Tag
The paper introduces attention manifolds — learned 2D B-spline surfaces that modulate value dimensions in transformer attention — showing 2–2.5 point perplexity improvements on LLaMA 3.2 with just 0.3% parameter overhead, plus mechanically editable surfaces enabling steering and attention walls for safety.
该论文揭示了音频-视觉大语言模型中"源混淆接地幻觉"的内在机制——问题中继(question-relay)会携带干扰模态的线索,并提出训练无关的SECRET方法,在保持生成位置通路的同时将问题状态导向所需的源证据,在CMM和AVHBench基准上相比基线模型最高提升18.0和7.1个百分点。
This study explores whether large language models have distinct internal representations of pain, finds that they do, and examines the functional consequences through experiments, concluding with implications for AI safety and welfare.
This paper introduces an information-asymmetric spot-the-difference task to measure epistemic vigilance in vision-language models, finding that models often overlook private evidence to agree with partners. Model steering to reduce sycophancy improves reliability in cooperative tasks.
HARP is a training-free interpretability method that uses an LLM agent with a vector database of activations and tools for manipulation, outperforming training-based methods on concept discovery, detection, steering, and secret elicitation.
The author expresses skepticism about linear activation model steering methods like J-space, arguing that related research often relies on cherry-picked special cases and is therefore unreliable.
This paper shows that in chat models, refusal behavior is gated by a compliant model persona direction at late layers, rather than being an isolated mechanism. Steering persona suppresses refusal, and reintroducing refusal partially restores it only at late layers, revealing a coupling between persona and safety representations.
DenseSteer is a training-free inference-time framework that improves small language models' math reasoning by steering their internal representations towards dense reasoning patterns, achieving accuracy gains without increasing token-level negative log-likelihood.
This paper introduces geometric stability measures—based on pairwise distance consistency in representations—to predict language model steerability and detect structural drift. Supervised variants achieve near-perfect correlation (ρ=0.89-0.97) with linear steerability across 35-69 embedding models, while unsupervised variants outperform CKA and Procrustes for post-deployment drift detection.