamil Thai Vietnamese Question Language 0.780.220.000.000.000.000.00 0.001.000.000.000.000.000.00 0.000.560.440.000.000.000.00 0.000.070.000.930.000.000.00 0.000.080.000.000.920.000.00 0.000.070.000.000.000.930.00 0.000.060.000.000.000.000.94 Ours-Qwen3-VL-8B 0.0 0.2 0.4 0.6 0.8 1.0 Probability of Language Fallback Figure 11: Main language fallback matrices comparing base reasoning models and OSCD post-trained models across the HMMT25 benchmark and 7 target languages (ZH,EN,Fi,IN,TA,TH,VI). 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 6 × 10 1 Total Backpropagated Loss 0.00.20.40.60.81.0 Epoch 10 0 Gradient Norm 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 6 × 10 1 Cross-Entropy Metric 0.00.20.40.60.81.0 Epoch 10 2 10 1 Cosine Distance Metric (a) Ours-SmolLM3-3B 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 6 × 10 1 Total Backpropagated Loss 0.00.20.40.60.81.0 Epoch 10 0 Gradient Norm 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 6 × 10 1 Cross-Entropy Metric 0.00.20.40.60.81.0 Epoch 10 3 10 2 Cosine Distance Metric (b) Ours-Qwen3-4B 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 5 × 10 1 Total Backpropagated Loss 0.00.20.40.60.81.0 Epoch 10 0 Gradient Norm 0.00.20.40.60.81.0 Epoch 3 × 10 1 4 × 10 1 5 × 10 1 Cross-Entropy Metric 0.00.20.40.60.81.0 Epoch 10 2 Cosine Distance Metric (c) Ours-Qwen3-VL-8B Figure 12: Training dynamics for models (a) SmolLM3-3B, (b) Qwen3-4B-Thinking-2507, and (c) Qwen3-VL-8B-Thinking using the OSCD post-training framework. −60−40−2002040 −60 −40 −20 0 20 40 60 80 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−2002040 −60 −40 −20 0 20 40 60 80 Chinese English Filipino Indonesian Tamil Thai Vietnamese (a) SmolLM3-3B (left) and Ours-SmolLM3-3B (right) −60−40−2002040 −60 −40 −20 0 20 40 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−2002040 −60 −40 −20 0 20 40 Chinese English Filipino Indonesian Tamil Thai Vietnamese (b) Qwen3-4B-Thinking-2507 (left) and Ours-Qwen3-4B (right) −60−40−200204060 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−200204060 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (c) Qwen3-VL-8B-Thinking (left) and Ours-Qwen3-VL-8B (right) Figure 13: 2D t-SNE visualizations of the first layer (Layer 1) multilingual representations for (a) SmolLM3-3B, (b) Qwen3-4B-Thinking-2507, and (c) Qwen3-VL-8B-Thinking before (left) and after (right) post-training with OSCD. −80−60−40−20020406080 −80 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −80−60−40−20020406080 −80 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (a) SmolLM3-3B (left) and Ours-SmolLM3-3B (right) −60−40−200204060 −75 −50 −25 0 25 50 75 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−200204060 −75 −50 −25 0 25 50 75 Chinese English Filipino Indonesian Tamil Thai Vietnamese (b) Qwen3-4B-Thinking-2507 (left) and Ours-Qwen3-4B (right) −60−40−200204060 −80 −60 −40 −20 0 20 40 60 80 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−200204060 −80 −60 −40 −20 0 20 40 60 80 Chinese English Filipino Indonesian Tamil Thai Vietnamese (c) Qwen3-VL-8B-Thinking (left) and Ours-Qwen3-VL-8B (right) Figure 14: 2D t-SNE visualizations of the middle layer (Layer 13) multilingual representations for (a) SmolLM3-3B, (b) Qwen3-4B-Thinking-2507, and (c) Qwen3-VL-8B-Thinking before (left) and after (right) post-training with OSCD. −60−40−2002040 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−2002040 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (a) SmolLM3-3B (left) and Ours-SmolLM3-3B (right) −60−40−20020406080 −80 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−20020406080 −80 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (b) Qwen3-4B-Thinking-2507 (left) and Ours-Qwen3-4B (right) −60−40−200204060 −80 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−200204060 −80 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (c) Qwen3-VL-8B-Thinking (left) and Ours-Qwen3-VL-8B (right) Figure 15: 2D t-SNE visualizations of the middle layer (Layer 25) multilingual representations for (a) SmolLM3-3B, (b) Qwen3-4B-Thinking-2507, and (c) Qwen3-VL-8B-Thinking before (left) and after (right) post-training with OSCD. −60−40−200204060 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−200204060 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (a) SmolLM3-3B (left) and Ours-SmolLM3-3B (right) −40−200204060 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −40−200204060 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (b) Qwen3-4B-Thinking-2507 (left) and Ours-Qwen3-4B (right) −60−40−200204060 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese −60−40−200204060 −60 −40 −20 0 20 40 60 Chinese English Filipino Indonesian Tamil Thai Vietnamese (c) Qwen3-VL-8B-Thinking (left) and Ours-Qwen3-VL-8B (right) Figure 16: 2D t-SNE visualizations of the last layer (Layer 36) multilingual representations for (a) SmolLM3-3B, (b) Qwen3-4B-Thinking-2507, and (c) Qwen3-VL-8B-Thinking before (left) and after (right) post-training with OSCD.