A no-finetune method aligns heterogeneous LLM embeddings for agent routing
Super_Designer7952 · reddit · 2026-07-26
A Reddit post describes a method for aligning heterogeneous LLM embedding spaces without fine-tuning.
- The author targets decentralized multi-agent systems that mix models such as Llama, Mistral, and Phi.
- The proposed approach uses a fixed set of text anchors, orthogonalizes them with Löwdin symmetric orthogonalization, and maps model outputs into a shared anchor space.
- It then standardizes the similarity profiles and uses cosine distance for routing tasks across mismatched models.
- The post claims this avoids weight mutation, handles anisotropic embeddings, and enables more selective agent routing.
More from coding & agent
- Hermes adds a built-in skill for installing the open-source Obliteratus tool — Teknium · 2026-07-26
- YC Startup School argues that Markdown has become executable code for AI systems — ycombinator · 2026-07-26
- OpenAI Codex is blamed for maxing out SSDs, and Teknium points users to Hermes Agent — Teknium · 2026-07-26
- Users are splitting coding tools into planner, workhorse, fast-iteration, and frontend roles — nijfranck · 2026-07-26
- ChatGPT has been running a prompt overnight to automate ML, PINNs and LLM fine-tuning — burhop · 2026-07-26
- CCAR-P is an architect exam, not an API quiz, with integration and design dominating — Mauve_Tess · 2026-07-26