TALAN: Task-Aligned Latent Adaptation Networks for Targeted Post-Training of Large Language Models

In three linesTALAN (Task-Aligned Latent Adaptation Networks) combines a low-rank adapter with a sequence-conditioned latent side path inserted into the transformer's residual stream. Tested on four Qwen3 backbones and four STEM/code benchmarks, TALAN improves LoRA (+1.41 points) and DoRA (+1.85 points) baselines with <1% additional parameters and 1.01-1.02x inference overhead.

Summary generated by Claude — human-verified