Dissociating Direct Access from Inference in AI Introspection

Introspection is a foundational cognitive ability, but its mechanism is not well understood. Recent work has shown that AI models can introspect. We study their mechanism of introspection, first extensively replicating Lindsey et al. (2025)'s thought injection detection paradigm in large open-source models. We show that these models detect injected representations via two separable mechanisms: (i) probability-matching (inferring from perceived anomaly of the prompt) and (ii) direct access to internal states. The direct access mechanism is content-agnostic: models detect that an anomaly occurred but cannot reliably identify its semantic content. The two model classes we study confabulate injected concepts that are high-frequency and concrete (e.g., "apple'"); for them correct concept guesses typically require significantly more tokens. This content-agnostic introspective mechanism is consistent with leading theories in philosophy and psychology.

From sunblock to softblock: Analyzing the correlates of neology in published writing and on social media

Living languages are shaped by a host of conflicting internal and external evolutionary pressures. While some of these pressures are universal across languages and cultures, others differ depending on the social and conversational context: language use in newspapers is subject to very different constraints than language use on social media. Prior distributional semantic work on English word emergence (neology) identified two factors correlated with creation of new words by analyzing a corpus consisting primarily of historical published texts (Ryskina et al., 2020, arXiv:2001.07740). Extending this methodology to contextual embeddings in addition to static ones and applying it to a new corpus of Twitter posts, we show that the same findings hold for both domains, though the topic popularity growth factor may contribute less to neology on Twitter than in published writing. We hypothesize that this difference can be explained by the two domains favouring different neologism formation mechanisms.

LLMs and people both learn to form conventions -- just not with each other

Humans align to one another in conversation -- adopting shared conventions that ease communication. We test whether LLMs form the same kinds of conventions in a multimodal communication game. Both humans and LLMs display evidence of convention-formation (increasing the accuracy and consistency of their turns while decreasing their length) when communicating in same-type dyads (humans with humans, AI with AI). However, heterogenous human-AI pairs fail -- suggesting differences in communicative tendencies. In Experiment 2, we ask whether LLMs can be induced to behave more like human conversants, by prompting them to produce superficially humanlike behavior. While the length of their messages matches that of human pairs, accuracy and lexical overlap in human-LLM pairs continues to lag behind that of both human-human and AI-AI pairs. These results suggest that conversational alignment requires more than just the ability to mimic previous interactions, but also shared interpretative biases toward the meanings that are conveyed.

Mechanisms of Prompt-Induced Hallucination in Vision-Language Models

Large vision-language models (VLMs) are highly capable, yet often hallucinate by favoring textual prompts over visual evidence. We study this failure mode in a controlled object-counting setting, where the prompt overstates the number of objects in the image (e.g., asking a model to describe four waterlilies when only three are present). At low object counts, models often correct the overestimation, but as the number of objects increases, they increasingly conform to the prompt regardless of the discrepancy. Through mechanistic analysis of three VLMs, we identify a small set of attention heads whose ablation substantially reduces prompt-induced hallucinations (PIH) by at least 40% without additional training. Across models, PIH-heads mediate prompt copying in model-specific ways. We characterize these differences and show that PIH ablation increases correction toward visual evidence. Our findings offer insights into the internal mechanisms driving prompt-induced hallucinations, revealing model-specific differences in how these behaviors are implemented.

Convergence and Divergence of Language Models under Different Random Seeds

In this paper, we investigate the convergence of language models (LMs) trained under different random seeds, measuring convergence as the expected per-token Kullback--Leibler (KL) divergence across seeds. By comparing LM convergence as a function of model size and training checkpoint, we identify a four-phase convergence pattern: (i) an initial uniform phase, (ii) a sharp-convergence phase, (iii) a sharp-divergence phase, and (iv) a slow-reconvergence phase. Further, we observe that larger models reconverge faster in later training stages, while smaller models never actually reconverge; these results suggest that a certain model size may be necessary to learn stable distributions. Restricting our analysis to specific token frequencies or part-of-speech (PoS) tags further reveals that convergence is uneven across linguistic categories: frequent tokens and function words converge faster and more reliably than their counterparts (infrequent tokens and content words). Overall, our findings highlight factors that influence the stability of the learned distributions in model training.

Privileged Self-Access Matters for Introspection in AI

Whether AI models can introspect is an increasingly important practical question. But there is no consensus on how introspection is to be defined. Beginning from a recently proposed ''lightweight'' definition, we argue instead for a thicker one. According to our proposal, introspection in AI is any process which yields information about internal states through a process more reliable than one with equal or lower computational cost available to a third party. Using experiments where LLMs reason about their internal temperature parameters, we show they can appear to have lightweight introspection while failing to meaningfully introspect per our proposed definition.

semantic-features: A User-Friendly Tool for Studying Contextual Word Embeddings in Interpretable Semantic Spaces

We introduce semantic-features, an extensible, easy-to-use library based on Chronis et al. (2023) for studying contextualized word embeddings of LMs by projecting them into interpretable spaces. We apply this tool in an experiment where we measure the contextual effect of the choice of dative construction (prepositional or double object) on the semantic interpretation of utterances (Bresnan, 2007). Specifically, we test whether "London" in "I sent London the letter." is more likely to be interpreted as an animate referent (e.g., as the name of a person) than in "I sent the letter to London." To this end, we devise a dataset of 450 sentence pairs, one in each dative construction, with recipients being ambiguous with respect to person-hood vs. place-hood. By applying semantic-features, we show that the contextualized word embeddings of three masked language models show the expected sensitivities. This leaves us optimistic about the usefulness of our tool.

Is It JUST Semantics? A Case Study of Discourse Particle Understanding in LLMs

Discourse particles are crucial elements that subtly shape the meaning of text. These words, often polyfunctional, give rise to nuanced and often quite disparate semantic/discourse effects, as exemplified by the diverse uses of the particle "just" (e.g., exclusive, temporal, emphatic). This work investigates the capacity of LLMs to distinguish the fine-grained senses of English "just", a well-studied example in formal semantics, using data meticulously created and labeled by expert linguists. Our findings reveal that while LLMs exhibit some ability to differentiate between broader categories, they struggle to fully capture more subtle nuances, highlighting a gap in their understanding of discourse particles.

Causal Interventions Reveal Shared Structure Across English Filler-Gap Constructions

Large Language Models (LLMs) have emerged as powerful sources of evidence for linguists seeking to develop theories of syntax. In this paper, we argue that causal interpretability methods, applied to LLMs, can greatly enhance the value of such evidence by helping us characterize the abstract mechanisms that LLMs learn to use. Our empirical focus is a set of English filler-gap dependency constructions (e.g., questions, relative clauses). Linguistic theories largely agree that these constructions share many properties. Using experiments based in Distributed Interchange Interventions, we show that LLMs converge on similar abstract analyses of these constructions. These analyses also reveal previously overlooked factors -- relating to frequency, filler type, and surrounding context -- that could motivate changes to standard linguistic theory. Overall, these results suggest that mechanistic, internal analyses of LLMs can push linguistic theory forward.

A suite of LMs comprehend puzzle statements as well as humans

Recent claims suggest that large language models (LMs) underperform humans in comprehending minimally complex English statements (Dentella et al., 2024). Here, we revisit those findings and argue that human performance was overestimated, while LLM abilities were underestimated. Using the same stimuli, we report a preregistered study comparing human responses in two conditions: one allowed rereading (replicating the original study), and one that restricted rereading (a more naturalistic comprehension test). Human accuracy dropped significantly when rereading was restricted (73%), falling below that of Falcon-180B-Chat (76%) and GPT-4 (81%). The newer GPT-o1 model achieves perfect accuracy. Results further show that both humans and models are disproportionately challenged by queries involving potentially reciprocal actions (e.g., kissing), suggesting shared pragmatic sensitivities rather than model-specific deficits. Additional analyses using Llama-2-70B log probabilities, a recoding of open-ended model responses, and grammaticality ratings of other sentences reveal systematic underestimation of model performance. We find that GPT-4o can align with either naive or expert grammaticality judgments, depending on prompt framing. These findings underscore the need for more careful experimental design and coding practices in LLM evaluation, and they challenge the assumption that current models are inherently weaker than humans at language comprehension.