The widespread success of LLMs on NLP benchmarks has been accompanied by concerns that LLMs function primarily as stochastic parrots that reproduce texts similar to what they saw during pre-training, often erroneously. But what is the nature of their errors, and do these errors exhibit any regularities? In this work, we examine irrelevant context hallucinations, in which models integrate misleading contextual cues into their predictions. Through behavioral analysis, we show that these errors result from a structured yet flawed mechanism that we term _class-based (mis)generalization_, in which models combine abstract class cues with features extracted from the query or context to derive answers. Furthermore, mechanistic interpretability experiments on Llama-3, Mistral, and Pythia across 39 factual recall relation types reveal that this behavior is reflected in the model’s internal computations: (i) abstract class representations are constructed in lower layers before being refined into specific answers in higher layers, (ii) feature selection is governed by two competing circuits — one prioritizing direct query-based reasoning, the other incorporating contextual cues — whose relative influences determine the final output. Our findings provide a more nuanced perspective on the stochastic parrot argument: through form-based training, LLMs can exhibit generalization leveraging abstractions, albeit in unreliable ways based on contextual cues — what we term _stochastic chameleons_.

Long sequence modeling has gained broad interest as large language models (LLMs) continue to advance. Recent research has identified that a large portion of hidden states within the key-value caches of Transformer models can be discarded (also termed evicted) withoutaffecting the perplexity performance in generating long sequences. However, we show that these methods, despite preserving perplexity performance, often drop information that is important for solving downstream tasks, a problem which we call information neglect. To address this issue, we introduce Chunked Instruction-aware State Eviction (CItruS), a novel modeling technique that integrates the attention preferences useful for a downstream task into the eviction process of hidden states. In addition, we design a method for chunked sequence processing to further improve efficiency. Our training-free method exhibits superior performance on long sequence comprehension and retrieval tasks over several strong baselines under the same memory budget, while preserving language modeling perplexity. The code and data have been released at https://github.com/ybai-nlp/CItruS.

There is an increasing interest in developing text-based relational reasoning systems, which are capable of systematically reasoning about the relationships between entities mentioned in a text. However, there remains a substantial performance gap between NLP models for relational reasoning and models based on graph neural networks (GNNs), which have access to an underlying symbolic representation of the text. In this work, we investigate how the structured knowledge of a GNN can be distilled into various NLP models in order to improve their performance. We first pre-train a GNN on a reasoning task using structured inputs and then incorporate its knowledge into an NLP model (e.g., an LSTM) via knowledge distillation. To overcome the difficulty of cross-modal knowledge transfer, we also employ a contrastive learning based module to align the latent representations of NLP models and the GNN. We test our approach with two state-of-the-art NLP models on 13 different inductive reasoning datasets from the CLUTRR benchmark and obtain significant improvements.

We analyzed the outputs of multiple question answering (QA) models applied to the Stanford Question Answering Dataset (SQuAD) to identify the core challenges for QA systems on this data set. Through an iterative process, challenging aspects were hypothesized through qualitative analysis of the common error cases. A classifier was then constructed to predict whether SQuAD test examples were likely to be difficult for systems to answer based on features associated with the hypothesized aspects. The classifier’s performance was used to accept or reject each aspect as an indicator of difficulty. With this approach, we ensured that our hypotheses were systematically tested and not simply accepted based on our pre-existing biases. Our explanations are not accepted based on human evaluation of individual examples. This process also enabled us to identify the primary QA strategy learned by the models, i.e., systems determined the acceptable answer type for a question and then selected the acceptable answer span of that type containing the highest density of words present in the question within its local vicinity in the passage.