The Hallucination Tax of Reinforcement Finetuning

• URL: http://arxiv.org/abs/2505.13988v1
• Date: Tue, 20 May 2025 06:36:45 GMT
• Title: The Hallucination Tax of Reinforcement Finetuning
• Authors: Linxin Song, Taiwei Shi, Jieyu Zhao,
• Abstract summary: Reinforcement finetuning (RFT) has become a standard approach for enhancing the reasoning capabilities of large language models (LLMs)<n>We study a critical side effect of RFT, which we term the hallucination tax: a degradation in refusal behavior causing models to produce hallucinated answers confidently.<n>Our results show that standard RFT training could reduce model refusal rates by more than 80%, which significantly increases model's tendency to hallucinate.
• Score: 11.558277688244354
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reinforcement finetuning (RFT) has become a standard approach for enhancing the reasoning capabilities of large language models (LLMs). However, its impact on model trustworthiness remains underexplored. In this work, we identify and systematically study a critical side effect of RFT, which we term the hallucination tax: a degradation in refusal behavior causing models to produce hallucinated answers to unanswerable questions confidently. To investigate this, we introduce SUM (Synthetic Unanswerable Math), a high-quality dataset of unanswerable math problems designed to probe models' ability to recognize an unanswerable question by reasoning from the insufficient or ambiguous information. Our results show that standard RFT training could reduce model refusal rates by more than 80%, which significantly increases model's tendency to hallucinate. We further demonstrate that incorporating just 10% SUM during RFT substantially restores appropriate refusal behavior, with minimal accuracy trade-offs on solvable tasks. Crucially, this approach enables LLMs to leverage inference-time compute to reason about their own uncertainty and knowledge boundaries, improving generalization not only to out-of-domain math problems but also to factual question answering tasks.

Related papers

• SwS: Self-aware Weakness-driven Problem Synthesis in Reinforcement Learning for LLM Reasoning [95.28059121743831]
  Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for training large language models (LLMs) on complex reasoning tasks.<n>We introduce a Self-aware Weakness-driven problem Synthesis framework (SwS) that systematically identifies model deficiencies and leverages them for problem augmentation.<n>SwS enables robust generalization byempowering the model to self-identify and address its weaknesses in RL, yielding average performance gains of 10.0% and 7.7% on 7B and 32B models.
  arXiv  Detail & Related papers  (2025-06-10T17:02:00Z)
• RULE: Reinforcement UnLEarning Achieves Forget-Retain Pareto Optimality [24.299312059430704]
  Unlearning is the task of selectively removing specific information from a model without retraining from scratch or degrading overall utility.<n>Existing methods often rely on large-scale forget and retain datasets, and suffer from unnatural responses, poor generalization, or catastrophic utility loss.<n>We propose Reinforcement UnLearning (RULE), an efficient framework that formulates unlearning as a refusal boundary optimization problem.
  arXiv  Detail & Related papers  (2025-06-08T14:38:39Z)
• Balancing Truthfulness and Informativeness with Uncertainty-Aware Instruction Fine-Tuning [79.48839334040197]
  Instruction fine-tuning (IFT) can increase the informativeness of large language models (LLMs), but may reduce their truthfulness.<n>In this paper, we empirically demonstrate how unfamiliar knowledge in IFT datasets can negatively affect the truthfulness of LLMs.<n>We introduce two new IFT paradigms, $UNIT_cut$ and $UNIT_ref$, to address this issue.
  arXiv  Detail & Related papers  (2025-02-17T16:10:30Z)
• Model Surgery: Modulating LLM's Behavior Via Simple Parameter Editing [63.20133320524577]
  We show that editing a small subset of parameters can effectively modulate specific behaviors of large language models (LLMs)<n>Our approach achieves reductions of up to 90.0% in toxicity on the RealToxicityPrompts dataset and 49.2% on ToxiGen.
  arXiv  Detail & Related papers  (2024-07-11T17:52:03Z)
• LLM Uncertainty Quantification through Directional Entailment Graph and Claim Level Response Augmentation [5.255129053741665]
  Large language models (LLMs) have showcased superior capabilities in sophisticated tasks across various domains, stemming from basic question-answer (QA) This paper presents a novel way to evaluate the uncertainty that captures the directional instability, by constructing a directional graph from entailment probabilities. We also provide a way to incorporate the existing work's semantics uncertainty with our proposed layer.
  arXiv  Detail & Related papers  (2024-07-01T06:11:30Z)
• Low-rank finetuning for LLMs: A fairness perspective [54.13240282850982]
  Low-rank approximation techniques have become the de facto standard for fine-tuning Large Language Models. This paper investigates the effectiveness of these methods in capturing the shift of fine-tuning datasets from the initial pre-trained data distribution. We show that low-rank fine-tuning inadvertently preserves undesirable biases and toxic behaviors.
  arXiv  Detail & Related papers  (2024-05-28T20:43:53Z)
• R-Tuning: Instructing Large Language Models to Say `I Don't Know' [66.11375475253007]
  Large language models (LLMs) have revolutionized numerous domains with their impressive performance but still face their challenges. Previous instruction tuning methods force the model to complete a sentence no matter whether the model knows the knowledge or not. We present a new approach called Refusal-Aware Instruction Tuning (R-Tuning) Experimental results demonstrate R-Tuning effectively improves a model's ability to answer known questions and refrain from answering unknown questions.
  arXiv  Detail & Related papers  (2023-11-16T08:45:44Z)
• Improving the Reliability of Large Language Models by Leveraging Uncertainty-Aware In-Context Learning [76.98542249776257]
  Large-scale language models often face the challenge of "hallucination" We introduce an uncertainty-aware in-context learning framework to empower the model to enhance or reject its output in response to uncertainty.
  arXiv  Detail & Related papers  (2023-10-07T12:06:53Z)
• Advancing Counterfactual Inference through Nonlinear Quantile Regression [77.28323341329461]
  We propose a framework for efficient and effective counterfactual inference implemented with neural networks. The proposed approach enhances the capacity to generalize estimated counterfactual outcomes to unseen data. Empirical results conducted on multiple datasets offer compelling support for our theoretical assertions.
  arXiv  Detail & Related papers  (2023-06-09T08:30:51Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.