LLMs learn to abstain on demand with confidence targets
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Researchers have tackled a fundamental reliability problem in LLMs: models that confidently generate false answers. This work introduces explicit confidence targets, allowing users to specify acceptable accuracy thresholds at inference time while the model learns to abstain strategically rather than guess. Prior methods used fixed penalties that couldn't adapt across different confidence requirements. The breakthrough addresses a critical gap in deployment: production systems often need tunable trade-offs between coverage and correctness, yet existing RL approaches force a one-size-fits-all penalty structure. This technique enables models to meet user-specified reliability contracts dynamically, which matters for high-stakes applications where abstention is preferable to confident hallucination.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The key innovation isn't just learning when to abstain, but doing so while respecting user-specified accuracy thresholds at inference time. Prior RL methods baked in a single penalty coefficient during training, forcing a fixed abstention rate. This work decouples training from deployment requirements, letting the same model serve different reliability contracts without retraining.
This directly extends the abstention-focused work from the past week. The DeclineCraft benchmark (Oct 20) and the legal verification study (Oct 21) both identified that current models force answers when they should refuse. Those papers measured the problem; this one proposes a training mechanism to solve it. The connection to 'Reward Misspecification Degrades Chain-of-Thought Monitorability' (Oct 20) is also relevant: explicit confidence targets reduce the risk of reward hacking by making the objective transparent and verifiable at inference time, rather than relying on opaque penalty tuning.
If this approach maintains calibration across held-out confidence thresholds (e.g., 90% accuracy target vs. 95% target on the same model), that confirms the method generalizes beyond the training distribution. If instead performance degrades when users request thresholds far from training targets, the approach is limited to narrow operational windows and the fixed-penalty baseline remains more practical.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
New benchmark measures when LLMs correctly refuse to answer
DeclineCraft establishes a new benchmark for measuring when language models should refuse to answer rather than hallucinate. The work targets a critical failure mode in RAG systems: distinguishing between retrievable facts, learned parameters, and genuine knowledge gaps. Testing 27 models reveals that retrieval alone boosts abstention accuracy substantially, though iterative self-assessment shows promise for further…
MentionsLarge language models
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Reinforcement Learning Done Right for Explicit Confidence Targets”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.