Research
Driving hallucination to zero was a knowledge problem, not a behaviour problem.
Sixty days on 8×H100, measuring whether an open LLM can stand beside a CNC machine and answer in under three seconds — with citations, and with a refusal when it does not know.
- Program
- NVIDIA Innovation Lab · 2026-05-12 → 2026-07-11 (60 days)
- Cluster
- 8×H100 SXM5 80GB NVLink
- Report
- TR-2026-01 · 2026-07-13
- Domain
- CNC / NC code · 7 controller dialects
01 — Summary
From a 4–10% closed-book ceiling to 89.9% grounded correctness in real time.
A model deployed beside a running machine owes a stricter contract than a general assistant: one unsupported feed, speed, alarm or canned-cycle answer can cause physical damage, so a false answer is strictly worse than a refusal. We therefore measured false-answer rate as a first-class metric alongside correctness.
89.9
72B-FP8 grounded correctness (of 100)
1.73s
72B-FP8 p99 latency (TP4, CUDA graph)
4.8%
72B false-answer rate
−81pt
8B hallucination cut (SFT 90.5% → RAFT 9.5%)
1.5%
Held-out hallucination (n=200 unseen)
220,813
RAFT corpus rows (v3, every row cited)
02 — Diagnosis
The finding is a diagnosis, not just a number.
The residual hallucination we chased for four weeks was a knowledge problem, not a behaviour problem — the fact simply was not in the weights.
That is why 457 closed-book LoRA adapters could not lift trap-set correctness past a 4–10 out of 100 parametric ceiling. No amount of behavioural training conjures knowledge that is not there.
Retrieval-Augmented Fine-Tuning trains the model the way the product serves: over a verified knowledge base that supplies documents at inference time, so the model learns to read → ground → cite, else abstain. The ceiling fell at every scale — 8B, 32B and 72B alike.
03 — Method
Read · ground · cite — else abstain.
01
Cited corpus
Controller manuals, alarm dictionaries and cutting data normalised into 220,813 rows, every row carrying its source.
02
RAFT training
Qwen-family 8B, 32B and 72B trained alongside retrieval so they cite when a document exists and refuse when it does not.
03
FP8 quantization
Offline per-channel FP8 (W8A8) put the 72B tier on a real-time envelope with no measurable quality loss.
04
Retrieval-confidence gate
In-weights refusal saturates near 60% regardless of no-answer training share, so a retrieval gate sits in front of generation and drives production hallucination toward zero.
04 — Results
The ceiling fell at every scale.
Same base model, same cost. The only change is whether knowledge is supplied at inference time.
| Configuration | Correct (of 100) | False answers | Note |
|---|---|---|---|
| Closed-book LoRA (457 adapters) | 4 – 10 | — | Parametric ceiling, never broken |
| 8B · SFT | 59.7 | 90.5% | Fluent but ungrounded |
| 8B · RAFT | 75.2 | 9.5% | Unchanged cost, −81pt hallucination |
| 32B · RAFT | — | 1.5% | Held-out n=200, 98.5% correct refusal |
| 72B · RAFT (FP8) | 88.4 – 89.9 | 4.8% | 100% source-grounded |
The held-out evaluation used an alarm distribution excluded entirely from training (n=200). False answers falling to 1.5% with 98.5% correct refusal is evidence of learned read → ground → cite, else abstain behaviour rather than memorisation.
05 — Serving
A 72B tier inside a real-time envelope.
Accuracy only counts on a shop floor if it arrives inside the time an operator can wait beside the machine. Offline per-channel FP8 (W8A8) placed the 72B tier there with no measurable quality loss.
p99 1.73 seconds
At TP4 with CUDA graph. The 32B RAFT tier reaches p99 1.1s on a 52-token response.
≈18 GiB per GPU
About a 13× serving-cost reduction — it fits on a single on-premise node.
Per-tensor dynamic FP8 collapses
The same FP8 budget applied per-tensor and dynamically produced degenerate output. Quantization scheme is design, not preference.
06 — Production
Refusal is guaranteed by the gate, not by the weights.
Refusal trained into weights saturated near 60% no matter how we tuned the no-answer share. So in production, retrieval confidence is judged before generation: when the supporting documents do not clear the bar, the model does not compose an answer — it says so. That is the contract DION AI keeps on the floor.
Test it on your own floor.
Bring the parts and the controllers you run today, and ask the same questions.
Download the technical report (PDF)U2DIA AI Research (Suyong Yun). Retrieval-Augmented Fine-Tuning Bypasses the Parametric Knowledge Ceiling: Benchmarking Zero-Hallucination Machining-Copilot LLMs on 8×H100. Technical Report TR-2026-01, 2026-07-13.