arXiv:2607. 25663v1 Announce Type: new Abstract: Transformer adaptation is typically distributed across model depth, even when the intended change is narrow.
arXiv:2602. 23164v2 Announce Type: replace Abstract: Foundation models must handle multiple generative processes, yet mechanistic interpretability largely studies capabilities in isolation; it remains unclear how a single transformer organizes multiple, potentially conflicting "world models".
The paper investigates how transformer models learn latent structure by training a small decoder-only transformer on three variants of the Alchemy benchmark. It finds that the model acquires different components of latent structure in discrete stages, with a notable asymmetry: it robustly composes fundamental transitions but struggles to decompose complex examples into atomic transitions. Layer‑specific causal interventions reveal plasticity windows where freezing layers delays or prevents stage completion, offering a detailed view of capability evolution during training.
arXiv:2601. 22510v2 Announce Type: replace-cross Abstract: Large language models (LLMs) often achieve strong benchmark accuracy yet remain brittle under small distribution shifts.
arXiv:2609.37371v1 Announce Type: cross Abstract: Model adaptation is typically governed by a fixed recipe, even though different update programs can produce substantially different behavioral outcom...
The paper introduces the AGI Maze Prediction Datasets and Benchmark, a lightweight, procedurally generated grid‑world testbed for evaluating predictive models, particularly Transformers, on tasks such as per‑step transition prediction, fixed‑horizon state prediction, and sequential textual‑observation prediction. It compares byte‑level Transformer baselines with two memory‑augmented architectures, showing that a pseudo‑video spatial‑memory Transformer achieves perfect validation accuracy on selected tasks and improves sequential text‑trace prediction, while a generic auxiliary latent‑memory Transformer does not consistently help. The study highlights that structured, task‑aligned working memory can be more effective than merely increasing latent capacity, and positions the benchmark as a compact setting for testing architectures that couple textual interfaces to learned structured state.
The paper investigates how large language models (LLMs) share a common Fisher‑Rao geometry in their next‑token probability distributions, revealing that behaviour largely determines this geometry while activation geometry depends on coordinate choices. Across transformer, state‑space, and recurrent architectures, output geometries align more closely than activation geometries, and this shared structure facilitates semantic‑category transfer and improves agreement with human word choices as models scale and train. The study further demonstrates that geometry can guide minimum‑disturbance interventions, enabling reusable control that preserves behaviour better than Euclidean methods and enhances steering, editing, attribution, dictionary learning, and fine‑tuning.
arXiv:2607. 01232v1 Announce Type: new Abstract: Reinforcement learning (RL) has become a central component of post-training large language models (LLMs), yet little is understood about how RL adaptation is distributed across transformer layers.
The paper introduces the AGI Maze Prediction Datasets and Benchmark, a lightweight testbed for evaluating how Transformers and other models learn world dynamics. The benchmark, built from procedurally generated grid worlds, includes per‑step transition prediction, fixed‑horizon state prediction, and sequential textual‑observation prediction, with source‑maze‑disjoint training and validation splits to test transferable action‑conditioned dynamics. Experiments show that a pseudo‑video spatial‑memory Transformer, which initializes and updates a two‑dimensional latent workspace from the input map and action history, achieves perfect validation accuracy on selected tasks and improves sequential text‑trace prediction, outperforming byte‑level and unstructured‑memory baselines and suggesting that structured, task‑aligned working memory is more effective than additional latent capacity alone.
MetaSteer is a new method for steering large language models that learns nonlinear, context-dependent interventions applied to attention projection matrices. Unlike traditional linear, context-independent techniques, MetaSteer adapts its effects based on the input, requiring no linear concept-geometry assumption. Trained once on a pooled preference corpus, it transfers zero‑shot to unseen concepts and out‑of‑distribution contexts, matching or surpassing strong task‑specific baselines on multiple benchmarks and model families.
arXiv:2609.07876v1 Announce Type: cross Abstract: Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linea...
arXiv:2607. 11875v1 Announce Type: cross Abstract: We present a theoretical framework to explain the emergence of inductive reasoning abilities in Transformer language models.
WorldAgen is a unified framework that jointly learns world modeling and action prediction using a shared Transformer backbone with two specialized heads. It introduces a Mixed Unidirectional Attention Mask to separate the world model and agent model, and enables Test-Time Training (TTT) by sampling exploratory actions and updating the world model with real state transitions. Experiments on CALVIN and LIBERO show that WorldAgen matches or surpasses state‑of‑the‑art methods, especially when TTT is applied to a few samples.