General LLM

Measuring and Narrowing the Compositionality Gap in Language Models (2023)

We measure how often models can correctly answer all sub-problems but not generate the overall solution, a ratio we call the compositionality gap. … In the GPT-3 family of models, as model size increases we show that the single-hop question answering performance improves faster than the multi-hop performance does, therefore the compositionality gap does not decrease. … while more powerful models memorize and recall more factual knowledge, they show no corresponding improvement in their ability to perform this kind of compositional reasoning.

We present a new method, self-ask, that further improves on chain of thought. In our method, the model explicitly asks itself (and answers) follow-up questions before answering the initial question.

we find that the compositionality gap remains at a roughly constant 40% between different model sizes and training techniques, with no apparent improvement from scale

as the perplexities assigned to correct sub-question answers decrease (i.e., the model becomes more confident about the correct answers), the probability of answering the compositional question correctly increases. For example, when the maximal perplexity assigned to the correct sub-question answer (i.e., assigned to the correct answer that the model is less confident about of the two sub-questions) is between 1.232 and 6.738, the model answers 42.6% of the compositional question correctly. However, when the maximal perplexity is between 1.000 and 1.002, the model answers 81.1% of the compositional question correctly. We observed a similar pattern when sorting sub-question pairs by average instead of worse perplexity. Being able to answer a 1-hop question correctly in a prompt containing similar 1-hop questions does not mean that the model has fully “learned” the given fact. Our results indicate that a model can compose facts at a much higher rate when it can recall these facts more confidently.

additionally reporting the perplexity assigned to the correct answer might be a better way to evaluate LMs on downstream question answering tasks as opposed to just reporting accuracy. Concurrent papers (Srivastava et al., 2022; Wei et al., 2022a) recently advocate for this, but from a different perspective. They show evidence that as model size scales, cross-entropy loss can indicate an improvement on downstream task performance even when actual task metrics (i.e., accuracy) do not improve.

Self-ask (depicted in Figure 2) requires a one- or few-shot prompt that demonstrates how to answer the questions. Our prompt starts with those examples, after which we append the inference-time question. We then insert the phrase “Are follow up questions needed here:” at the end of the prompt since we found that doing so slightly improves results.4 The model then outputs a response. In most cases it first outputs “Yes.”, meaning that follow-up questions are necessary. The LM then outputs the first follow-up question, answers it, and continues asking and answering follow-up questions until it decides it has sufficient information; at this point, it outputs “So the final answer is:” before providing the final answer

Self-ask improves over chain of thought by smaller margins on 2WikiMultiHopQA and Musique but by a large 11% (absolute) on Bamboogle. We hypothesize that the much more varied nature of Bamboogle, and the fact that most questions are not similar to those in the few-shot prompt, might make it harder for chain of thought to decompose the questions, whereas our self-ask model, which explicitly decomposes questions before answering them, deals much better with novel inference questions.

LLaMA: Open and Efficient Foundation Language Models (2023)

TBC

PaLM: Scaling Language Modeling with Pathways (2022)

TBC

GENERALIZATION THROUGH MEMORIZATION: NEAREST NEIGHBOR LANGUAGE MODELS (2020)

We introduce kNN-LM, an approach that extends a pre-trained LM by linearly interpolating its next word distribution with a k-nearest neighbors (kNN) model. The nearest neighbors are computed according to distance in the pre-trained embedding space and can be drawn from any text collection, including the original LM training data. This approach allows rare patterns to be memorized explicitly, rather than implicitly in model parameters.

This can be done with a single forward pass over a text collection (potentially including the original LM training set), where the resulting context-target pairs are stored in a key-value datastore that is queried during inference. knn-llm

To search over this large datastore, we use FAISS (Johnson et al., 2017), an open source library for fast nearest neighbor retrieval in high dimensional spaces.

A FAISS index is then created using 1M randomly sampled keys to learn 4096 cluster centroids. For efficiency, keys are quantized to 64-bytes. During inference, we retrieve k = 1024 neighbors, and the index looks up 32 cluster centroids while searching for the nearest neighbors.

retrieving nearest neighbors from the corpus outperforms training on it.

performance monotonically improves as more neighbors are returned, and suggests that even larger improvements may be possible with a higher value of k. Nonetheless, even a small number of neighbors (k = 8) is enough to achieve a new state of the art.

Interpolating the memorizing LM with the original LM improves validation perplexity by just 0.1 – compared to 1.9 from kNN-LM. This result suggests that although the Transformer is expressive enough to memorize all training examples, learning to do so does not result in context representations that generalize. In contrast, kNN-LM memorizes training data while improving generalization.

Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models (2022)

we propose test-time prompt tuning (TPT), a method that can learn adaptive prompts on the fly with a single test sample. For image classification, TPT optimizes the prompt by minimizing the entropy with confidence selection so that the model has consistent predictions across different augmented views of each test sample.

TBC

Towards Revealing the Mystery behind Chain of Thought: A Theoretical Perspective (NeurIPS 2023)

TBC

Random-Access Infinite Context Length for Transformers (NeurIPS 2023)

TBC

LIMA: Less Is More for Alignment (NeurIPS 2023)

TBC

Towards Optimizing the Costs of LLM Usage (2024)

we propose optimizing the usage costs of LLMs by estimating their output quality (without actually invoking the LLMs), and then solving an optimization routine for the LLM selection to either keep costs under a budget, or minimize the costs, in a quality and latency aware manner. Additionally, we propose several deterministic heuristics for reducing tokens in a quality aware manner, and study the related optimization problem of applying the heuristics optimizing the quality and cost trade-off.

For instance, we have empirically observed that there is a significant difference in the summarization capabilities of GPT-3.5-Turbo and Text-Davinci on documents containing data in certain formats, such as tables versus lists.

existing methods:

  • Model Selection and Cascade: sequentially querying for the next model in the cascade if the previous model’s performance was not satisfactory. Examples: cascade architecture, triage, API Selection
  • Prompt Length Reduction: compress instruction, or demonstrations, or reduce their number. Examples: sentence compression, dynamic ICL, ICL pruning, LLMLingua
  • Caching Based approaches: GPTCache

A key insight that we have empirically observed is that there might not be a clear hierarchy of the models in terms of their performance [13, 15]. In other words, the largest or most expensive or most popular model might or might not perform the best for a given task and context