Neural Information Processing Systems http://nips.cc/

Growing evidence suggests that large language models do not use their depth uniformly, yet we still lack a fine-grained understanding of their layer-wise prediction dynamics. In this paper, we trace the intermediate representations of several open-weight models during inference and reveal a structured and nuanced use of depth. Specifically, we propose a "Guess-then-Refine" framework that explains how LLMs internally structure their computations to make predictions. We first show that the top-ranked predictions in early LLM layers are composed primarily of high-frequency tokens, which act as statistical guesses proposed by the model early on due to the lack of appropriate contextual information. As contextual information develops deeper into the model, these initial guesses get refined into contextually appropriate tokens. Even high-frequency token predictions from early layers get refined > 70% of the time, indicating that correct token prediction is not "one-and-done". We then go beyond frequency-based prediction to examine the dynamic usage of layer depth across three case studies. Together, our results provide a detailed view of depth usage in LLMs, shedding light on the layer-by-layer computations that underlie successful predictions and providing insights for future works to improve computational efficiency in transformer-based models. Despite the remarkable performance of large language models (LLMs), their internal computations remain poorly understood. One critical question is: how do LLMs internally structure their computations during inference and use their depth layer-by-layer to arrive at predictions? Are specific token predictions always computed at the last layer or does the model settle on predictable tokens early on and simply propagate these predictions? These questions have implications both for interpreting the internal computations of these models and for building more efficient LLM that can use their compute dynamically.

In the framework specified in section 2.2, we use a first-order Markov chain with two states as our Figure S1: For four different cells, the posterior distribution function is computed and depicted. Here, we concentrate only on trials within original environments, where we know the correct environment and hence can assess how well is the decoding. In this approach, instead of using a state-space structure, we use the likelihoods given by Eq. (1) of Each plot shows a histogram of the average probability (over time) of correctly decoding the trials within unambiguous environments. Fig. S3 shows the decoded environment for a few sample trials based on the neural activity of the whole population. In some trials (e. g. trials 65 & 25) we observe few fluctuations, while in other In Eq. 6, we use a history dependent, gamma-distributed generalized linear model with identity link.

We thank all reviewers for their careful reviews and many positive comments. We feel that the typos and minor issues are easily addressable and will be corrected. We will incorporate this analysis into a revision of the paper. We thank R1 for bringing this highly related work to our attention. That work focuses on environments for which mice have previously developed spatial maps.

It turns out that Scry is a "social forecasting platform." Users join for free and can enter their personal estimates of the probabilities that certain events will happen, with Scry calculating the average probability. For example, one question is, "Will Apple launch a commercial self-driving electric vehicle before the end of 2024?" As I write this, there are 18 responses, entered up to six months ago. Eight answers are 50-50 and two are 100% yes.

The two-sample hypothesis testing problem is studied for the challenging scenario of high dimensional data sets with small sample sizes. We show that the two-sample hypothesis testing problem can be posed as a one-class set classification problem. In the set classification problem the goal is to classify a set of data points that are assumed to have a common class. We prove that the average probability of error given a set is less than or equal to the Bayes error and decreases as a power of $n$ number of sample data points in the set. We use the positive definite Set Kernel for directly mapping sets of data to an associated Reproducing Kernel Hilbert Space, without the need to learn a probability distribution. We specifically solve the two-sample hypothesis testing problem using a one-class SVM in conjunction with the proposed Set Kernel. We compare the proposed method with the Maximum Mean Discrepancy, F-Test and T-Test methods on a number of challenging simulated high dimensional and small sample size data. We also perform two-sample hypothesis testing experiments on six cancer gene expression data sets and achieve zero type-I and type-II error results on all data sets.

Digital crowdsourcing (CS) is a modern approach to perform certain large projects using small contributions of a large crowd. In CS, a taskmaster typically breaks down the project into small batches of tasks and assigns them to so-called workers with imperfect skill levels. The crowdsourcer then collects and analyzes the results for inference and serving the purpose of the project. In this work, the CS problem, as a human-in-the-loop computation problem, is modeled and analyzed in an information theoretic rate-distortion framework. The purpose is to identify the ultimate fidelity that one can achieve by any form of query from the crowd and any decoding (inference) algorithm with a given budget. The results are established by a joint source channel (de)coding scheme, which represent the query scheme and inference, over parallel noisy channels, which model workers with imperfect skill levels. We also present and analyze a query scheme dubbed k-ary incidence coding and study optimized query pricing in this setting.