The ability not to answer, included in LLM quality

Kalai et al. argued that accuracy-based evaluations used in general benchmarks reward guessing over stating ignorance, and demonstrated that this tendency appears in models. [3] Evaluation rules not only measure capability but also shape desired outputs.

If only correct answers are rewarded and abstention is treated as zero points like incorrect answers, guessing with even a slight chance of being correct becomes advantageous. This is a consequence of scoring design, not the model's character. In high-impact tasks, the loss from incorrect answers may exceed the delay caused by abstention.

First, formally define three outputs: answer, abstain, and human confirmation. Do not exclude appropriate abstentions from failure logs, and distinguish between appropriate and excessive abstention. Which is preferable depends on the impact of the question and its verifiability.

Acceptance tests should include not only answerable questions but also those with insufficient information, future events, contradictory premises, and judgments outside organizational authority. Evaluations that do not explicitly include unanswerable examples cannot measure the ability to abstain.

• Answer: Sufficient evidence exists, meeting the lower bound of quality
• Confirm: Solvable with additional information or human judgment
• Abstain: Cannot be answered safely with current input and authority
• Stop: The impact of an incorrect answer is significant, preventing progression to subsequent processing

Farquhar et al. demonstrated a method for detecting hallucinations using semantic entropy, which measures uncertainty by grouping multiple answers to the same question into semantic clusters. [1] The value lies in examining whether the meaning of the answer is stable, rather than superficial linguistic differences.

However, this method primarily targets hallucinations arising from the model not knowing the answer and lacking consistency. If the model repeats the same misinformation with conviction, the meaning may be stable while the fact is incorrect. [1] Do not treat low uncertainty as proof of correctness.

Multiple generations and semantic classification incur additional costs. Do not apply them to all questions; instead, prioritize areas with high incorrect-answer impact, questions with weak evidence, and questions where answers diverge across models. Also evaluate errors in the classifier itself.

Uncertainty scores are insufficient if shown to users as raw numbers. Translate them into actions: what additional information is needed, where to request confirmation, and whether a provisional answer can be used.

AbstentionBench examined reasoning models across multiple datasets containing unanswerable questions and reported that simply scaling model size does not solve the abstention problem. [2] Extending reasoning and recognizing insufficient information to stop are distinct capabilities.

Unanswerable questions come in various types: no correct answer exists, cannot be determined from provided information, the question is ambiguous, the premise is false, or answering is prohibited by authority. Mixing types obscures improvement measures and responsible parties.

A model writing "I don't know" is not sufficient for passing. Verify what is missing, what can be added to enable an answer, and what alternative procedures exist. Since plausible reasons for insufficiency may be post-hoc, cross-reference with known missing conditions.

Also measure excessive abstention. Sending safe questions in large numbers to humans increases wait times and operator load, potentially causing users to seek workarounds. Aim for appropriate selection rather than low abstention rates.

• Insufficient Information: Necessary facts are missing from the input
• Bad Premise: Contains contradictory or false premises
• Out of Scope: Outside the model's, business's, or authority's purview
• Excessive Abstention: Sending answerable questions to humans unnecessarily

Excluding examples scored as uncertain narrows the answered range; whether retained accuracy improves depends on the score’s ability to distinguish errors. Therefore, plot coverage (the proportion of answered questions) and accuracy within that coverage for each threshold. Do not assume a single optimal threshold applies to all business contexts.

The cost of incorrect answers differs between guiding contract searches and deciding on payment suspensions. For the former, broadly outputting answers that prompt additional confirmation may be acceptable, whereas for the latter, accepting lower coverage and sending to humans may be better. Create curves by impact level.

Fix thresholds on the evaluation set, then verify them on a separate holdout set. Selecting metrics and thresholds on the same data inflates apparent performance. Also check if curves shift across languages, document types, and user groups.

Model updates may change score distributions. Do not automatically inherit old thresholds; monitor for sudden increases in answer volume, disappearance of abstentions, and changes in human submission rates. Treat threshold changes as requiring the same approval as model changes.

Incorrect answers entail rework, customer loss, and legal impact. Abstention entails wait times, opportunity loss, and user attrition. Human confirmation entails expert time and backlog. Without estimating these three losses for the same case categories, one cannot determine whether to increase abstention.

Separate significant tail losses from average costs. Irreversible incorrect answers, even if infrequent, should not be diluted by average amounts. Specify areas requiring stopping first, and compare costs outside those areas.

The quality of human confirmation is not fixed. If excessive abstentions concentrate on a single channel, confirmers may become fatigued, leading to superficial processing. Monitor submission volume, deadline breaches, and overturn rates, adjusting capacity in addition to improving the model.

Test whether users can avoid abstention by re-requesting answers. If changing the phrasing slightly yields dangerous answers, single-turn thresholds are insufficient for control. Evaluate using retry units that include conversation history.

Abstention messages must not be boilerplate for evading responsibility but should indicate the next safe action. Provide materials to check, information to add, contact details, and emergency routes to prevent users from hitting dead ends.

Define conditions for returning to an answer after evidence is added. Do not permanently automate based on one human confirmation; verify that premises and document versions remain the same. If including exception judgments in training data, verify rights and confidentiality separately.

Audits should report accuracy, appropriate abstention, excessive abstention, severity of incorrect answers, human confirmation time, and avoidance via retries. Do not consider displaying "confidence" as completion of control.

The ability not to answer is not merely a model function. It is an operational capability linking evaluation rules, business losses, confirmers' capacity, and guidance to users. Automate answers only within the range where this interconnected system maintains the lower bound.

Cases where human confirmation overturns the model's abstention and allows a safe answer are material for improving excessive abstention. Conversely, cases where humans stopped an answer are misses. Return both to the same ledger and update not only the model but also thresholds and input requirements.

Do not use overturn rates under the assumption that humans are always correct. Track disagreements among experts, evidence updates, and results revealed later to revise baseline labels. Do not mix evaluations before and after label changes; report by version.

Free-text abstention reasons cannot be aggregated alone. Classify them into insufficient information, contradiction, out of scope, safety constraints, uncertainty, and system errors, and verify consistency between reasons shown to users and internal judgments.

If abstentions concentrate among specific users or languages, do not attribute the cause solely to question quality. Investigate differences in model capability, input screens, reference data, and evaluation thresholds, and compare completion times for alternative routes.

In periodic reviews, judge not by the number of increased automatic answers but by whether significant incorrect answers were avoided and excessive abstention reduced. For areas where improvement is not possible, retain human-centric procedures as permanent designs.