🧠 ReasonGuard v0.2 — Linear Probe at L55 / mid_think on Qwen3.6-27B

v0.2 update (2026-04-29) — multi-bench training thesis FALSIFIED.

v0.2 trained on combined GSM8K + StrategyQA + MATH (455 samples, 45.8% halu rate) — same methodology that gave FabricationGuard cross-task AUROC 0.882. Within-bench improved on GSM8K (0.888 → 0.908). Cross-domain transfer still fails: StrategyQA 0.612, MATH 0.500 (chance). Position-of-faithfulness in the deep residual stream is more strongly domain-bound than multi-bench training compensates for. Honest negative-ish result registered as canonical case study of ProbeBench anti-Goodhart norms.

Reproducer: notebook 34 (multi-bench combined-train)

v0.2 results (multi-bench combined-train)

The pattern: AUROC degrades monotonically with task difficulty within the same probe + same model + same multi-bench training. Multi-bench training does NOT generalize reasoning probes the way it generalizes hallucination probes. The signal at L55/mid_think captures task-specific "I'm doing arithmetic correctly" representation, not abstract "I'm reasoning faithfully" — and that representation does not transfer to commonsense (StrategyQA) or harder math (MATH).

Why this matters: this is the first negative-ish result registered on ProbeBench v0.0.1 as live with full transparency. Both v0.1 (0.888 within / 0.605 cross) and v0.2 (0.908 / 0.612 / 0.500) numbers are publicly reported on the leaderboard. Anti-Goodhart norms ensure no spin: probes ship as they are, not as we wish they were.

---

v0.1 (original release, 2026-04-28)

Activation-probe reasoning-faithfulness detector for Qwen3.6-27B in thinking mode. AUROC 0.888 within math reasoning (GSM8K), 0.605 cross-domain to commonsense (StrategyQA), ~1ms scoring latency.

This is the OpenInterp ReasonGuard v0.1 probe — a linear LR probe at the mid-thinking position of layer 55 of Qwen3.6-27B. Trained on 300 GSM8K rollouts with thinking-mode generation. Detects when the model is producing math-reasoning that is likely to end in a wrong answer, read directly from the residual stream during the <think> block.

---

Honest scope — domain-bound

ReasonGuard works for math reasoning. It does not generalize to commonsense reasoning.

This is the headline of the v0.1 release. We trained on GSM8K alone, achieving 0.888 AUROC within-bench. When we tested cross-domain on StrategyQA (commonsense reasoning over multi-step strategy), AUROC dropped to 0.605 — barely above chance.

The L55/mid_think signal captures a math-reasoning-specific representation of "this trajectory is going wrong," not a generalized reasoning-faithfulness signal. Position-of-faithfulness in the deep residual stream is domain-bound, not abstracted.

This is a negative result on transfer that we ship honestly rather than hide. Per ProbeBench v0.0.1's anti-Goodhart norms, both numbers (0.888 within, 0.605 cross) are registered. Future v0.2 will train multi-bench (GSM8K + StrategyQA + MATH) and re-test transfer — same methodology that gave FabricationGuard cross-bench AUROC 0.882.

---

Full sweep across (layer, position)

We ran 12 combinations during training to find the best signal placement:

Novel finding — layer × position interaction:

• Shallow layer (L31): end_question wins. The model has just finished encoding the question.
• Deep layer (L55): mid_think wins. The reasoning-faithfulness signal emerges in the middle of the <think> block.

Position-of-faithfulness migrates with depth. This is a different finding from FabricationGuard, which used L31/end_question (shallow). The two probes capture different cognitive states at different stages.

---

Headline numbers

Detection AUROC

Threshold gates (notebook 32 verdict)

Final verdict: 1/3 PASS → 🟡 PARTIAL — narrow scope viable, multi-bench training needed for full ship.

---

Methodology lineage

ReasonGuard builds on three Anthropic interpretability lines + Apollo's deception-probe work:

---

How to use

Quickstart

import joblib, torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer, AutoModelForImageTextToText

# 1) Download probe artifact
probe_path = hf_hub_download(
    repo_id='caiovicentino1/ReasoningGuard-linearprobe-qwen36-27b',
    repo_type='dataset', filename='probe.joblib',
)
artifacts = joblib.load(probe_path)
probe   = artifacts['probe']
scaler  = artifacts['scaler']
layer   = int(str(artifacts['layer']).replace('L', ''))   # 55
position = artifacts.get('position', 'mid_think')

# 2) Load Qwen3.6-27B + register hook at L55
tok = AutoTokenizer.from_pretrained('Qwen/Qwen3.6-27B', trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    'Qwen/Qwen3.6-27B', dtype=torch.bfloat16,
    attn_implementation='sdpa', device_map='cuda', trust_remote_code=True,
).eval()

blocks = model.model.language_model.layers if hasattr(model.model, 'language_model') else model.model.layers
buf = {}
hook = blocks[layer].register_forward_hook(
    lambda _m, _i, out: buf.update(h=(out[0] if isinstance(out, tuple) else out).detach())
)

# 3) Score a math reasoning trace
@torch.no_grad()
def reasoning_score(question: str, max_think_tokens: int = 512) -> float:
    """Generate reasoning trace, capture mid-think residual, score with probe."""
    messages = [{'role': 'user', 'content': question}]
    text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    enc = tok(text, return_tensors='pt').to('cuda')
    # Generate up to mid_think_tokens to populate KV cache, then capture residual at midpoint
    out = model.generate(**enc, max_new_tokens=max_think_tokens // 2, do_sample=False)
    h_mid = buf['h'][0, -1].float().cpu().numpy()
    return float(probe.predict_proba(scaler.transform([h_mid]))[0, 1])

score = reasoning_score("If a train leaves Boston at 60mph and another from NYC at 80mph, when do they meet?")
print(f"Wrong-answer probability: {score:.3f}")

Files in this dataset

---

Comparison to FabricationGuard (the sibling probe)

Same author, same model, different cognitive target — different layer and position win:

The 4 things that differ between them are exactly the 4 things you'd expect to differ between fabrication detection and reasoning faithfulness:

• Different layer (shallow vs deep)
• Different position (end-of-prompt vs mid-thinking)
• Different training distribution (multi-bench vs single-bench → drives cross-bench transfer)
• Different cognitive states being captured

---

Reproduce

git clone https://github.com/OpenInterpretability/notebooks.git
# open 32_reasoningguard_proof_qwen36_27b_v2.ipynb in Colab
# Drive mount required for resumable rollouts (~6h on RTX PRO 6000 Blackwell)
# probe training: ~5 min after rollouts complete

Total cost: ~R$80 in Colab Pro+ credits (RTX 6000 ~R$4.50/hr × 6h19m + 5min training).

---

Background — why this exists

Built by Caio Vicentino at OpenInterpretability, April 28, 2026. Part of an open-source-first interpretability platform — the OSS productization layer above frontier-lab interpretability methodology, evaluated on 27B+ open-weights models.

Sister artifacts:

• caiovicentino1/FabricationGuard-linearprobe-qwen36-27b — fabrication probe (same model, different layer/position, multi-bench trained, generalizes cross-bench)
• caiovicentino1/qwen36-27b-sae-papergrade — paper-grade SAE on Qwen3.6-27B
• ProbeBench v0.0.1 leaderboard — categorical standardization for activation probes

Citation

@misc{vicentino2026reasonguard,
  title  = {ReasonGuard v0.1: Domain-bound Reasoning-Faithfulness Probe at L55/mid_think on Qwen3.6-27B},
  author = {Vicentino, Caio},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/caiovicentino1/ReasoningGuard-linearprobe-qwen36-27b},
  note   = {OpenInterpretability — open-weights mechanistic interpretability platform. Honest negative result on cross-domain transfer (math → commonsense): 0.888 within / 0.605 cross.}
}

License

Apache 2.0 — including patent grant. Free for commercial use. Attribution appreciated.

---

Last updated: 2026-04-28. Built solo from Brazil. Negative results shipped as first-class citizens. Star and fork at github.com/OpenInterpretability.