| # Prior art: LLMs as decision models / self-verification (scan 2026-10-06) |
|
|
| Scope: before building the "generator verifies itself" experiments, check what |
| is already established for (a) reading typed decisions out of general LLMs, |
| (b) self-verification vs an independent verifier, (c) whether model size |
| matters, and (d) whether readouts survive quantization. Confidence tags: |
| [verified] = primary abstract/repo read here; [single-source] = one write-up; |
| [inferred] = our synthesis. |
|
|
| ## 1. Typed decisions from general LLMs — settled, recent |
|
|
| - **[verified] LLM-as-Jev** (Li & Wagle, `arXiv:2610.02076`, 1 Oct 2026): |
| architecture-preserving readout of bracketed option identifiers from |
| next-token probabilities. Qwen3.5-4B **without training matches community |
| Jev-style models built on the same backbone**; fine-tuning gives targeted |
| benefits (weaker models, many-option intent routing) and diminishing returns |
| on strong backbones; KL anchors protect conversational generation. |
| - **[verified] AnyJev** (Nokia, `arXiv:2610.00831`, 30 Sep 2026): L0 |
| (rotation marginalisation + label-prior correction, zero labels) and L2 |
| (closed-form LDA/ridge head on a hidden state ~2/3 depth, 100-300 labels per |
| question). Qwen3-1.7B L2 0.730 ~ Jev's published 0.727; Qwen3-4B 0.786 > |
| fine-tuned Laya 0.768; zero-label L0 is weak at small scale (1.7B 0.494, |
| 8B 0.647, 32B 0.700). Heads do not transfer across questions; labels are |
| needed per question. |
|
|
| Consequence: the premise of the Clef-shrinking line (train a separate |
| decision model on a capable base) is superseded. The readout is the product. |
|
|
| ## 2. Self-verification vs independent verification — the live debate |
|
|
| - **[verified] Kadavath et al. 2022** (`arXiv:2207.05221`): P(True) |
| self-evaluation of own samples works, is well calibrated few-shot, and |
| improves with model size; models find their own samples more plausible. |
| This is the origin of "the ability is latent". |
| - **[verified] Panickssery et al. 2024** (`arXiv:2404.13076`): self-preference |
| and self-recognition bias are real and causally linked; evaluators favour |
| their own outputs. |
| - **[verified] Huang et al. 2023** (`arXiv:2310.01798`): prompted intrinsic |
| self-correction of reasoning fails, at times degrading. |
| - **[verified] Lu et al. 2025/26** (`arXiv:2512.02304`): across 37 models and |
| 9 benchmarks, **cross-family verification is more effective than |
| self-verification; benefits decrease as solver and verifier become more |
| similar**; reasoning post-training weakens self-improvement. This is the |
| strongest argument to keep an independent verifier in the cascade. |
| - **[single-source] Zhang et al. 2025** (`arXiv:2504.05419`): probes on |
| reasoning-model hidden states verify answers with high accuracy and |
| calibration, and encode future correctness. |
| - **[single-source] Gnosis** (`arXiv:2512.20578`): ~5M-parameter head on |
| frozen internal traces (hidden-state evolution + attention), backbones |
| 1.7B-20B, outperforms large external judges on accuracy and calibration; |
| generalises to partial generations. |
|
|
| Synthesis: *readout-based* self-evaluation (logits, probes) >> prompted |
| self-critique; independent verifiers still beat self-verification. The |
| "generator judges itself for free" hypothesis is the weakest variant. |
|
|
| ## 3. Does size matter? |
|
|
| - **[verified]** LLM-as-Jev: 4B is the practical knee; 0.6B needs fine-tuning. |
| - **[verified]** AnyJev: L2 lifts 1.7B to Jev-level; zero-label L0 needs size. |
| - **[verified]** Kadavath: P(True) improves with scale. |
| - **[verified]** Lu et al.: verifier *diversity* matters more than raw size for |
| selection gains. |
| - **[inferred]** For our verifier: a 4B-class base + readout is the expected |
| sweet spot; below that, fine-tuning or a task-specific head is required. |
|
|
| ## 4. Readouts under quantization — our former blocker, largely answered |
|
|
| - **[verified] Aiersilan 2026** (`arXiv:2606.02628`): on three 7-8B |
| instruction-tuned models in **4-bit NF4**, a single *linear* probe on a |
| mid-network layer reaches **0.904-1.000 AUROC** on held-out hallucination |
| benchmarks; MLP probes rarely beat linear by >0.01 AUROC; peak layers are |
| consistent (Qwen2.5-7B blocks 19-25 of 28). Sampling-based detectors were |
| far worse under the paired-label protocol. |
| - **[single-source] SAPLMA** (Azaria & Mitchell 2023): linear probes on |
| hidden states detect statement truthfulness; middle layers best. |
| - **[single-source] `arXiv:2510.09033`** (counterpoint): hidden states may |
| reflect knowledge recall rather than truthfulness; unknown-unknown |
| hallucinations overlap with factual outputs. |
| - **[single-source] `arXiv:2609.31796`**: probes are sensitive to inference |
| numerics/precision; fix instrumentation across comparisons. |
|
|
| Consequence for us: the earlier "ternary kills the decision signal" |
| conclusion was measured through *Clef's frozen final-layer head*. Mid-layer |
| linear readouts are documented to survive 4-bit quantization; **2-bit ternary |
| survival is the genuinely open question**, and our QAT/V2 artifacts are |
| ready-made subjects. |
|
|
| ## 5. What remains for this project |
|
|
| 1. **Self vs independent on our task**: cascade verifier with checker gold, |
| mixed math/code/QA, on the actual local generator. Prediction (Lu et al.): |
| independent wins; if it does, the "no separate model" claim fails for the |
| hallucination lever even though typed decisions need no dedicated model. |
| 2. **Ternary/2-bit readout survival** (novel): linear probes at mid-depth on |
| 2-bit ternary bodies, fitted on bf16 and refit on ternary. |
| 3. **Footprint fit** on this box (RAM/VRAM floor beside the generator) — |
| systems question, not in the literature. |
| 4. **Calibration/threshold** for the cascade's reject/escalate lever. |
|
|
| Test plan (frozen set first): see `PLAN.md` in this directory. |
|
|
