Selected work

Representative Papers

LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR

Jan 20, 2026

This work proposes a billion-parameter, end-to-end multilingual vision-language model that directly converts document images into well-structured, naturally ordered text while enabling precise localization of embedded images. Addressing the error-proneness and inefficiency of traditional multi-stage OCR pipelines in handling multilingual documents, the model introduces a novel curriculum-based bounding box localization strategy during pretraining. It further incorporates reinforcement learning with an IoU-based reward mechanism (RLVR) and enhances robustness through checkpoint averaging and task arithmetic fusion. Evaluated on OlmOCR-Bench, the model achieves state-of-the-art performance while being nine times smaller than the previous best model and offering significantly faster inference. The authors publicly release the model, training data, and a new evaluation benchmark, LightOnOCR-bbox-bench.

1 citationsRead paper

DenseOn with the LateOn: Fully Open Dense and Late-Interaction Models for Multilingual, Long-Context, and Code Search

Jul 29, 2026

This work addresses the poor reproducibility and limited multilingual transferability of existing retrieval models, which often rely on closed-source data. We propose a fully open-source, end-to-end training framework leveraging a reconstructed dataset of 665 million English contrastive pairs and 1.88 million supervised pairs to train both DenseOn (a single-vector dense retriever) and LateOn (a ColBERT-style late-interaction model). These models are extended to eight languages, marking the first public release of large-scale multilingual retrieval data and models. Experimental results show that LateOn significantly outperforms DenseOn on unseen languages, achieving average nDCG@10 scores of 56.20 and 57.22 respectively on BEIR—setting new state-of-the-art results at this scale. Furthermore, our analysis reveals that translate-train functions as a general multilingual generalization mechanism rather than merely a target-language expansion strategy. All code, data, and models are publicly released.

0 citationsRead paper

Rethinking the Multilingual Reasoning Gap with Layer Swap

May 26, 2026

This study addresses the significant performance degradation of multilingual large language models in non-English reasoning, where maintaining both target-language chain-of-thought and reasoning accuracy remains challenging. The authors construct a six-language long-reasoning dataset and train native-language reasoning and English-pivot models based on Qwen3-8B-Base. Through weight-space analysis, they identify that core reasoning capabilities are concentrated in intermediate layers. Leveraging this insight, they propose an innovative Layer Swap method that exchanges intermediate-layer parameters to enhance native-language reasoning. Experiments demonstrate that this approach reduces the average reasoning gap to 1.9–3.5% across five non-English languages, effectively closing the performance gap while preserving target-language chain-of-thought throughout. This work further reveals, for the first time, a language-agnostic reasoning core alongside language-specific peripheral layer structures.

0 citationsRead paper

Internalized Reasoning for Long-Context Visual Document Understanding

Mar 31, 2026

This work addresses the lack of effective internalized reasoning mechanisms in existing visual long-document understanding methods by introducing, for the first time, an approach that embeds reasoning capabilities directly into vision-language models. The authors construct a synthetic data pipeline to generate reasoning traces comprising page relevance scores, evidence extraction, and ranking. These traces are integrated into the model through supervised fine-tuning (SFT), augmented with <think> tags and <cot> control tokens, alongside a low-intensity model merging strategy. Evaluated on MMLongBenchDoc, the resulting Qwen3-VL-32B achieves a score of 58.3—surpassing the performance of Qwen3-VL-235B, which has over seven times more parameters—while a Mistral-based variant reduces reasoning output length by 12.4× and simultaneously improves performance by 3.8 points.

0 citationsRead paper

ColBERT-Zero: To Pre-train Or Not To Pre-train ColBERT models

Feb 18, 2026

This work investigates whether multi-vector retrieval models inherently require large-scale pretraining or can achieve competitive performance solely through knowledge distillation from strong single-vector models. To address this, we introduce ColBERT-Zero, a model pretrained end-to-end exclusively on publicly available data, and conduct a systematic analysis of the interplay among pretraining, supervised fine-tuning, and knowledge distillation. We demonstrate for the first time that large-scale multi-vector pretraining significantly outperforms distilled-only counterparts—even when the latter leverage stronger but proprietary data—and highlight the critical importance of aligning pretraining and fine-tuning configurations. Under comparable model scales, ColBERT-Zero establishes a new state of the art for multi-vector retrieval trained solely on public data, surpassing both GTE-ModernColBERT and its base encoder, GTE-ModernBERT.

0 citationsRead paper

Recent publications

Latest Papers

DenseOn with the LateOn: Fully Open Dense and Late-Interaction Models for Multilingual, Long-Context, and Code Search

Jul 29, 2026

This work addresses the poor reproducibility and limited multilingual transferability of existing retrieval models, which often rely on closed-source data. We propose a fully open-source, end-to-end training framework leveraging a reconstructed dataset of 665 million English contrastive pairs and 1.88 million supervised pairs to train both DenseOn (a single-vector dense retriever) and LateOn (a ColBERT-style late-interaction model). These models are extended to eight languages, marking the first public release of large-scale multilingual retrieval data and models. Experimental results show that LateOn significantly outperforms DenseOn on unseen languages, achieving average nDCG@10 scores of 56.20 and 57.22 respectively on BEIR—setting new state-of-the-art results at this scale. Furthermore, our analysis reveals that translate-train functions as a general multilingual generalization mechanism rather than merely a target-language expansion strategy. All code, data, and models are publicly released.

0 citationsRead paper

Rethinking the Multilingual Reasoning Gap with Layer Swap

May 26, 2026

This study addresses the significant performance degradation of multilingual large language models in non-English reasoning, where maintaining both target-language chain-of-thought and reasoning accuracy remains challenging. The authors construct a six-language long-reasoning dataset and train native-language reasoning and English-pivot models based on Qwen3-8B-Base. Through weight-space analysis, they identify that core reasoning capabilities are concentrated in intermediate layers. Leveraging this insight, they propose an innovative Layer Swap method that exchanges intermediate-layer parameters to enhance native-language reasoning. Experiments demonstrate that this approach reduces the average reasoning gap to 1.9–3.5% across five non-English languages, effectively closing the performance gap while preserving target-language chain-of-thought throughout. This work further reveals, for the first time, a language-agnostic reasoning core alongside language-specific peripheral layer structures.

0 citationsRead paper

Internalized Reasoning for Long-Context Visual Document Understanding

Mar 31, 2026

This work addresses the lack of effective internalized reasoning mechanisms in existing visual long-document understanding methods by introducing, for the first time, an approach that embeds reasoning capabilities directly into vision-language models. The authors construct a synthetic data pipeline to generate reasoning traces comprising page relevance scores, evidence extraction, and ranking. These traces are integrated into the model through supervised fine-tuning (SFT), augmented with <think> tags and <cot> control tokens, alongside a low-intensity model merging strategy. Evaluated on MMLongBenchDoc, the resulting Qwen3-VL-32B achieves a score of 58.3—surpassing the performance of Qwen3-VL-235B, which has over seven times more parameters—while a Mistral-based variant reduces reasoning output length by 12.4× and simultaneously improves performance by 3.8 points.

0 citationsRead paper

ColBERT-Zero: To Pre-train Or Not To Pre-train ColBERT models

Feb 18, 2026

This work investigates whether multi-vector retrieval models inherently require large-scale pretraining or can achieve competitive performance solely through knowledge distillation from strong single-vector models. To address this, we introduce ColBERT-Zero, a model pretrained end-to-end exclusively on publicly available data, and conduct a systematic analysis of the interplay among pretraining, supervised fine-tuning, and knowledge distillation. We demonstrate for the first time that large-scale multi-vector pretraining significantly outperforms distilled-only counterparts—even when the latter leverage stronger but proprietary data—and highlight the critical importance of aligning pretraining and fine-tuning configurations. Under comparable model scales, ColBERT-Zero establishes a new state of the art for multi-vector retrieval trained solely on public data, surpassing both GTE-ModernColBERT and its base encoder, GTE-ModernBERT.

0 citationsRead paper

How to Train Your Long-Context Visual Document Model

Feb 16, 2026

This work addresses the lack of reproducible training methodologies in existing long-context vision-language models, which hinders effective visual question answering over lengthy documents. The study systematically investigates efficient training strategies—combining continual pretraining, supervised fine-tuning, and preference optimization—for models at 24B and 32B parameter scales, enabling visual document understanding with context lengths up to 344K tokens. Key innovations include aligning training sequence lengths with evaluation settings for improved efficacy, introducing page-index embeddings to significantly boost performance, leveraging synthetic data for self-augmented training, and demonstrating for the first time that long-context visual training can enhance pure textual long-context capabilities. The approach achieves state-of-the-art results on MMLongBenchDoc and releases a refined, high-quality benchmark, MMLBD-C, to support future research.

0 citationsRead paper