MathOctopus Models: LLMs & Soft Robotics

• MathOctopus Models are a dual approach combining multilingual large language models for math reasoning with advanced computational models for cephalopod biomechanics.
• LLMs in this family leverage supervised fine-tuning with both parallel and cross-training strategies to enhance performance across multiple languages using formula-verified datasets.
• Neuromechanical and Cosserat-rod models simulate soft-body dynamics of octopus arms through PDEs and optimal control, informing robotics and biologically inspired design.

MathOctopus Models comprise a family of mathematical and computational approaches spanning (1) multilingual mathematical reasoning in LLMs, and (2) neuromechanical and control-theoretic modeling of octopus arms and tentacles. The term "MathOctopus" refers both to LLMs developed for cross-linguistic math reasoning (Chen et al., 2023), and to a distinctive lineage of mathematical models for soft, continuum, and musculoskeletal control in cephalopods, foundational for robotics, biomechanics, and optimal control (Wang et al., 2022, Chang et al., 2022, Cacace et al., 2018). These models share the theme of unifying sophisticated mathematical reasoning or control with high-dimensional, non-rigid, or linguistically complex systems.

1. Multilingual MathOctopus LLMs for Mathematical Reasoning

MathOctopus LLMs are multilingual mathematical reasoning models based on fine-tuning LLaMA2 backbones. Architecturally, MathOctopus-7B uses LLaMA-2-7B (32 layers, hidden size 4096, 32 attention heads), and MathOctopus-13B employs LLaMA-2-13B (40 layers, hidden size 5120, 40 attention heads). No architectural modifications are made; full-parameter fine-tuning is found superior to LoRA adapters. A distinctive xRFT (multilingual rejection-sampling fine-tuning) variant augments the data with multiple independently-sampled correct reasoning paths per problem, filtered for correctness and formulaic uniqueness, yielding on average 3.8–5.2 distinct paths depending on the language (Chen et al., 2023).

2. MGSM8KInstruct: Dataset Construction and Verification

The MGSM8KInstruct corpus is constructed by translating GSM8K's 7,473 English math word problems and annotated chain-of-thought (CoT) solutions into ten languages: English, Bengali, Chinese, French, German, Japanese, Russian, Spanish, Swahili, and Thai. Rigorous translation prompts enforce consistency in entity names, formula preservation (via "<<"...">>" markers), and numeric representation. Formula consistency is machine-verified by extracting and evaluating all inline mathematical expressions post-translation; translations failing to preserve numerical correctness after five attempts are discarded (Chen et al., 2023). After filtering, each language subset contains approximately 7,400 problems, with a total multilingual scale of 73.6k problems.

3. Training Paradigms and Multilingual SFT/XRFT

MathOctopus models are trained using supervised fine-tuning (SFT) with two strategies:

• Parallel-training: Problem and CoT answer in the same language; optimizes in-domain performance on MGSM-like tasks.
• Cross-training: English problem with target-language CoT solution; improves out-of-domain generalization (tested on MSVAMP) and also notably boosts monolingual (English) GSM8K accuracy (+8.4 pp gain over English-only SFT for 7B models).

Multilingual rejection-sampling fine-tuning (xRFT) further augments SFT datasets with unique, valid reasoning paths sampled at temperature $\tau=0.9$, then numerically and structurally filtered. Gains from xRFT are quantifiably modest (typically +1–2 pp in-domain, <1 pp out-of-domain), and excessive rejection sampling can degrade generalization (Chen et al., 2023).

4. Experimental Evaluation, Key Metrics, and Comparative Performance

MathOctopus models are evaluated using both zero-shot and few-shot (two CoT exemplars) prompting on test sets including MGSM (across all ten languages) and GSM8K/SVAMP/MSVAMP (English and translated).

• On MGSM, MathOctopus-13B cross-trained and xRFT variants reach the state-of-art among open-source LLMs, matching or slightly exceeding ChatGPT (e.g., 47.6% vs. 46.3% 2-shot English CoT baseline).
• On GSM8K (English), MathOctopus-7B (cross, multilingual) substantially outperforms its English-only counterpart (50.8% vs. 42.2%), and MathOctopus-13B (cross) reaches 56.6%, exceeding baseline LLaMA2-13B.
• Monolingual SFT on non-English languages is consistently outperformed by multilingual SFT in the same language.

Statistical significance or confidence intervals are not reported, and mixing more than ∼70k SFT examples per language or naively combining all language pairs does not yield further improvements and may degrade accuracy (Chen et al., 2023).

5. Neuromechanical and Cosserat-Rod Models of Octopus Arms

Complementing the LLMs, "MathOctopus Models" also refers to a class of mathematical models for soft-bodied octopus appendages. Principal frameworks include:

• Planar Neuromechanical Model (Wang et al., 2022): Coupled PDEs for electrical (cable equation with adaptation variable for parallel nerve cords) and mechanical (planar Cosserat rod) domains. Muscle activation is mapped from local membrane voltage via a saturating function, producing an active bending couple. Control laws for reaching (reference-tracking and sensory-feedback) are rigorously analyzed for local asymptotic stability with Lyapunov and singular perturbation arguments.
• 3D Muscular Cosserat-Rod Model (Chang et al., 2022): Full 3D rod model using $SE(3)$ configuration, stored elastic and muscle energy densities, Hill-type muscle actuation, and a two-level optimization architecture. Task-specific equilibrium is found via an optimization constrained by the Cosserat rod equilibrium, solved using Pontryagin's Maximum Principle and a forward-backward sweep algorithm, yielding converged spatial activation profiles.
• Inextensible String Tentacle Model with Optimal Control (Cacace et al., 2018): Fourth-order nonlinear PDE for $(q,\sigma,u)$ with pointwise curvature control, inextensibility constraint, and reachability optimal control formulated as stationarity with respect to both state and control. The stationary reachability problem is related to the Dubins car system in control theory. Evolutionary optimal control leverages an adjoint-based gradient descent, with explicit algorithmic steps and boundary treatment.

6. Practical Insights, Observations, and Recommendations

Key findings across the MathOctopus lineages include:

• In LLMs, supervised fine-tuning with high-quality, formula-verified, parallel multilingual corpora substantially enhances mathematical reasoning both in target languages and monolingually.
• Task format alignment (question/answer language) is critical for performance—parallel-training for parallel evaluation, cross-training for cross-lingual evaluation.
• Rejection-sampling data augmentation achieves only limited incremental improvements and may worsen generalization if over-applied.
• In neuromechanical models, cascaded fast-slow (neural-mechanical) dynamics are structurally stable under moderate gain conditions. High-fidelity muscle modeling in 3D enables robust energy-shaping for both reach and grasp strategies, while in planar models, feedback based purely on local sensory cues suffices to generate biologically-plausible bend-propagation.
• Computational complexity is addressed through symplectic integrators and efficient PDE or ODE solvers; orders are explicitly reported (per (Cacace et al., 2018), $O(N\cdot M)$ per time-step).

7. Impact, Scope, and Research Opportunities

MathOctopus Models set benchmarks for multilingual mathematical reasoning in LLMs, highlighting strategies for equitable model performance across languages and the importance of cross-lingual dataset curation (Chen et al., 2023). In computational biomechanics and robotics, the Cosserat-rod and neuromechanical ecosystems provide a rigorous foundation for soft manipulator design, control, and biological emulation (Wang et al., 2022, Chang et al., 2022, Cacace et al., 2018). The interplay between large-scale data-driven and mechanistic modeling in the MathOctopus corpus presents opportunities for hybrid methodologies, interpretable AI control of soft robots, and deeper investigation into the neural-grounded mechanics of cephalopod movement.