The problem is not that a hallucination detector loses a few percentage points. The problem is that the ranking flips. On two open 3-billion-parameter models, Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct, loaded in 4-bit NF4 quantization on a single Colab T4 GPU, detector rankings show a Spearman rho of -0.80 and a Kendall tau-b of -0.667: the method that looks best on one model can become the worst on the other. If the goal is to ship a detector with an on-device LLM, benchmarks built on 7B-70B research models offer no compass.
The reproducible audit compares four detectors across three paradigms: length-normalized token-margin calibration, a lexical n-gram baseline, and two sampling-based consistency methods, semantic entropy and SelfCheckGPT-NLI. The test uses closed-book TriviaQA. The small models that actually ship on devicesโPhi-3-mini, Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, and Gemma-2-2B-Instruct, usually at 4-bit precisionโsit outside the evidence base. The most uncomfortable result is not the low correlation: the cheap, simple lexical baseline is statistically indistinguishable from the sampling-based consistency methods on both models, with bootstrap 95 percent confidence intervals. On a T4 GPU, sampling methods must generate multiple completions for the same input; if they deliver no measurable advantage, that extra compute has no justification. For local inference on constrained hardware, the signal is direct: detector complexity does not automatically translate into reliability.
The labeling short-circuit
There is a second, methodological short-circuit. The standard labeler for TruthfulQA-Generation, which turns generation into multiple choice through token overlap, produces near-chance AUROC: 0.46-0.59 for all four detectors on both models. It looks like detector collapse, but it is not. The same detectors, with the same pipeline, reach 0.69-0.89 AUROC on TriviaQA. What fails is the labeling, not the detector. The distinction has concrete consequences: a team could discard a good detector because a badly built benchmark produces numbers that look like chance. Evaluation becomes a problem of instrumentation validity before it is a problem of model quality.
The group releases the harness, prompts, and per-item scores, and frames detector choice as a fairness-of-error-rates allocation problem: on resource-scarce devices, not all errors have the same cost. A detector that reduces average error can shift mistakes toward more damaging categories. With rankings flipping between models, every choice has to be redone on the real model and the specific workload. For teams evaluating self-hosted or edge deployments, AI-RADAR offers analytical frameworks at /llm-onpremise to navigate these trade-offs, but the structural message remains that general-purpose benchmarks are not enough: local, reproducible, open evaluation is needed. The source does not promise a universal solution, but it makes the problem checkableโand that may be the only result that really matters when a hallucination detector has to work inside a device that cannot afford to be wrong.
๐ฌ Comments (0)
๐ Log in or register to comment on articles.
No comments yet. Be the first to comment!