We propose a method to teach multiple large language models (LLM) to collaborate by interleaving their generations at the token level. We model the decision of which LLM generates the next token as a latent variable. By optimizing the marginal likelihood of a training set under our latent variable model, the base LLM automatically learns when to generate itself and when to call on one of the “assistant” language models to generate, all without direct supervision. Token-level collaboration during decoding allows for a fusion of each model’s expertise in a manner tailored to the specific task at hand. Our collaborative decoding is especially useful in cross-domain settings where a generalist base LLM learns to invoke domain ex- pert models. On instruction-following, domain- specific QA, and reasoning tasks, we show that the performance of the joint system exceeds that of the individual models. Through qualitative analysis of the learned latent decisions, we show models trained with our method exhibit several interesting collaboration patterns, e.g., template-filling.
more »
« less
Boosting Dialog Response Generation
Neural models have become one of the most important approaches to dialog response generation. However, they still tend to generate the most common and generic responses in
the corpus all the time. To address this problem, we designed an iterative training process
and ensemble method based on boosting. We combined our method with different training
and decoding paradigms as the base model, including mutual-information-based decoding
and reward-augmented maximum likelihood learning. Empirical results show that our approach can significantly improve the diversity and relevance of the responses generated by all base models, backed by objective measurements and human evaluation.
more »
« less
- Award ID(s):
- 1722897
- PAR ID:
- 10106807
- Date Published:
- Journal Name:
- Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics
- Page Range / eLocation ID:
- 38-43
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
null (Ed.)Abstract Background Hidden Markov models (HMM) are a powerful tool for analyzing biological sequences in a wide variety of applications, from profiling functional protein families to identifying functional domains. The standard method used for HMM training is either by maximum likelihood using counting when sequences are labelled or by expectation maximization, such as the Baum–Welch algorithm, when sequences are unlabelled. However, increasingly there are situations where sequences are just partially labelled. In this paper, we designed a new training method based on the Baum–Welch algorithm to train HMMs for situations in which only partial labeling is available for certain biological problems. Results Compared with a similar method previously reported that is designed for the purpose of active learning in text mining, our method achieves significant improvements in model training, as demonstrated by higher accuracy when the trained models are tested for decoding with both synthetic data and real data. Conclusions A novel training method is developed to improve the training of hidden Markov models by utilizing partial labelled data. The method will impact on detecting de novo motifs and signals in biological sequence data. In particular, the method will be deployed in active learning mode to the ongoing research in detecting plasmodesmata targeting signals and assess the performance with validations from wet-lab experiments.more » « less
-
null (Ed.)Nanopore genome sequencing is the key to enabling personalized medicine, global food security, and virus surveillance. The state-of-the-art base-callers adopt deep neural networks (DNNs) to translate electrical signals generated by nanopore sequencers to digital DNA symbols. A DNN-based base-caller consumes 44.5% of total execution time of a nanopore sequencing pipeline. However, it is difficult to quantize a base-caller and build a power-efficient processing-in-memory (PIM) to run the quantized base-caller. Although conventional network quantization techniques reduce the computing overhead of a base-caller by replacing floating-point multiply-accumulations by cheaper fixed-point operations, it significantly increases the number of systematic errors that cannot be corrected by read votes. The power density of prior nonvolatile memory (NVM)-based PIMs has already exceeded memory thermal tolerance even with active heat sinks, because their power efficiency is severely limited by analog-to-digital converters (ADC). Finally, Connectionist Temporal Classification (CTC) decoding and read voting cost 53.7% of total execution time in a quantized base-caller, and thus became its new bottleneck. In this paper, we propose a novel algorithm/architecture co-designed PIM, Helix, to power-efficiently and accurately accelerate nanopore base-calling. From algorithm perspective, we present systematic error aware training to minimize the number of systematic errors in a quantized base-caller. From architecture perspective, we propose a low-power SOT-MRAM-based ADC array to process analog-to-digital conversion operations and improve power efficiency of prior DNN PIMs. Moreover, we revised a traditional NVM-based dot-product engine to accelerate CTC decoding operations, and create a SOT-MRAM binary comparator array to process read voting. Compared to state-of-the-art PIMs, Helix improves base-calling throughput by 6x, throughput per Watt by 11.9x and per mm2 by 7.5x without degrading base-calling accuracy.more » « less
-
Automatic short answer grading is an important research di- rection in the exploration of how to use artificial intelligence (AI)-based tools to improve education. Current state-of-the- art approaches use neural language models to create vector- ized representations of students responses, followed by clas- sifers to predict the score. However, these approaches have several key limitations, including i) they use pre-trained lan- guage models that are not well-adapted to educational sub- ject domains and/or student-generated text and ii) they al- most always train one model per question, ignoring the link- age across question and result in a significant model storage problem due to the size of advanced language models. In this paper, we study the problem of automatic short answer grad- ing for students’ responses to math questions and propose a novel framework for this task. First, we use MathBERT, a variant of the popular language model BERT adapted to mathematical content, as our base model and fine-tune it on the downstream task of student response grading. Sec- ond, we use an in-context learning approach that provides scoring examples as input to the language model to provide additional context information and promote generalization to previously unseen questions. We evaluate our framework on a real-world dataset of student responses to open-ended math questions and show that our framework (often signif- icantly) outperform existing approaches, especially for new questions that are not seen during training.more » « less
-
Lawrence, Neil (Ed.)Autoregressive Large Language Models (LLMs) have achieved impressive performance in language tasks but face two significant bottlenecks: (1) quadratic complexity in the attention module as the number of tokens increases, and (2) limited efficiency due to the sequential processing nature of autoregressive LLMs during generation. While linear attention and speculative decoding offer potential solutions, their applicability and synergistic potential for enhancing autoregressive LLMs remain uncertain. We conduct the first comprehensive study on the efficacy of existing linear attention methods for autoregressive LLMs, integrating them with speculative decoding. We introduce an augmentation technique for linear attention that ensures compatibility with speculative decoding, enabling more efficient training and serving of LLMs. Extensive experiments and ablation studies involving seven existing linear attention models and five encoder/decoder-based LLMs consistently validate the effectiveness of our augmented linearized LLMs. Notably, our approach achieves up to a 6.67 reduction in perplexity on the LLaMA model and up to a 2× speedup during generation compared to prior linear attention methods. Codes and models are available at https://github.com/GATECH-EIC/Linearized-LLM.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
