Loom: Weaving Diagnostic Strands into Free-Text Consensus via Embedding-Space Reweighting
Loom turns multiple diagnostic hypotheses into one RCA answer with a single LLM synthesis call.
The paper describes a deployed framework that embeds open-form heuristic outputs, reweights them through an iterative centroid method, and uses those weights to ground a lightweight LLM response. On OpenRCA, it matches a state-of-the-art autonomous agent on Bank and Market-2, while trailing on Market-1 and Telecom. The authors report about 26x faster inference, or about 33x with an 8B synthesizer. They also note deployment lessons around latency trade-offs, failed redundancy detection, and deterministic consensus improving SME trust. ArXiv · AI/CL/LG's note
The paper describes a deployed framework that embeds open-form heuristic outputs, reweights them through an iterative centroid method, and uses those weights to ground a lightweight LLM response. On OpenRCA, it matches a state-of-the-art autonomous agent on Bank and Market-2, while trailing on Market-1 and Telecom. The authors report about 26x faster inference, or about 33x with an 8B synthesizer. They also note deployment lessons around latency trade-offs, failed redundancy detection, and deterministic consensus improving SME trust. ArXiv · AI/CL/LG's note
score 4