LaMR: Multi-Rubric Latent Reasoning for Coding Agent Context Pruning
A recent publication on arXiv (2605.15315) presents LaMR (Latent Multi-Rubric), a structured pruning framework designed to enhance context compression for coding agents powered by LLMs. Traditional pruners rely on a single-objective sequence labeler with a CRF transition prior, which has difficulty managing diverse retention patterns such as contiguous semantic spans and sparse structural support lines. LaMR breaks down code relevance into two clear dimensions: semantic evidence and dependency support, each represented by a specific CRF with tailored transition dynamics. A gating network utilizing a mixture of experts adjusts the weights of emissions based on the query, while a concluding CRF layer on the combined emissions determines the pruning outcome, tackling the limitations of single-objective modeling to minimize token waste from irrelevant repository files.
Key facts
- arXiv paper 2605.15315 introduces LaMR (Latent Multi-Rubric).
- LaMR is a structured pruning framework for coding agent context.
- Existing pruners use a single-objective sequence labeler with one CRF transition prior.
- Single-objective formulation creates a modeling bottleneck for heterogeneous retention patterns.
- LaMR decomposes code relevance into semantic evidence and dependency support.
- Each dimension is modeled by a dedicated CRF with specific transition dynamics.
- A mixture-of-experts gating network weights per-rubric emissions conditioned on the query.
- Final CRF layer on fused emission produces the pruning decision.
Entities
Institutions
- arXiv