Towards Minimal Targeted Updates of Language Models with Targeted Negative Training

TL;DR

This work formalizes the notion of a minimal targeted update and proposes a method to achieve such updates using negative examples from a model's generations, resulting in updates that keep the new distribution close to the original, unlike existing losses for negative signal which push down probability but do not control what the updated distribution will be.

Abstract

Generative models of language exhibit impressive capabilities but still place non-negligible probability mass over undesirable outputs. In this work, we address the task of updating a model to avoid unwanted outputs while minimally changing model behavior otherwise, a challenge we refer to as a minimal targeted update. We first formalize the notion of a minimal targeted update and propose a method to achieve such updates using negative examples from a model's generations. Our proposed Targeted Negative Training (TNT) results in updates that keep the new distribution close to the original, unlike existing losses for negative signal which push down probability but do not control what the updated distribution will be. In experiments, we demonstrate that TNT yields a better trade-off between reducing unwanted behavior and maintaining model generation behavior than baselines, paving the way towards a modeling paradigm based on iterative training updates that constrain models from generating undesirable outputs while preserving their impressive capabilities.

Figures (11)

• Figure 1: Summary of tnt (): For negative tokens (i.e. those flagged as undesirable given the preceding tokens), optimizes for a distribution that matches the original, renormalized after the offending token probability is set to zero. For all other tokens, encourages the new distribution to match the original.
• Figure 2: Across nearly all values of reducing unwanted behavior, the suite of objectives is able to achieve a comparable or better trade-off than baseline methods ( + and + ) between reducing unwanted behavior and minimizing change relative to the original model's generations. On the y-axis, higher is better.
• Figure 3: Baseline methods introduce more obvious disfluencies (word repeats and random ?? tokens) than methods, especially as the rate of unwanted behavior is reduced to small amounts. For readability and a more direct comparison to \ref{['fig:auc_curves']}, only points that are located on the frontier of the similarity vs. reduction curves are plotted.
• Figure 4: Results on XSUM and PaLM 2 1B. methods yield a better trade-off between similarity vs. reduction than baselines up to a 50% reduction rate. methods struggle to reduce the hallucination rate past this point, while baseline methods do so but at the expense of increasing obvious disfluencies. For readability and a more direct comparison to \ref{['fig:auc_curves']}, only points that are located on the frontier of the similarity vs. reduction curves are plotted.
• Figure 5: Similarity vs. reduction results as dataset size varies (results are on PaLM 2 1b with $\alpha=1$). Numbers signify percentage of the original dataset used for training and validation, with test set for evaluation held constant. With less data, methods are generally less effective at reducing hallucination rate. However, stands out as method that is able to minimize the changes to the original model behavior even with less data (positive slope in the results), suggesting its practical efficacy for minimal targeted updates in low-data regimes. Black dot signifies the original model's metrics.