# Masking Frequent Tokens Sharpens Direct Preference Optimization

arXiv: https://arxiv.org/abs/2609.32445 · HTML: https://mlsanity.com/paper/2609.32445/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 28 of its 29 references that are in the map.
- First posted on arXiv 26 Sep 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [LLM preference optimization methods](https://mlsanity.com/alignment/llm-preference-optimization-methods/)

## Builds on (most cited first, 28 of 28)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2204.05862/) (2022)
- [Distributed Representations of Words and Phrases and their Compositionality](https://mlsanity.com/paper/1310.4546/) (2013)
- [Learning to summarize from human feedback](https://mlsanity.com/paper/2009.01325/) (2020)
- [Scaling Laws for Reward Model Overoptimization](https://mlsanity.com/paper/2210.10760/) (2022)
- [SimPO: Simple Preference Optimization with a Reference-Free Reward](https://mlsanity.com/paper/2405.14734/) (2024)
- [A General Theoretical Paradigm to Understand Learning from Human Preferences](https://mlsanity.com/paper/2310.12036/) (2023)
- [Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators](https://mlsanity.com/paper/2404.04475/) (2024)
- [LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion](https://mlsanity.com/paper/2306.02561/) (2023)
- [ORPO: Monolithic Preference Optimization without Reference Model](https://mlsanity.com/paper/2403.07691/) (2024)
- [UltraFeedback: Boosting Language Models with Scaled AI Feedback](https://mlsanity.com/paper/2310.01377/) (2023)
- [SLiC-HF: Sequence Likelihood Calibration with Human Feedback](https://mlsanity.com/paper/2305.10425/) (2023)
- [Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts](https://mlsanity.com/paper/2406.12845/) (2024)
- [Representation Degeneration Problem in Training Natural Language Generation Models](https://mlsanity.com/paper/1907.12009/) (2019)
- [Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive](https://mlsanity.com/paper/2402.13228/) (2024)
- [Self-Play Preference Optimization for Language Model Alignment](https://mlsanity.com/paper/2405.00675/) (2024)
- [Disentangling Length from Quality in Direct Preference Optimization](https://mlsanity.com/paper/2403.19159/) (2024)
- [Token-level Direct Preference Optimization](https://mlsanity.com/paper/2404.11999/) (2024)
- [Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization](https://mlsanity.com/paper/2410.08847/) (2024)
- [Anisotropy Is Inherent to Self-Attention in Transformers](https://mlsanity.com/paper/2401.12143/) (2024)
- [Selective Preference Optimization via Token-Level Reward Function Estimation](https://mlsanity.com/paper/2408.13518/) (2024)
- [AlphaPO: Reward Shape Matters for LLM Alignment](https://mlsanity.com/paper/2501.03884/) (2025)
- [A Common Pitfall of Margin-based Language Model Alignment: Gradient Entanglement](https://mlsanity.com/paper/2410.13828/) (2024)
- [ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Preference Optimization](https://mlsanity.com/paper/2506.08712/) (2025)
- [SparsePO: Controlling Preference Alignment of LLMs via Sparse Token Masks](https://mlsanity.com/paper/2410.05102/) (2024)
- [Your UnEmbedding Matrix is Secretly a Feature Lens for Text Embeddings](https://mlsanity.com/paper/2606.07502/) (2026)
- [Normalized Rewards for Preference Optimization](https://mlsanity.com/paper/2607.16240/) (2026)

## Abstract (arXiv)

Direct Preference Optimization (DPO) aligns language models by optimizing over sequence-level sums of token-wise implicit reward differences. However, we identify a pervasive pathology in this formulation: a disproportionately small subset of high-frequency token types dominates cumulative sequence scores while appearing symmetrically across both preferred and dispreferred responses. Specifically, under canonical Qwen tokenization on Anthropic HH-RLHF, merely 69 token types account for 55.1% of all response tokens and 85.9% of within-pair shared token mass, exhibiting substantially lower preference-side specificity than the remaining vocabulary. This symmetric ubiquity induces gradient entanglement and dilutes the discriminative preference signal propagated through the objective. To resolve this issue, we introduce Anisotropic DPO (textsfADPO) and its canonical realization, Frequency-Hard DPO. Using a fixed, label-agnostic vocabulary mask, our method zeroes the implicit reward contribution of high-frequency response tokens while assigning unit weight to informative positions, thereby suppressing gradient interference without modifying preference pairs, discarding context, or introducing learned parameters. Here, anisotropy designates non-uniform token-level objective weighting rather than representational geometry. Extensive empirical evaluations on AlpacaEval, MT-Bench, and Arena-Hard demonstrate that Frequency-Hard DPO consistently outperforms standard DPO across Qwen-2.5-7B-Instruct and Llama-3-8B-Instruct, establishing that selectively masking shared high-frequency tokens offers an effective, zero-overhead mechanism for robust preference alignment.