11 papers · 1 filter
Understanding and Exploiting Anisotropy in Post-Training
Samyak Jha, Harshvardhan Saini, Yizhen Liao +2
LLM post-training combines supervised fine-tuning (SFT), a mode-covering forward-KL objective, with reinforcement learning (RL), a mode-seeking reverse-KL objective. Frequency-weig…
Masking Frequent Tokens Sharpens Direct Preference Optimization
Harshvardhan Saini, Samyak Jha, Yiming Tang +1
Direct Preference Optimization (DPO) aligns language models by optimizing over sequence-level sums of token-wise implicit reward differences. However, we identify a pervasive patho…
Capsule Lens: Locating and Tracking Concept Geometry in Model Representations
Yiming Tang, Harshvardhan Saini, Samyak Jha +3
Understanding how concepts are encoded in the internal representations of machine learning models is a central problem in mechanistic interpretability, essential both for the scien…
Emergent Hierarchical Monosemantic Neurons from the Group-Contrastive Forward-Forward Algorithm
Yiming Tang, Qinglin Qi, Zhaoqian Yao +2
Mechanistic interpretability has made significant strides in understanding neural network representations, with sparse dictionary learning (SDL) methods, most prominently sparse au…
When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models
Harshvardhan Saini, Samyak Jha, Yiming Tang +1
Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing conten…
JEDI: Joint Embedding Diffusion World Model for Online Model-Based Reinforcement Learning
Jing Yu Lim, Rushi Shah, Zarif Ikram +4
Diffusion world models have recently become competitive for online model-based reinforcement learning, but current approaches expose a tension: pixel diffusion is effective but com…
AB-Sparse: Sparse Attention with Adaptive Block Size for Accurate and Efficient Long-Context Inference
Di Liu, Ruitian Wang, Chen Chen +6
As large language models scale to longer contexts, loading the growing KV cache during attention computation becomes a critical bottleneck. Previous work has shown that attention c…
Absurd World: A Simple Yet Powerful Method to Absurdify the Real-world for Probing LLM Reasoning Capabilities
Ryan Albright, Golam Md Muktadir, Zarif Ikram +3
While extremely powerful and versatile at various tasks, the thinking capabilities of large language models (LLMs) are often put under scrutiny as they sometimes fail to solve prob…
Mitigating Premature Discretization with Progressive Quantization for Robust Vector Tokenization
Wenhao Zhao, Qiran Zou, Zhouhan Lin +1
Vector Quantization (VQ) has become the cornerstone of tokenization for many multimodal Large Language Models and diffusion synthesis. However, existing VQ paradigms suffer from a…
VQKV: High-Fidelity and High-Ratio Cache Compression via Vector-Quantization
Yixuan Wang, Qingyu Shi, Jiayu Zhou +3
The growing context length of Large Language Models (LLMs) enlarges the Key-Value (KV) cache, limiting deployment in resource-limited environments. Prior training-free approaches f…
Bridging Mechanistic Interpretability and Prompt Engineering with Gradient Ascent for Interpretable Persona Control
Harshvardhan Saini, Yiming Tang, Dianbo Liu
Controlling emergent behavioral personas (e.g., sycophancy, hallucination) in Large Language Models (LLMs) is critical for AI safety, yet remains a persistent challenge. Existing s…