Selected work

Representative Papers

In-Place Tokenizer Expansion for Pre-trained LLMs

Jul 16, 2026

This work addresses the inefficiency of fixed vocabularies in pretrained large language models when supporting new languages, which leads to tokenization redundancy, increased latency, and higher energy consumption—particularly hindering on-device multilingual capabilities. The authors propose an in-place tokenizer expansion method that preserves the original vocabulary while extending it through continued application of Byte Pair Encoding (BPE) merge rules on multilingual corpora. New subword embeddings are initialized as the mean of their constituent subword embeddings, followed by a two-stage training process involving embedding fine-tuning and full-model continued pretraining to recover performance. This approach achieves the first lossless, in-place tokenizer expansion for already-trained large models, ensuring new tokens decompose precisely into atomic units. Evaluated on the LFM2-8B-A1B model with a 128K-token vocabulary, the method reduces token counts for Hindi and Vietnamese by 2.4–2.6× (up to 4× for Thai), yielding an estimated 2.2–3.7× speedup in character-level decoding. The extended model and tokenizer are publicly released.

0 citationsRead paper

Multi-Agent Robotic Control with Onboard Vision-Language Models

Jul 08, 2026

This work addresses the limitations of existing vision-language models in robotic control—namely, poor interpretability, weak generalization, and reliance on cloud-based computation—by proposing a fully onboard multi-agent architecture. The system deploys lightweight vision-language models (3–20B parameters) alongside vision-language-action models on an AMD Ryzen AI mini PC, enabling autonomous mobile manipulation without external support through fine-tuning and hardware-in-the-loop simulation. A novel “Megamind” coordinating agent is introduced to mitigate the challenge of context retention in long-horizon tasks faced by smaller models. The architecture’s feasibility in terms of cost, performance, and real-world transferability is validated across five industrial warehouse tasks, and the associated simulation environment is open-sourced.

0 citationsRead paper

Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers

Jun 16, 2026

Deep recurrent models often struggle with signal propagation difficulties and fixed computational budgets in compositional reasoning tasks. This work proposes the Fixed-Point Reasoning Model (FPRM), built upon a recurrent Transformer architecture, which integrates pre-normalization and residual scaling to mitigate gradient degradation. Crucially, FPRM employs fixed-point convergence as an end-to-end halting mechanism, enabling the number of reasoning steps to adapt dynamically to task complexity. Evaluated on benchmarks including Sudoku, Maze solving, state tracking, and ARC-AGI, the model demonstrates substantial improvements in both reasoning performance and stability, confirming its effectiveness and strong generalization capability across diverse compositional reasoning domains.

0 citationsRead paper

Low-Pass Flow Matching

Jun 01, 2026

This work addresses a key limitation of conventional flow matching methods, which employ white noise as the source distribution despite its flat spectrum being inconsistent with the frequency-decaying nature of real-world data. To better align with the spectral characteristics of natural signals, the authors propose low-pass flow matching—a novel approach that introduces, for the first time in flow matching, a time-varying spectral bias mechanism. This is achieved through operator-modulated interpolation paths that dynamically steer the generative process toward low-frequency components during evolution. Combined with an adaptive ODE solver, the method enables highly efficient sampling. Experiments on unconditional image generation tasks, such as Galaxy10, demonstrate that the proposed approach significantly reduces sampling cost while maintaining or even improving sample quality.

0 citationsRead paper

Looped SSMs: Depth-Recurrence and Input Reshaping for Time Series Classification

May 15, 2026

This work explores two underexplored design dimensions of state space models (SSMs) for time series classification: deep recurrence and input reshaping. By recurrently applying the same SSM module across depth (i.e., deep recurrence) and integrating temporal concatenation or feature-time rechunking strategies at the input stage, the proposed approach substantially enhances model performance. The study introduces deep recurrence into SSMs for the first time, revealing its role as an effective inductive bias, and systematically demonstrates the consistent benefits of input reshaping across both low- and high-dimensional time series. Evaluated on six benchmarks, the method matches or surpasses existing large models with fewer parameters, achieving accuracy gains of 1–6%. The combined effect of both techniques is additive and consistently validated across multiple random seeds.

0 citationsRead paper

Recent publications

Latest Papers

In-Place Tokenizer Expansion for Pre-trained LLMs

Jul 16, 2026

This work addresses the inefficiency of fixed vocabularies in pretrained large language models when supporting new languages, which leads to tokenization redundancy, increased latency, and higher energy consumption—particularly hindering on-device multilingual capabilities. The authors propose an in-place tokenizer expansion method that preserves the original vocabulary while extending it through continued application of Byte Pair Encoding (BPE) merge rules on multilingual corpora. New subword embeddings are initialized as the mean of their constituent subword embeddings, followed by a two-stage training process involving embedding fine-tuning and full-model continued pretraining to recover performance. This approach achieves the first lossless, in-place tokenizer expansion for already-trained large models, ensuring new tokens decompose precisely into atomic units. Evaluated on the LFM2-8B-A1B model with a 128K-token vocabulary, the method reduces token counts for Hindi and Vietnamese by 2.4–2.6× (up to 4× for Thai), yielding an estimated 2.2–3.7× speedup in character-level decoding. The extended model and tokenizer are publicly released.

0 citationsRead paper

Multi-Agent Robotic Control with Onboard Vision-Language Models

Jul 08, 2026

This work addresses the limitations of existing vision-language models in robotic control—namely, poor interpretability, weak generalization, and reliance on cloud-based computation—by proposing a fully onboard multi-agent architecture. The system deploys lightweight vision-language models (3–20B parameters) alongside vision-language-action models on an AMD Ryzen AI mini PC, enabling autonomous mobile manipulation without external support through fine-tuning and hardware-in-the-loop simulation. A novel “Megamind” coordinating agent is introduced to mitigate the challenge of context retention in long-horizon tasks faced by smaller models. The architecture’s feasibility in terms of cost, performance, and real-world transferability is validated across five industrial warehouse tasks, and the associated simulation environment is open-sourced.

0 citationsRead paper

Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers

Jun 16, 2026

Deep recurrent models often struggle with signal propagation difficulties and fixed computational budgets in compositional reasoning tasks. This work proposes the Fixed-Point Reasoning Model (FPRM), built upon a recurrent Transformer architecture, which integrates pre-normalization and residual scaling to mitigate gradient degradation. Crucially, FPRM employs fixed-point convergence as an end-to-end halting mechanism, enabling the number of reasoning steps to adapt dynamically to task complexity. Evaluated on benchmarks including Sudoku, Maze solving, state tracking, and ARC-AGI, the model demonstrates substantial improvements in both reasoning performance and stability, confirming its effectiveness and strong generalization capability across diverse compositional reasoning domains.

0 citationsRead paper

Low-Pass Flow Matching

Jun 01, 2026

This work addresses a key limitation of conventional flow matching methods, which employ white noise as the source distribution despite its flat spectrum being inconsistent with the frequency-decaying nature of real-world data. To better align with the spectral characteristics of natural signals, the authors propose low-pass flow matching—a novel approach that introduces, for the first time in flow matching, a time-varying spectral bias mechanism. This is achieved through operator-modulated interpolation paths that dynamically steer the generative process toward low-frequency components during evolution. Combined with an adaptive ODE solver, the method enables highly efficient sampling. Experiments on unconditional image generation tasks, such as Galaxy10, demonstrate that the proposed approach significantly reduces sampling cost while maintaining or even improving sample quality.

0 citationsRead paper

Looped SSMs: Depth-Recurrence and Input Reshaping for Time Series Classification

May 15, 2026

This work explores two underexplored design dimensions of state space models (SSMs) for time series classification: deep recurrence and input reshaping. By recurrently applying the same SSM module across depth (i.e., deep recurrence) and integrating temporal concatenation or feature-time rechunking strategies at the input stage, the proposed approach substantially enhances model performance. The study introduces deep recurrence into SSMs for the first time, revealing its role as an effective inductive bias, and systematically demonstrates the consistent benefits of input reshaping across both low- and high-dimensional time series. Evaluated on six benchmarks, the method matches or surpasses existing large models with fewer parameters, achieving accuracy gains of 1–6%. The combined effect of both techniques is additive and consistently validated across multiple random seeds.

0 citationsRead paper