Grounded question answering systems have a simple-sounding obligation: answer only when the context supports it. In multi-hop tasks, this is harder because partial evidence can make an unsupported answer look plausible. The challenge is not just deciding whether an answer is correct, but locating the exact moment when silence should give way to a response.

A research group addressed this with Evidence Sufficiency Boundary Training, a generation-native training framework that builds ordered evidence chains and directly supervises the transition from abstention to answering. The method combines level supervision, a boundary flip margin, post-boundary stability, and answer recall protection. The chains are built from HotpotQA, 2WikiMultiHopQA, and MuSiQue; the model is Qwen2.5-3B-Instruct adapted with LoRA.

Locating the boundary, not just classifying

The results show that the framework achieves the strongest boundary localization among the tested systems, with flip accuracy of 0.807 compared with 0.781 for a token-level abstention baseline. On external non-answerable sets, the overall unsupported-answer rate drops to 0.095 versus 0.101 for the same baseline, while retaining competitive raw QA F1.

The less visible point is that the training does not only teach when to answer. Boundary supervision also requires that, once evidence sufficiency is reached, the answer remains stable even when redundant context is added. In real RAG systems, documents arrive in a messy order: first an insufficient passage, then the full chain, then superfluous material. A model that changes its answer when the context becomes redundant is hard to manage and audit. Post-boundary stability therefore addresses a concrete failure mode of retrieval pipelines, not a benchmark case.

What changes for locally deployed models

For teams running LLMs in self-hosted mode, the result has an architectural consequence. Selective abstention does not have to rely only on external filters or text instructions, which can be bypassed or poorly calibrated. It can be baked into the model during fine-tuning, using a 3-billion-parameter model adapted with LoRA. This reduces pressure on local inference infrastructure and makes refusal behavior more directly controllable.

The framework signals a shift from accuracy-only benchmarks toward refusal calibration metrics. In domains where a formally safe but unsupported answer is riskier than abstention, a system that learns the evidence boundary has an operational advantage. Teams building assistants for regulated domains benefit from a model that knows when to stay silent without needing an external gate. Approaches that treat abstention as a problem separate from training, by contrast, remain dependent on arbitrary thresholds.

For those evaluating on-premise deployment, there are trade-offs between local control and integration complexity; AI-RADAR collects analytical frameworks at /llm-onpremise for those who want to explore them. The boundary between silence and answer is not a fixed line: it depends on context and the order in which evidence arrives. Training it explicitly changes model behavior more than a simple answer classifier does.