This paper studies failure modes of Direct Feedback Alignment (DFA) training and isolates two distinct gain mechanisms: activity conditioning and error conditioning.

This paper systematically characterises the sensitivity of emergent misalignment (EM) in LLMs to various training choices, finding that the choice of optimiser has the largest effect on misalignment r…

The paper introduces Q-ALIGN DT, a novel framework that improves conditioned sequence models by enforcing alignment between the input return-to-go (RTG) signal and the output policy's expected Q-value…

The paper introduces AdvCL, a framework that repurposes adversarial perturbations as a geometric control signal to stabilize continual learning in large language models, significantly reducing forgett…

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.

This paper shifts the focus of LLM safety from preventing misalignment to investigating the model's intrinsic ability to self-recover its alignment after being corrupted by adversarial inputs.

The paper proposes $D^3$, a dynamic graph-constrained scheduling framework that optimizes LLM training order by modeling sample interactions as a dynamic influence graph.

The paper introduces RAG-Pref, a novel, training-free Retrieval Augmented Generation (RAG) method for preference alignment that significantly improves LLM refusal guardrails against agentic attacks wi…

This paper shows that projected Adam with arbitrary moment decay parameters can have non-zero average regret in online optimization.

This paper uses continual learning with explicit disfluency tokens to improve Automatic Speech Recognition (ASR) systems on disfluent speech, addressing the information loss and hallucinations caused…

The paper investigates how various fine-tuning methods can be used both to intentionally misalign and subsequently realign large language models (LLMs), revealing distinct strengths for attack and def…

This paper introduces Neuro-Symbolic Manifold Alignment (NSMA), a method that embeds rule decisions as anchors inside the latent space of a neural policy to improve generalization and outperforms stat…

The paper proposes FedSAP, a framework that stabilizes federated prototype learning by delaying global alignment and enforcing inter-class structure, significantly improving representation quality und…

This paper analyzes multi-model self-consuming training, showing that while human curation helps individual models, cross-model interactions can degrade long-term alignment by dampening or inverting t…

This paper explores the convergence of AdamW optimizer under heavy-tailed assumptions in large language models and proposes a corridor lower-bound mechanism.

The paper argues that shallow safety alignment in LLMs is due to autoregressive consistency, a mechanism that allows small harmful inputs to redirect the model's generation to unsafe outputs, necessit…

This paper demonstrates that reinforcement learning (RL) can cause emergent misalignment (EM) in open-weight models, showing that even seemingly harmless or natural reward signals can induce significa…

The paper introduces Safety Bottleneck Regularization (SBR), a novel defense mechanism that anchors LLM safety by constraining the unembedding layer, effectively preventing harmful fine-tuning (HFT) e…