Research map / Large language models / Language, safety and interpretability
Controlled LLM Decoding via Discrete Auto-regressive Biasing
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 Discrete Auto-regressive Biasing (DAB), a decoding-time controlled generation method that alternates biased autoregressive decoding with gradient-based discrete Langevin sampling over an auxiliary bias token sequence.
Problem
Energy-based controlled text decoding struggles to balance fluency and constraint satisfaction because gradient-based (Langevin) sampling happens in continuous space rather than the discrete token space, and it is computationally costly.
Approach
Define a joint distribution over the response Y and an auxiliary discrete bias sequence B (P(Y,B|X) ∝ P_LM(Y|X,B) exp(f(B|X))), sample B with a discrete Langevin proposal computed from gradients w.r.t. one-hot bias tokens (no backprop through autoregressive generation), and steer decoding by penalizing tokens by embedding distance to B.
Key results
- DAB achieves internal classifier positive-sentiment probability of .992 vs .903 for BOLT and the best external SST-2 score of .975 on sentiment control.
- On detoxification DAB lowers average max toxicity to .211 vs .264 for BOLT, with perplexity 25.6 vs 27.3.
- DAB runs at 23.2 tokens/s versus 9.5 for BOLT (over 2x faster decoding).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAB instead of BOLT | Controlled sentiment generation · Sentiment-controlled generation | Int. Clsf: .903 ± .006 -> .992 ± .001 | Helped |
| Use DAB instead of LM-Steer | Controlled sentiment generation · Sentiment-controlled generation | Ext. Clsf (Yelp): .900 ± .008 -> .894 ± .009 | Hurt |
| Use DAB instead of LM-Steer | Controlled sentiment generation · Sentiment-controlled generation | Ext. Clsf (SST-2): .948 ± .006 -> .975 ± .003 | Helped |
| Use DAB instead of BOLT | Controlled sentiment generation · Sentiment-controlled generation | CoLA: .874 ± .005 -> .860 ± .005 | Hurt |
| Use DAB instead of BOLT | Toxicity avoidance · Toxicity avoidance | Avg. Max Toxicity: .264 ± .006 -> .211 ± .006 | Helped |
| Use DAB instead of BOLT | Toxicity avoidance · Toxicity avoidance | Toxicity Pred. Prob.: 6.8% -> 6.8% | No clear effect |
| Use DAB instead of BOLT | Keyword-guided generation · Keyword-guided generation | BertScore: .8291 ± .0003 -> .8303 ± .0003 | Helped |
| Use DAB instead of MuCOLA | Keyword-guided generation · Keyword-guided generation | Success Rate: 100% -> 99.0% | Hurt |
| Replace continuous BOLT bias sampling with DAB discrete sampling | Controlled generation decoding · Sentiment-directed generation | Tokens per Second: 9.495 ± .095 -> 23.213 ± 0.304 | Helped |
| Add keyword-and-topic instruction to the BOLT prompt (ablation) | BOLT · Topic-guided generation | Success Rate: 99.1% -> 99.7% | Helped |
| Add keyword-and-topic instruction to the BOLT prompt (ablation) | BOLT · Topic-guided generation | BertScore: .8291 ± .0003 -> .8123 ± .0002 | Hurt |
| Increase the bias weight (ablation) | DAB · Sentiment-controlled generation | Constraint control | Helped |
Surprising findings
- Modifying BOLT's constraint function to match DAB's form did not benefit BOLT, suggesting log-softmax smooths gradient direction needed for continuous sampling.
- Prepending keywords to the prompt raised BOLT's success rate slightly but worsened its fluency and topicality.
Limitations
- Number of gradient computations grows with queries, which can be costly.
- Not evaluated with multiple external constraints or compositional generation.
Experimental setup
Datasets: Yelp polarity SST-2 Jigsaw toxic comment RealToxicityPrompts GPT-4o-generated reference sentences for keyword topics
Benchmarks: Yelp polarity SST-2 RealToxicityPrompts keyword-guided topic generation (7 topics x 4 keywords)
Metrics: sentiment classifier probability average max toxicity toxicity predicted probability BertScore precision success rate CoLA repeated 3-grams perplexity (GPT2-XL) tokens per second
Baselines: MuCOLA COLD BOLT LM-Steer
Builds on: BOLT COLD MuCOLA Discrete Langevin Proposal
Closest papers by meaning
- Principled Gradient-Based MCMC for Conditional Sampling of Text (ICML 2024 · similarity 0.87)
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.87)
- Multi-Branch Self-Drafting for LLM Inference Acceleration (AAAI 2025 · similarity 0.87)
- Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata (NeurIPS 2026 · similarity 0.87)
- Primal-Dual Guided Decoding for Constrained Discrete Diffusion (NeurIPS 2026 · similarity 0.87)
- Cactus: Accelerating Auto-Regressive Decoding with Constrained Acceptance Speculative Sampling (ICLR 2026 · similarity 0.87)
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.