REVIEW 12 cited by
Controlled Decoding from Language Models
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
KL-regularized reinforcement learning (RL) is a popular alignment framework to control the language model responses towards high reward outcomes. We pose a tokenwise RL objective and propose a modular solver for it, called controlled decoding (CD). CD exerts control through a separate prefix scorer module, which is trained to learn a value function for the reward. The prefix scorer is used at inference time to control the generation from a frozen base model, provably sampling from a solution to the RL objective. We empirically demonstrate that CD is effective as a control mechanism on popular benchmarks. We also show that prefix scorers for multiple rewards may be combined at inference time, effectively solving a multi-objective RL problem with no additional training. We show that the benefits of applying CD transfer to an unseen base model with no further tuning as well. Finally, we show that CD can be applied in a blockwise decoding fashion at inference-time, essentially bridging the gap between the popular best-of-K strategy and tokenwise control through reinforcement learning. This makes CD a promising approach for alignment of language models.
Forward citations
Cited by 12 Pith papers
-
Selective Safety Steering via Value-Filtered Decoding
Value-filtered decoding steers LLM outputs for safety at decoding time using a value criterion with an explicit bound on false interventions controlled by one threshold hyperparameter.
-
Efficient Controllable Diffusion via Optimal Classifier Guidance
SLCD provably converges, under no-regret learning and a strong score-estimation assumption, to the KL-regularized optimal distribution using only supervised classification oracles.
-
XGrammar-2: Dynamic and Efficient Structured Generation Engine for Agentic LLMs
XGrammar-2 makes dynamic, tag-triggered structured generation fast by dispatching to cached substructure grammars with an Earley-based adaptive token-mask cache.
-
Simultaneous Multi-objective Alignment Across Verifiable and Non-verifiable Rewards
MAHALO aligns LLMs to multiple objectives in one model via per-objective action heads and PRM-guided decoding, improving math, value, and tutoring metrics jointly.
-
From Outcomes to Processes: Guiding PRM Learning from ORM for Inference-Time Alignment
SP-PRM trains process reward models from outcome reward models via score and preference consistency, improving reward-guided search on dialogue, summarization, and reasoning tasks.
-
Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time
SITAlign is an inference-time constrained decoder that maximizes a primary reward while enforcing thresholds on secondary rewards, and it reports better primary-reward win-tie rates than weighted-objective decoding.
-
Safe Inference-Time Alignment via Lagrangian Reward Augmentation
Dualizing Safe RLHF yields a one-dimensional convex calibration of λ that defines a drop-in safety-aware reward for Best-of-N and token-level inference-time decoders.
-
Reward Guided Decoding for Generative Recommendation
Reward-guided decoding reweights generative recommender probabilities by a learned business-value score, log P + R/β, and is deployed at Kuaishou.
-
Reward Shaping for (Inference-Time) Alignment: A Stackelberg Game Perspective
The optimal reward for KL-regularized LLM alignment is a threshold function—reward B above a prompt-dependent cutoff, 0 below—which can be estimated from base-model samples and integrated into decoding-time alignment.
-
QueryBandits for Hallucination Mitigation: Exploiting Semantic Features for No-Regret Rewriting
A contextual bandit that chooses among five query-rewrite strategies, conditioned on 17 linguistic features, reduces LLM hallucination on QA benchmarks and beats static prompting and no-rewrite baselines.
-
Advancing Decoding Strategies: Enhancements in Locally Typical Sampling for LLMs
ASTS extends locally typical sampling with semantic scoring and dynamic thresholds, reporting improved perplexity, MAUVE, and diversity on story and summarization tasks.
-
Safety Game: Inference-Time Alignment of Black-Box LLMs via Constrained Optimization
A linear-programming 'safety game' selects among LLM candidate answers to maximize helpfulness under a self-reported risk cap, improving safety-benchmark accuracy over reranking baselines in multiple-choice settings.
Discussion (0). Sign in to comment.