Mamba language models process sequences recurrently, but recurrent processing does not imply that all tokens are retained equally. We study this gap using structured recall tasks that expose primacy, recency, and lost-in-the-middle behavior. Our results suggest that Mamba uses two functional memory regimes: a short-lived state that supports recency, and a sparse long-lived channel subspace that preserves selected early inputs. Targeted interventions on this long-memory subspace selectively degrade early recall and transfer to long-context QA and summarization, indicating that the same channels support real long-context use. We then ask why middle items are forgotten despite being processed sequentially. We find that the central failure is a short-to-long memory transition bottleneck: middle items are no longer recent enough for short memory, but are often not promoted into the sparse long-lived subspace. Repeated input structure strengthens this bottleneck by changing the input-dependent discretization dynamics. These findings provide a mechanistic account of Mamba’s U-shaped recall curve as the interaction between long memory, short memory, and a fragile transition between them.

We analyze the learned input-output behavior of GLU-based neurons in large language models (LLMs). We propose a simple analysis method: For each neuron, we compute the cosine similarities between its input (reading) and output (writing) weight vectors. In this scheme, a strong negative cosine similarity indicates the neuron *weakens* the direction it detects in the residual stream, so we call this a *weakening neuron*. This allows us to gain a number of novel insights. First, we show that nine different LLMs have similar patterns: weakening neurons appear mostly in late layers whereas their counterparts, *(conditional) strengthening* neurons, are frequent in early-middle layers. Second, we find that weakening neurons display surprising behavior: even though there are few, they activate often and have a large influence on model behavior. Third, weakening neurons have a strong effect on model output when gate values are negative – which is surprising since negative gate values are not expected to encode functionality.

We find that correct-to-incorrect sycophancy signals are most linearly accessible within multi-head attention activations. Motivated by the linear representation hypothesis, we train linear probes across the residual stream, multilayer perceptron (MLP), and attention layers to analyze where these signals emerge. Although separability appears in the residual stream and MLPs, steering using these probes is most effective in a sparse subset of middle-layer attention heads. Using TruthfulQA as the base dataset, we find that probes trained on it transfer effectively to other factual QA benchmarks. Furthermore, comparing our discovered direction to previously identified “truthful” directions reveals limited overlap, suggesting that factual accuracy, and deference resistance, arise from related but distinct mechanisms. Attention-pattern analysis further indicates that the influential heads attend disproportionately to expressions of user doubt, contributing to sycophantic shifts. Overall, these findings suggest that sycophancy can be mitigated through simple, targeted linear interventions that exploit the internal geometry of attention activations. Code will be released upon publication.

This paper investigates whether large language models (LLMs) utilize numerical attributes encoded in a low-dimensional subspace of theembedding space when answering questions involving numeric comparisons, e.g., Was Cristiano born before Messi? We first identified,using partial least squares regression, these subspaces, which effectively encode the numerical attributes associated with the entities in comparison prompts. Further, we demonstrate causality, by intervening in these subspaces to manipulate hidden states, thereby altering the LLM’s comparison outcomes. Experiments conducted on three different LLMs showed that our results hold across different numerical attributes, indicating that LLMs utilize the linearly encoded information for numerical reasoning.

We introduce the concept of the self-referencing causal cycle (abbreviated ReCall )—a mechanism that enables large language models (LLMs) to bypass the limitations of unidirectional causality, which underlies a phenomenon known as the reversal curse. When an LLM is prompted with sequential data, it often fails to recall preceding context. For example, when we ask an LLM to recall the line preceding “O say does that star-spangled banner yet wave” in the U.S. National Anthem, it often fails to correctly return “Gave proof through the night that our flag was still there”—this is due to the reversal curse. It occurs because language models such as ChatGPT and Llama generate text based on preceding tokens, requiring facts to be learned and reproduced in a consistent token order. While the reversal curse is often viewed as a limitation, we offer evidence of an alternative view: it is not always an obstacle in practice. We find that ReCall is driven by what we designate as cycle tokens—sequences that connect different parts of the training data, enabling recall of preceding tokens from succeeding ones. Through rigorous probabilistic formalization and controlled experiments, we demonstrate how the cycles they induce influence a model’s ability to reproduce information. To facilitate reproducibility, we provide our code and experimental details at https://anonymous.4open.science/r/remember-B0B8/.

In an era where information is quickly shared across many cultural and language contexts, the neutrality and integrity of news media are essential. Ensuring that media content remains unbiased and factual is crucial for maintaining public trust. With this in mind, we introduce SAFARI (CroSs-lingual BiAs and Factuality Detection in News MediA and News ARtIcles), a novel corpus of news media and articles for predicting political bias and the factuality of reporting in a multilingual and cross-lingual setup. To the best of our knowledge, this corpus is unprecedented in its collection and introduces a dataset for political bias and factuality for three tasks: (i) media-level, (ii) article-level, and (iii) joint modeling at the article-level. At the media and article levels, we evaluate the cross-lingual ability of the models; however, in joint modeling, we evaluate on English data. Our frameworks set a new benchmark in the cross-lingual evaluation of political bias and factuality. This is achieved through the use of various Multilingual Pre-trained Language Models (MPLMs) and Large Language Models (LLMs) coupled with ensemble learning methods.

We introduce our systems developed for two subtasks in the shared task “Wojood” on Arabic NER detection, part of ArabicNLP 2023. For Subtask 1, we employ the XLM-R model to predict Flat NER labels for given tokens using a single classifier capable of categorizing all labels. For Subtask 2, we use the XLM-R encoder by building 21 individual classifiers. Each classifier corresponds to a specific label and is designed to determine the presence of its respective label. In terms of performance, our systems achieved competitive micro-F1 scores of 0.83 for Subtask 1 and 0.76 for Subtask 2, according to the leaderboard scores.