The paper investigates whether language models can be fine‑tuned to produce outputs that better match intended probability distributions, even when the valid outputs cannot be enumerated in advance. Two calibration fine‑tuning methods—soft‑target and hard‑target—are evaluated across 12 models from four families, showing substantial improvements in structured‑sampling fidelity on unseen distribution families and parameter settings. The gains also transfer to broader stochastic generation tasks, expanding first‑token support and improving NoveltyBench scores, though some downstream capabilities such as arithmetic reasoning may be slightly affected.
arXiv:2606. 09856v1 Announce Type: cross Abstract: Post-training Large Language Models (LLMs) for reasoning typically focuses on deductive tasks such as mathematics and coding where correctness is verifiable.
The paper introduces Test-Time Calibration Learning (TTCL), a label‑free framework that adapts both reasoning accuracy and verbalized confidence of large language models directly on unlabeled target‑task data. TTCL generates self‑supervision signals from multiple model responses, enabling calibration without ground‑truth labels and proving theoretically as a bounded surrogate for the ideal calibration objective. Experiments on mathematical reasoning and factual question answering show consistent improvements, with base models gaining an average 40.13% in accuracy and 70.80% in ECE reduction across eight benchmarks, and further gains under domain shift.
arXiv:2608. 07827v1 Announce Type: new Abstract: Confidence estimation for large language models (LLMs) aims to estimate the probability that a generated answer is correct, while calibration aligns these estimates with empirical accuracy.
The paper introduces Dual-Seed Comparison (DSC), a protocol that uses two independent LLM-generated seeds to reduce systematic bias in probabilistic sampling. DSC constructs a bit sequence from the character-level ordinal values of the seeds, normalizes it into a pseudo-uniform variate, and maps it to the target distribution via the inverse cumulative distribution function. Empirical results show DSC outperforms existing methods in 96% of evaluated settings and enhances distributional control in tasks like MCQ generation and attribute-constrained text-to-image prompting.
arXiv:2604. 01170v2 Announce Type: replace-cross Abstract: While test-time scaling has enabled large language models to solve highly difficult tasks, state-of-the-art results come at exorbitant compute costs.
The paper introduces Scalable Power Sampling, a training‑free method that sharpens a large language model’s generative distribution by approximating the autoregressive power distribution with a token‑level scaled low‑temperature distribution. This approach eliminates the need for iterative Markov chain Monte Carlo, achieving performance comparable to or better than one‑shot GRPO on math, QA, and code tasks while cutting inference latency by more than tenfold. The method relies solely on the base model’s outputs, requiring no external rewards or verifiers.
arXiv:2501. 11790v5 Announce Type: replace-cross Abstract: Recent studies have raised significant concerns regarding the reliability of current mathematics benchmarks, highlighting issues such as simplistic design and potential data contamination.
arXiv:2502. 04646v2 Announce Type: replace-cross Abstract: Weighted sampling -- sampling from a probability density function (PDF) proportional to the product of a base PDF and a weight function -- is a fundamental technique with wide-ranging applications in variance reduction, biased sampling, data augmentation, and more.
The paper introduces Drift-Constrained Optimization (DCO), a framework that treats behavioral drift during fine‑tuning of instruction models as a bounded constraint rather than an uncontrolled side effect. By defining a drift budget, the authors reformulate fine‑tuning as a direction‑selection problem, showing that choosing different update directions can qualitatively change outcomes. Experiments on Qwen3 models demonstrate that carefully selected directions improve scientific reasoning and multilingual translation while preserving reasoning capabilities and general performance.
arXiv:2608. 15448v1 Announce Type: cross Abstract: Large language models increasingly rely on sampling as a driver of their own improvement, making the fidelity of their learned distributions more critical than ever.
The paper introduces Grammar‑Aligned Decoding (GAD), addressing the issue that conventional grammar‑constrained decoding (GCD) can distort a large language model’s probability distribution, yielding grammatical but low‑likelihood outputs. GAD proposes an adaptive sampling method, Approximate Expected Futures (ASAp), which uses prior samples to over‑approximate future grammaticality, ensuring outputs remain both grammatical and faithful to the model’s conditional probabilities. Experiments on code generation and structured NLP tasks demonstrate that ASAp often produces higher‑likelihood outputs than existing GCD techniques while still enforcing the required grammatical constraints.
As the chain-of-thought reasoning capabilities of large language models improve, evaluating and calibrating their reasoning confidence is becoming increasingly important for quantifying the uncertaint...