Chain-of-Self-Questioning (CoSQ) is a proposed prompt-level way to help an AI model decide whether it has enough information to answer—or should abstain rather than make an unsupported claim. In a 2026 benchmark study, its Grounded-CoSQ variant reduced wrong commitments compared with chain-of-thought prompting while still answering most questions. The result is promising, but it is evidence from a specific evaluation, not a guarantee against hallucinations in everyday use.
What Chain-of-Self-Questioning asks a model to do
CoSQ puts an explicit information check before an answer. Rather than treating a model’s apparent confidence as sufficient, the prompt asks it to assess what information the question requires and whether it has an adequate basis to commit. The response can then be an answer or an abstention.
As an Amazon Associate I earn from qualifying purchases.
That choice matters when the cost of an unsupported answer is higher than the cost of sending a question to a person or another review process. Abstaining is not the same as solving the question: it is a way to avoid presenting an uncertain claim as established.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →What the paper tested
Ali Şenol’s paper, “When Should LLMs Abstain? Chain-of-Self-Questioning for Selective Risk Control”, submitted to arXiv on September 15, 2026, describes CoSQ as a prompt-only framework. Its abstract reports tests across 17 conditions and 11 open-weight and hosted model families. The main evaluation used the 817-item TruthfulQA multiple-choice validation set.
The paper reports three variants: Grounded-CoSQ, Critical-CoSQ and Adaptive-CoSQ. The abstract’s headline comparison is Grounded-CoSQ at a threshold of τ=0.90 against chain-of-thought prompting, under what the authors call the final balanced-option protocol.
What the reported results mean
For that specific TruthfulQA condition, the paper reports that Grounded-CoSQ lowered the mean unconditional wrong-commitment rate from 13.1% with chain-of-thought prompting to 8.9%. That is a reported 32.1% relative reduction. Answered accuracy increased from 86.9% to 89.7%, while coverage was 87.6%—the system answered 87.6% of questions rather than committing on all of them.
These measures describe different parts of the trade-off:
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →- Wrong-commitment rate tracks incorrect answers among all questions, including those on which the system abstains.
- Answered accuracy measures correctness among the questions the system did answer.
- Coverage is the share of questions answered. Greater coverage means fewer abstentions, but it does not by itself mean answers are more reliable.
The abstract says the reported improvements held across all 11 evaluated models and every evaluated threshold. That is a claim about the study’s tested models and conditions, not about every AI system or deployment.
How the three variants compare
The abstract gives coverage figures for all three approaches, but not enough detail to rank them fully on accuracy or wrong-commitment risk across every operating point.
| Variant | Reported coverage | What can be concluded from the abstract |
|---|---|---|
| Grounded-CoSQ | 87.6% at τ=0.90 in the headline balanced-option condition (Ali Şenol, 2026) | The abstract reports its headline wrong-commitment and answered-accuracy comparison against chain-of-thought prompting. |
| Critical-CoSQ | 88.6% (Ali Şenol, 2026) | The abstract describes it as more reliable than the baseline; it does not give corresponding headline accuracy and wrong-commitment figures here. |
| Adaptive-CoSQ | 86.5% (Ali Şenol, 2026) | The abstract describes it as more reliable than the baseline; it does not give corresponding headline accuracy and wrong-commitment figures here. |
Coverage figures for Critical-CoSQ and Adaptive-CoSQ are neighboring reported operating points, not a complete like-for-like ranking against Grounded-CoSQ at τ=0.90.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What this does—and does not—establish
The abstract also names a Natural Questions short-answer evaluation as convergent evidence, but does not give its numeric results. The available abstract does not expose the exact prompt templates, full scoring procedure, uncertainty intervals or statistical tests. That limits what can be inferred from the reported figures.
In particular, the results do not establish that a model’s self-assessment is calibrated, that CoSQ prevents hallucinations generally, or that it will deliver the same trade-off in a live product. The evidence supports a narrower point: in the reported benchmark evaluation, prompting models to assess whether they had enough information was associated with fewer wrong commitments while retaining substantial answer coverage.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
