BidirLM: From Text to Omnimodal Bidirectional Encoders by Adapting and Composing Causal LLMs
This work provides a scalable and flexible method for creating bidirectional encoders from causal LLMs, which is incremental as it builds on existing adaptation techniques.
The paper tackled the problem of transforming causal generative language models into bidirectional encoders by addressing limitations like lack of consensus on training objectives, catastrophic forgetting, and inflexibility in integrating specialized models, resulting in BidirLM encoders that outperform alternatives on text, vision, and audio benchmarks.
Transforming causal generative language models into bidirectional encoders offers a powerful alternative to BERT-style architectures. However, current approaches remain limited: they lack consensus on optimal training objectives, suffer from catastrophic forgetting at scale, and fail to flexibly integrate the vast ecosystem of specialized generative models. In this work, through systematic ablations on the Gemma3 and Qwen3 families, we identify the key factors driving successful adaptation, highlighting the critical role of an often-omitted prior masking phase. To scale this process without original pre-training data, we introduce a dual strategy combining linear weight merging with a lightweight multi-domain data mixture that mitigates catastrophic forgetting. Finally, we augment our encoders by merging them with specialized causal models, seamlessly transferring modality- and domain-specific capabilities. This open-source recipe, designed for any causal decoder LLM, yields BidirLM, a family of five encoders that outperform alternatives on text, vision, and audio representation benchmarks.