Research map / Label-efficient and robust learning / Semi-supervised learning
Selective Generation for Controllable Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SGen, a supervised and semi-supervised selective generation method that controls a false discovery rate defined via textual entailment (FDR-E) with PAC guarantees, using conformal entailment sets for pseudo-labeling and neuro-selection functions.
Problem
Certified risk control (selective prediction) for generative language models is hampered by exact-match correctness metrics, which fail to account for multiple valid answers in open-ended generation, causing metric misalignment and hallucinations.
Approach
Measure correctness with textual entailment (FDR-E) instead of exact match; learn an entailment set via PAC conformal prediction to pseudo-label unlabeled data in a semi-supervised selective generator, and use multi-threshold neuro-selection functions over multiple scoring functions to improve selection efficiency.
Key results
- On GPT-3.5-Turbo with f_M1, SGen-Semi reaches efficiency 0.7334 and average efficiency is the best of the compared methods (0.7334 vs 0.6050 for heuristic PL).
- On Alpaca-7B, SGen-Semi achieves average efficiency 0.3173 versus 0.2061 for the best heuristic baseline, while keeping FDR-E at the target level in the reported setup.
- Fully supervised methods reach efficiency 0.7535 (GPT-3.5) and 0.2959 (Alpaca-7B); SGen-Semi matches this with only 75% of labeled data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use neuro-selection over candidate scoring functions instead of single-threshold selection with f_M1 | SGen-Semi on GPT-3.5-Turbo · Natural Questions (NQ) | efficiency: 0.2829 -> 0.7334 | Helped |
| Use neuro-selection over candidate scoring functions instead of single-threshold selection with f_M1 | SGen-Semi on Alpaca-7B · Natural Questions (NQ) | efficiency: 0.1580 -> 0.3173 | Helped |
| Use neuro-selection over candidate scoring functions instead of single-threshold selection with f_M1 | SGen-Semi on GPT-3.5-Turbo · Natural Questions (NQ) | FDR-E: 0.0609 -> 0.1589 | Hurt |
| Use neuro-selection over candidate scoring functions instead of single-threshold selection with f_M1 | SGen-Semi on Alpaca-7B · Natural Questions (NQ) | FDR-E: 0.0359 -> 0.0685 | Hurt |
| Use only 75% of labeled calibration data and add semi-supervised learning | SGen-Semi versus supervised methods · Natural Questions (NQ), GPT-3.5-Turbo and Alpaca-7B | efficiency: 0.7535 -> 0.7334; 0.2959 -> 0.3173 | Mixed |
| Increase the number of unlabeled samples | SGen-Semi and SGen-Semi-NoMS · Natural Questions (NQ) | efficiency | Helped |
| Use entailment labels rather than exact match to measure correctness | SGen-Sup versus SGen-EM · Natural Questions (NQ) | efficiency | Helped |
Surprising findings
- EM-based certified selection (SGen-EM) is overly conservative, rejecting correct answers that paraphrase the reference, e.g. 'Tom Felton' versus 'Thomas Andrew Felton'.
Limitations
- Guarantees require the i.i.d. assumption, which may be violated in practice.
- Empirical FDR-E is not tightly bounded below the target, especially for Alpaca-7B.
- Entailment labels are expensive; the semi-supervised method only partially mitigates this.
Experimental setup
Datasets: Natural Questions
Benchmarks: Natural Questions (entailment-labeled)
Metrics: FDR-E selection efficiency
Baselines: SGen-EM SGen-PL H-Semi SGen-PFL H-Semi SGen-Sup single-threshold SGen-Semi variants
Builds on: selective classification (certified threshold learning) PAC conformal prediction (binomial tail) self-consistency
Closest papers by meaning
- Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo (ICLR 2025 · similarity 0.86)
- TRACE Back from the Future: A Probabilistic Reasoning Approach to Controllable Language Generation (ICML 2025 · similarity 0.86)
- Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data (ICML 2025 · similarity 0.85)
- Conformal Language Modeling via Posterior Sampling (NeurIPS 2026 · similarity 0.85)
- Conformal Generative Modeling with Improved Sample Efficiency through Sequential Greedy Filtering (ICLR 2025 · similarity 0.85)
- Language Generation with Strictly Proper Scoring Rules (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.