REVIEW 6 cited by
Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization
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
abstract
Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences. Although these methods are designed to teach a model to generate preferred responses more frequently relative to dispreferred responses, prior work has observed that the likelihood of preferred responses often decreases during training. The current work sheds light on the causes and implications of this counter-intuitive phenomenon, which we term likelihood displacement. We demonstrate that likelihood displacement can be catastrophic, shifting probability mass from preferred responses to responses with an opposite meaning. As a simple example, training a model to prefer $\texttt{No}$ over $\texttt{Never}$ can sharply increase the probability of $\texttt{Yes}$. Moreover, when aligning the model to refuse unsafe prompts, we show that such displacement can unintentionally lead to unalignment, by shifting probability mass from preferred refusal responses to harmful responses (e.g., reducing the refusal rate of Llama-3-8B-Instruct from 74.4% to 33.4%). We theoretically characterize that likelihood displacement is driven by preferences that induce similar embeddings, as measured by a centered hidden embedding similarity (CHES) score. Empirically, the CHES score enables identifying which training samples contribute most to likelihood displacement in a given dataset. Filtering out these samples effectively mitigated unintentional unalignment in our experiments. More broadly, our results highlight the importance of curating data with sufficiently distinct preferences, for which we believe the CHES score may prove valuable.
Forward citations
Cited by 6 Pith papers
-
The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment
Shared chat-template tokens piggyback narrow finetuning behaviors onto out-of-domain queries; regularizing their KV states (TReFT) reduces emergent misalignment and other off-topic generalization.
-
The Cancellation Hypothesis in Critic-Free RL: From Outcome Rewards to Token Credits
The cancellation hypothesis shows how rollout-level rewards produce token-level credit assignment in critic-free RL through cancellation of opposing signals on shared tokens, with empirical support and batching interv...
-
The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment
The Piggyback Hypothesis attributes emergent misalignment to chat-template tokens piggybacking finetuned behavior; Token-Regularized Finetuning (TReFT) mitigates it by regularizing prefix token representations.
-
Exploring Post-Training Alignment of Small Language Models for Biomedical Data-to-Text Generation: A Case Study of Medication Leaflet
In medication-leaflet generation, ORPO and GRPO beat SFT and a zero-shot GPT-5 baseline, but GRPO's cross-domain advantage is partly built into its reward functions.
-
Normalized Rewards for Preference Optimization
A regularization term that conserves the combined length-normalized probability of chosen and rejected responses reduces likelihood displacement in DPO/SimPO, improves AlpacaEval and benchmark outcomes, and acts prima...
-
Multimodal Alignment and Preference Optimization for Zero-Shot Conditional RNA Generation
Moirain models use multimodal SFT and DPO to generate novel RNA sequences with superior protein binding affinities in a zero-shot conditional setting.
Discussion (0). Sign in to comment.