Model pruning in transformer-based language models, traditionally seen as a means of computational savings, can enhance the model’s reasoning capabilities. In this work, we uncover the surprising phenomenon that the selective pruning of certain attention heads leads to improvements in reasoning performance, particularly on challenging tasks. Motivated by this observation, we propose SPRINT, a novel contrastive learning framework that dynamically selects the optimal head and layer to prune during inference. By aligning question embeddings with head embeddings, our approach identifies those pruned-head configurations that result in more accurate reasoning. Extensive experiments on the MATH dataset demonstrate that our method significantly outperforms traditional best-of-N and random head selection strategies on the MATH500 and GSM8K datasets.

Language models, while capable of generating remarkably coherent and seemingly accurate text, can occasionally produce undesirable content including harmful or toxic outputs. In this paper, we present a new two-stage approach to detect and mitigate undesirable content generations by rectifying activations. First, we train an ensemble of layerwise classifiers to detect undesirable content using activations by minimizing a smooth surrogate of the risk-aware score. Then, for detected undesirable contents, we propose layerwise distributional steering policies that transform the attention heads. These policies are computed through principled semidefinite programming aims to minimally perturb the attention distribution while probabilistically guaranteeing the effectiveness of the editions. Empirical evaluations across multiple language models and datasets show that our method outperforms baselines in reducing the generation of undesirable output.

Language models (LMs) can produce texts that appear accurate and coherent but contain untruthful or toxic content. Inference-time interventions that edit the hidden activations have shown promising results in steering the LMs towards desirable generations. Existing activation intervention methods often comprise an activation probe to detect undesirable generation, triggering the activation modification to steer subsequent generation. This paper proposes a probe-free intervention method FLORAIN for all attention heads in a specific activation layer. It eliminates the need to train classifiers for probing purposes. The intervention function is parametrized by a sample-wise nonlinear low-rank mapping, which is trained by minimizing the distance between the modified activations and their projection onto the manifold of desirable content. Under specific constructions of the manifold and projection distance, we show that the intervention strategy can be computed efficiently by solving a smooth optimization problem. The empirical results, benchmarked on multiple base models, demonstrate that FLORAIN consistently outperforms several baseline methods in enhancing model truthfulness and quality across generation and multiple-choice tasks. Our implementation can be found at https://github.com/nguyenngocbaocmt02/EFI.

Modern language models (LMs) have significantly advanced generative modeling in natural language processing (NLP). Despite their success, LMs often struggle with adaptation to new contexts in real-time applications. A promising approach to task adaptation is activation intervention, which steers the LMs’ generation process by identifying and manipulating the activations. However, existing interventions rely heavily on heuristic rules or require many prompt inputs to determine effective interventions. In this paper, we propose a layer-wise additive activation intervention framework that optimizes the intervention process, thereby enhancing sample efficiency. We evaluate our framework on various datasets, demonstrating improvements in the accuracy of pretrained LMs and competing intervention baselines.

This paper describes our Speech-to-Text (STT) system for French, which was developed as part of our efforts in the Quaero program for the 2013 evaluation. Our STT system consists of six subsystems which were created by combining multiple complementary sources of pronunciation modeling including graphemes with various feature front-ends based on deep neural networks and tonal features. Both speaker-independent and speaker adaptively trained versions of the systems were built. The resulting systems were then combined via confusion network combination and crossadaptation. Through progressive advances and system combination we reach a word error rate (WER) of 16.5% on the 2012 Quaero evaluation data.