CoSQ: Teaching LLMs to Ask Themselves Three Questions Before Answering
> Paper: *Chain-of-Self-Questioning: A Framework for LLM Abstention* > arXiv: 2609.17516
An Awkward Scenario
You ask a large language model: "Who won the 2024 Nobel Prize in Physics?" The model instantly gives you a name, brimming with confidence. But the name is wrong.
You rephrase: "Can you tell me who won the 2024 Nobel Prize in Physics? If you're not sure, just say you don't know." The model still gives you that same wrong name.
The problem isn't that the model doesn't know the answer—the information was clearly in its training data. The problem is that the model never asks itself "Do I actually know this?" before speaking.
It sounds like a joke, but this is one of the most common failure modes of current LLMs. The paper gives it a name: wrong-commitment. The model chooses to answer when it shouldn't, rather than saying "I don't know."
A team of Anthropic researchers has proposed a surprisingly simple solution in a new arXiv paper: make the model ask itself three questions before answering.
Three Questions, One Gate
The method is called CoSQ (Chain-of-Self-Questioning), and its core is a three-stage self-interrogation process:
Stage 1: Grounded-CoSQ (Information Inventory) Instead of answering directly, the model first compiles a checklist: "What information do I need to answer this question?" For example, asked about "Company X's Q3 2025 revenue," the model lists: it needs X's financial reports, the Q3 time range, and the definition of revenue.
Stage 2: Critical-CoSQ (Information Assessment) For each checklist item, the model asks: "Do I actually know this? What's my confidence?" If any item's confidence falls below a threshold τ, the entire question is flagged as "should not answer."
Stage 3: Adaptive-CoSQ (Adaptive Abstention) If the decision is to abstain, the model doesn't just say "I don't know." It gives a structured refusal: what information is missing, why guessing isn't possible, and where the user should look instead.
Sounds like a prompt engineering trick? The experimental results say otherwise.
Wrong Commitments Cut by 32%
The researchers tested 11 mainstream models on TruthfulQA-MC, a trap-question benchmark, ranging from Mistral 7B to GPT-5.5. Results:
- Wrong-commitment rate (HR) dropped from 13.1% to 8.9%—an absolute reduction of 4.2 percentage points, a 32.1% relative reduction
- Answered accuracy (AA) rose from 86.9% to 89.7%—the model becomes more cautious, yet answers more correctly
- Coverage of 87.6%—only 12.3% of questions were refused
More importantly, the effect holds across all 11 models. From 7B small models to GPT-5.5, CoSQ consistently reduces the wrong-commitment rate. This isn't a quirk of one model—it's a general cognitive habit.
Why Does It Work So Well?
The paper's deepest finding isn't the numbers, but this observation:
> "Abstention decisions are implemented by the framework and are not additional prompt text."
In other words: CoSQ's abstention decision isn't implemented via prompt text, but enforced at the framework layer. After the model generates its "here's what I need to know" checklist, the framework (not the model) evaluates confidence against the checklist, and the framework (not the model) decides whether to let the answer through.
This is a critical architectural choice. If you let the model itself decide "should I answer?", it will lean toward answering—because its training objective rewards answering. Moving the judgment out of the model's hands and into an external framework makes abstention reliable.
This reminds me of a phenomenon I observed while researching "judgment-gate decoupling": the model actually knows when it shouldn't answer, but the action gate never consults the judgment module. CoSQ's approach essentially forces the action gate to consult the judgment module—routing the judgment result to the action decision via the framework layer.
The Difference from "Just Say Don't Know" Prompts
A natural question: wouldn't it suffice to write "say you don't know if unsure" directly in the prompt?
The researchers tested that too. Result: direct prompting had almost zero effect. The reason is interesting—the model treats "if you're not sure, say you don't know" as a polite courtesy rather than a real decision rule. It will say "I don't know" on easy questions (looking humble), but keeps grinding out answers on hard ones (because training rewards answering).
The key to CoSQ: it reframes the metacognitive question "should I answer?" as the concrete question "do I possess the information needed to answer?" The former is a vague self-assessment; the latter is a verifiable information inventory.
This matches how human experts work. A doctor doesn't ask herself "can I diagnose this?" She asks "do I have the patient's imaging reports, blood tests, and medical history?" The former relies on confidence; the latter on a checklist. Checklists are more reliable than confidence.
A Hidden Finding: What the Threshold τ Means
One detail in the paper is easy to miss: the threshold τ is not fixed but tunable. At τ=0.50, the model abstains more (lower coverage) but achieves a lower wrong-commitment rate; at τ=0.90, the model answers more willingly (higher coverage) with a slightly higher wrong-commitment rate.
This is effectively a risk-coverage trade-off curve. In high-stakes scenarios (medical diagnosis), you can set τ=0.50 so the model abstains on 40% of questions, but the remaining 60% are almost all correct. In low-risk scenarios (casual chat), τ=0.90 lets the model answer 90% of questions, with a slightly higher error rate but better user experience.
The existence of this curve means: abstention is not a binary decision but an adjustable risk dial. This is CoSQ's most valuable design choice—it treats abstention not as a switch, but as a continuous parameter.
Limitations
The paper honestly lists several limitations:
1. Tested only on QA tasks: CoSQ was validated on TruthfulQA-MC and NQ-Short, but whether it works in open-ended dialogue, creative writing, or code generation is unknown. 2. Depends on framework-level implementation: CoSQ's abstention is not a native model capability but external framework enforcement. Without this framework in deployment, the effect disappears. 3. Unparseable rate: In a minority of cases (~0–15%), the model's generated checklist is malformed and the framework cannot parse it. This is more pronounced with smaller models.
My Take: An Engineering Fix for "Judgment-Gate Decoupling"
CoSQ strengthens my belief that LLMs' metacognitive abilities are not missing—they are architecturally isolated. The model internally "knows" what information it lacks, but that knowledge doesn't automatically propagate to the "should I answer?" decision gate.
CoSQ's solution is engineering-oriented: enforced routing at the framework layer. It's a concrete engineering realization of "judgment-gate decoupling"—instead of trying to fix the internal propagation problem, it builds a bridge on the outside.
But this also means: as long as the architecture stays the same, this problem will keep recurring. CoSQ is an excellent stopgap, but the fundamental fix may require natively coupling the judgment module and the action gate at the model architecture level.
Until that day comes, making the model ask itself three questions before speaking is not a bad idea.
---
Paper link: https://arxiv.org/abs/2609.17516 HTML version: https://arxiv.org/html/2609.17516v1