SRCoT
An auditable closed-loop reasoning architecture for edge LLMs. Through explicit process control of retrieval, reflection, and correction, combined with four-layer storage with source labels and local BM25 retrieval, it makes the generation process traceable and auditable.
Timeline (last 90 days)
SRCoT (Self-Reflective Chain-of-Thought) is an architecture that makes the generation process of edge LLMs traceable and auditable through a 'retrieve—reflect—revise' closed-loop process.
The SRCoT architecture separates environment state, persistent knowledge, accumulated experience and short-term memory into four layers of storage units with provenance labels.
SRCoT uses BM25 retrieval to provide local context.
SRCoT's intermediate products (such as intermediate drafts, missing-information markers, follow-up queries and claim—evidence mappings) directly influence the next decision: continue retrieving, revise the draft, request clarification or abstain from answering.
The study examines the SRCoT architecture through a trace-based qualitative case study, in a fully synthetic emergency medical services scenario and an emergency operations simulation environment called Munich-Sim.
The SRCoT local prototype runs on an Apple M4 series laptop equipped with 128GB of memory.
The paper explicitly limits its scope: the case demonstrates report-level auditability and local execution capability, not comparative advantage, hallucination reduction, clinical effectiveness or operational safety.
The traces reveal failure modes that can still occur even with prompt constraints: source attribution errors, incomplete evidence coverage and unsafe assumptions.
SRCoT's main contribution is an auditable control procedure rather than a new model, and it is a research proof of concept requiring human review.
The paper acknowledges that the architecture's behavior across different models remains to be evaluated.
All Facts (10)
The paper acknowledges that the architecture's behavior across different models remains to be evaluated.
50%UnverifiedThe SRCoT architecture separates environment state, persistent knowledge, accumulated experience and short-term memory into four layers of storage units with provenance labels.
50%UnverifiedSRCoT uses BM25 retrieval to provide local context.
50%UnverifiedSRCoT's intermediate products (such as intermediate drafts, missing-information markers, follow-up queries and claim—evidence mappings) directly influence the next decision: continue retrieving, revise the draft, request clarification or abstain from answering.
50%UnverifiedSRCoT (Self-Reflective Chain-of-Thought) is an architecture that makes the generation process of edge LLMs traceable and auditable through a 'retrieve—reflect—revise' closed-loop process.
50%UnverifiedThe SRCoT local prototype runs on an Apple M4 series laptop equipped with 128GB of memory.
50%UnverifiedThe paper explicitly limits its scope: the case demonstrates report-level auditability and local execution capability, not comparative advantage, hallucination reduction, clinical effectiveness or operational safety.
50%UnverifiedThe traces reveal failure modes that can still occur even with prompt constraints: source attribution errors, incomplete evidence coverage and unsafe assumptions.
50%UnverifiedSRCoT's main contribution is an auditable control procedure rather than a new model, and it is a research proof of concept requiring human review.
50%UnverifiedThe study examines the SRCoT architecture through a trace-based qualitative case study, in a fully synthetic emergency medical services scenario and an emergency operations simulation environment called Munich-Sim.
50%