CacheTrap achieves 100% targeted attack success on five open-source LLMs by using an efficient search to locate and flip a single bit in the KV cache as a transient trigger, while preserving normal accuracy without the trigger.
hub
Massive Activations in Large Language Models
47 Pith papers cite this work. Polarity classification is still indexing.
47
Pith papers citing it
abstract
We observe an empirical phenomenon in Large Language Models (LLMs) -- very few activations exhibit significantly larger values than others (e.g., 100,000 times larger). We call them massive activations. First, we demonstrate the widespread existence of massive activations across various LLMs and characterize their locations. Second, we find their values largely stay constant regardless of the input, and they function as indispensable bias terms in LLMs. Third, these massive activations lead to the concentration of attention probabilities to their corresponding tokens, and further, implicit bias terms in the self-attention output. Last, we also study massive activations in Vision Transformers. Code is available at https://github.com/locuslab/massive-activations.
hub tools
citation-role summary
background 3
method 1
citation-polarity summary
representative citing papers
In 160M and 290M parameter models, a new residual-stream split into scratch and protected channels causes massive activations to re-emerge in the protected decode channel, more concentrated on the start token.
The normalized inverse-scale direction of LayerNorm's affine parameters is an exact algebraic kernel of the post-final-norm centred activation covariance for any input distribution in LayerNorm transformers.
Reroute turns irreversible visual-token pruning into recoverable routing that reuses existing attention scores, improving grounding performance under aggressive reduction on LLaVA-1.5 and Qwen while preserving TFLOPs and KV-cache budgets.
AVLLMs route audio-visual information sequentially in video tasks and via parallel streams for interleaved items, allowing early token discard with little performance loss across models and scales.
Dead directions recover Watanabe's RLCT contribution and triple (λ, m, ν) from directional Fisher curvature decay rates in original parameter space for singular models, extended via K-FAC to networks and gauge-equivariant optimizers.
Mechanistic analysis of GLMs shows graph sink tokens have high activation but low importance for predictions, indicating decoupling between saliency and graph-semantic utility.
YARD is a training-free method using Y-shaped decoder architecture and register tokens to improve contrastive decoding for hallucination reduction in LVLMs with lower latency.
Bayesian Filtering Transformer reframes attention as precision-weighted kriging and residual connections as Kalman updates, delivering gains on cold-start recommendation and noisy LLM fine-tuning tasks.
Massive activations first appear in a single ME Layer due to RMSNorm and FFN, remain invariant thereafter, and a simple softening method raises LLM performance while reducing attention sinks.
FlashAttention-3 achieves 1.5-2x speedup on H100 GPUs for attention, reaching 740 TFLOPs/s (75% utilization) in FP16 and near 1.2 PFLOPs/s in FP8 while cutting numerical error by 2.6x versus baseline FP8 attention.
K-sparse autoencoders with dead-latent fixes produce clean scaling laws and better feature quality metrics that improve with size, shown by training a 16-million-latent model on GPT-4 activations.
MiniMax Sparse Attention is a GQA-based block-sparse attention mechanism that selects top-k blocks independently per group and delivers 28.4x per-token compute reduction at 1M context with on-par performance plus 14.2x prefill and 7.6x decode speedups via co-designed GPU kernel.
DynamicPTQ uses new metrics of residual-stream dynamics to apply 8-bit activation precision only to quantization-sensitive layers in W4A4KV4 LLM inference, improving perplexity and QA performance over static smoothing baselines.
ICALens applies an optimized ICA workflow to LLM activations and recovers compact interpretable directions that match or exceed public SAEs on SAEBench probing and perturbation tasks without per-layer dictionary training.
A single dominant layer in LLMs, found by activation outliers, accounts for most ZO fine-tuning gains and can replace full-model updates across models and tasks.
Contribution Weights combine attention, value magnitude, and directional alignment to measure token influence more faithfully than attention alone, and show attention sinks actively suppress information via a convex sink-rate to output-norm relationship.
OScaR mitigates token norm imbalance via canalized rotation and omni-token scaling to enable near-lossless INT2 KV cache quantization with up to 3x decoding speedup and 5.3x memory reduction.
UniRefiner uses contrastive registers and a dual alignment objective to remove three categories of spurious tokens from pre-trained ViTs, yielding up to 9.4% mIoU gains on ADE20K and 22% zero-shot segmentation improvements.
A Weibull diagnostic framework classifies transformer weight matrices into consistent functional classes via the shape parameter k and tracks training progress via the scale parameter lambda across multiple architectures.
Register tokens enhance pixel-space DiT training and output quality via cleaner high-noise feature maps, and a dual-stream design adds further gains with little overhead.
LVLMs show vocabulary hijacking by inert tokens that decode to hijacking anchors; HABI locates them, NHAR finds resilient heads, and HAVAE boosts those heads to cut hallucinations.
Suppressing attention sinks in Stable Diffusion 3 does not degrade text-image alignment or preference metrics at mild intervention levels, though stronger suppression reveals sink-specific perceptual shifts larger than random masking.
Outlier tokens in DiTs are addressed with Dual-Stage Registers, which reduce artifacts and improve image generation on ImageNet and text-to-image tasks.
citing papers explorer
-
Reroute, Don't Remove: Recoverable Visual Token Routing for Vision-Language Models
Reroute turns irreversible visual-token pruning into recoverable routing that reuses existing attention scores, improving grounding performance under aggressive reduction on LLaVA-1.5 and Qwen while preserving TFLOPs and KV-cache budgets.
-
YARD: Y-Architecture Register Decoding for Efficient Hallucination Mitigation in Large Vision-Language Models
YARD is a training-free method using Y-shaped decoder architecture and register tokens to improve contrastive decoding for hallucination reduction in LVLMs with lower latency.
-
UniRefiner: Teaching Pre-trained ViTs to Self-Dispose Dross via Contrastive Register
UniRefiner uses contrastive registers and a dual alignment objective to remove three categories of spurious tokens from pre-trained ViTs, yielding up to 9.4% mIoU gains on ADE20K and 22% zero-shot segmentation improvements.
-
Registers Matter for Pixel-Space Diffusion Transformers
Register tokens enhance pixel-space DiT training and output quality via cleaner high-noise feature maps, and a dual-stream design adds further gains with little overhead.
-
Attention Sinks in Diffusion Transformers: A Causal Analysis
Suppressing attention sinks in Stable Diffusion 3 does not degrade text-image alignment or preference metrics at mild intervention levels, though stronger suppression reveals sink-specific perceptual shifts larger than random masking.
-
Taming Outlier Tokens in Diffusion Transformers
Outlier tokens in DiTs are addressed with Dual-Stage Registers, which reduce artifacts and improve image generation on ImageNet and text-to-image tasks.
-
Colinearity Decay: Training Quantization-Friendly ViTs with Outlier Decay
Colinearity-Decay regularizer trains ViTs that maintain or improve full-precision accuracy while delivering higher accuracy after low-bit quantization on ImageNet and COCO tasks.
-
Information-Regularized Attention for Visual-Centric Reasoning
IRA is a stochastic attention mechanism that regulates visual information injection in VLMs to yield smoother embedding trajectories and reduced attention sinks.
-
DuQuant++: Fine-grained Rotation Enhances Microscaling FP4 Quantization
DuQuant++ adapts outlier-aware fine-grained rotation to MXFP4 by matching block size to the 32-element microscaling group, enabling a single rotation that smooths distributions and achieves SOTA performance on LLaMA-3 with lower cost.
- When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models