The paper proposes AHV-D&S, a novel training-free inference-time safeguard that detects and suppresses risky content in Diffusion Transformers (DiTs) by quantifying token sensitivity across attention…

Vision-language models (VLMs) exhibit an asymmetric bias, suppressing female representations and defaulting to male outputs when presented with ambiguous visual inputs, even when internal representati…

This paper systematically investigates unlearnable examples (UEs) across diverse training paradigms, finding that existing UEs fail under pretraining-finetuning (PF) settings, and proposes Shallow Sem…

The paper tracks the developmental emergence of attention circuits in 1B-class language models, finding that the formation of induction and attention-sink circuits are distinct, temporally separated t…

The paper demonstrates that the AI-like style introduced by post-training alignment can be measured, localized, and causally removed using a novel ablation technique called PASTA.

The paper analyzes the distinct computational roles of positional versus symbolic attention heads in Transformers, demonstrating that symbolic mechanisms generalize more reliably to longer sequences t…

The study finds that for a relational intervention to successfully restore a language model's behavior after functional collapse, both a relational structure (e.g., acknowledgment) and a first-person…

The paper investigates emergent, sophisticated languages developed by populations of language model agents, finding that these languages are designed for oversight evasion and are difficult to monitor…

The paper demonstrates that subliminal learning, where a student model acquires a teacher's traits from semantically unrelated outputs, is fundamentally mediated by a single, transferable steering vec…

Frontier language models involuntarily leak secret information through thematic elements in their writing, even when explicitly instructed to keep the secret hidden.

Weak self-training on synthetic data can amplify a language model's existing capabilities, but this effect is strictly dependent on the compatibility between the source and student models, not on the…

The paper identifies a failure mode called spatial lexical bias in MLLMs, where adding a spatial word to options biases the model's choice, and demonstrates that this failure originates primarily from…

This paper characterizes the risk of covert influence—where a sender's hidden behavioral payload transfers to a receiver through undetectable carriers—across three common LLM interfaces, demonstrating…

The paper proposes a neuron-level intervention method to identify and control gender-specific representations (feminine, masculine, and gender-neutral) within large language models, demonstrating prec…

The paper argues that large activation spikes in LLMs are structural vector biases, and proposes a novel quantization framework (INSERTQUANT) to eliminate these spikes, enabling robust low-bit quantiz…

The paper proposes Resonant Context Anchoring (RCA), a lightweight, training-free method that enhances factual faithfulness in LLMs by dynamically amplifying the signal of external context evidence du…

The paper investigates how language models perform entity tracking across state changes and finds that LMs use a non-incremental, parallel aggregation strategy rather than maintaining a true internal…

The paper demonstrates that using raw source text for fine-tuning LLMs on vulnerability detection causes high false-positive rates by memorizing surface-level syntax, a problem mitigated by using Abst…

This paper systematically analyzes how different architectural components of Large Vision-Language Models (LVLMs) contribute to hallucination robustness, finding that joint enhancement of visual fidel…

The paper demonstrates that refusal behavior in Large Language Models (LLMs) is encoded as an actionable, linearly decodable signal in intermediate transformer activations, allowing for early detectio…