training-free-generate-critique-revise-loop-frozen-backbone-self-reported-numeric-confidence-tag CLM-0650In a training-free generate-critique-revise loop on a frozen backbone, a self-reported numeric confidence tag is useless as a stopping signal because the model declares high confidence on essentially every first generation (every threshold from 0.6 to 0.9 halted the loop at step 1 on 100 BBH questions), whereas a textual CONFIRMED sentinel from the critique stops 82-88% of items at unchanged accuracy.
Capability: Saying it does not know
Observed on
Pure-reasoning benchmarks (BBH, GSM8K, MATH), frozen API backbones deepseek-v4-flash and Qwen2.5-72B, depth-3 loop, prompt-only confidence elicitation..
Sources
- Measured, single paper, one 100-question BBH sweep for the threshold result plus a 500-item sentinel confusion matrix; the authors themselves note ~7pp session-to-session variance in single-shot accuracy on the same backbone and questions, so the accuracy comparisons are only read within Wilson intervals. Confidence calibration tested only via prompted tags, not log-probabilities.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For a frozen LLM run in a generate-critique-revise loop with no external feedback, the accuracy gain is concentrated where single-shot reasoning is already unreliable: on clean multi-step reasoning (BBH, ~74% single-shot) depth makes no difference beyond the confidence interval, while on harder math (MATH, 26.2% single-shot) the same loop gains +14.2 points, with GSM8K in between (+4.2).Fixing its own mistakes · reviewed by AI
- On SelfAware, a set of unanswerable questions (no scientific consensus, imagination, philosophical, too many variables, subjective) paired with answerable ones, models' ability to flag the unanswerable ones rises with parameter count, with instruction tuning and with in-context examples, yet the best model tested (GPT-4, F1 about 75) stays well below the human reference of about 85, so self-knowledge is present but incomplete.Saying it does not know · reviewed by AI
- Instruction tuning that forces an answer to every question teaches models to fabricate when the question lies outside their parametric knowledge; splitting the tuning data by whether the base model already answers correctly and appending "I am unsure" to the rest (refusal-aware tuning) raises average precision on known-versus-unknown separation over vanilla tuning, transfers to out-of-domain datasets, and yields better calibrated confidence than post-hoc uncertainty scoring.Saying it does not know · reviewed by AI
- For a frozen-backbone multi-agent scientific discovery pipeline on HypoBench tasks, distilling reusable procedural skills separately per stage — from discriminator critique of a whole hypothesis batch for generation, and from execution outcomes compared to ground-truth results for testing — beats no skill, an LLM-written skill, and a hand-written skill, and removing each stage's feedback signal costs most of the gain.Improving its own scaffolding from failures · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafter linked technique "stop-self-critique-loop-confirmed-sentinel", which does not list this capability in addresses -- recorded, not asserted. Drafted confidence: medium. Falsifier as drafted: A backbone on which prompted numeric confidence falls below the threshold on a meaningful share of first generations, producing a selective working point whose cost-accuracy trade-off beats the sentinel.