Hidden-state reliability
Verdict-access gap
A mechanistic hallucination project that began as a basin-trap hypothesis and ended with a stronger finding: hidden states encode unsupportedness far better than the final PASS/FAIL readout expresses.
Core hypothesis
LLMs may contain useful internal evidence about whether an answer is unsupported, but the final verdict channel underuses that evidence. The project tests whether this is measurable before generation and whether it can support safer abstention.
Steps taken
- Built a config-driven Qwen2.5-1.5B pipeline over HaluBench-style grounded verification: dataset preparation, hidden-state capture, geometry features, probe evaluation, verdict generation, and intervention tests.
- Falsified the original basin/slip/trap mechanism with near-chance basin metrics, then shifted to layer-wise evidence readout and matched false-pass/true-pass analysis.
- Tested detector + abstain, stricter reruns, PGLAC, ICAS v1/v2, DRAS, and domain-readout LoRA to separate detection from correction.
Results
Original basin-trap hypothesis was disconfirmed: mechanism_trap_pass_auc 0.4893, gen_basin_trap_pass_auc 0.4954, slip_effect 0.0.
Qwen2.5-1.5B HaluBench hidden-state probe reached AUPRC 0.9222 vs model readout AUPRC 0.4254 on predicted PASS cases.
Evidence signal appeared early at layer 3 MLP with delta +0.1254 AUPRC and p=2.22e-11; best detector was layer 17.
Detector + abstain reached 0.8403 answered accuracy at 45.6% coverage and 0.0728 false-pass rate; live selective accuracy improved to 0.6513 vs 0.4751 baseline.
Correction remains unproven: Heretic, PGLAC, ICAS v1/v2, DRAS, rerun-under-stricter-grounding, and domain-readout LoRA did not cleanly beat detector + abstain.
Caveat
This does not solve hallucination. The positive result is detection/access; the intervention story is mostly honest negative results and points toward attention-routing repair rather than simple activation steering.
Status
Public evidence packet exists; strongest candidate for a research writeup.