Search arXiv⌕ Search

arXiv · 2611.09709

RidgeFill: Training-Free Late-Layer Cache Fill for Faster Prefill

Abstract

RidgeFill shortens prefill by not computing most of a model's upper layers for the prompt. The lower layers run exactly up to a split; from the split layer's state a per-layer map predicts every later layer's input, and each late layer writes its cache from that prediction. The design follows YOCO and DeepSeek-V4.1-Flash, where the projection is trained into the model. RidgeFill retrofits it onto models that were never trained for it: the map is a ridge regression fitted once in closed form from one pass over calibration text, the layer's own projections run unchanged, and a chosen set of rows, the last few thousand tokens, and every decoding step run the full model. Nothing is trained and no token is dropped. Most of the work was in measurement, finding a repeatable methodology. Loss averaged over all tokens hides RidgeFill's error. Needle retrieval in the filled region, loss on tokens that depend on distant context, and splice tests that swap filled keys or values for exact ones show which component carries it: the keys on Llama-3.1-8B, half of it in their scale, and the values and the recurrent state on Qwen3.5-9B. On Qwen3.8-27B in a llama.cpp fork on an A100, time to first token is 1.25 to 1.31x faster at 16K to 120K tokens with the maps in VRAM; RULER changes by -1.2 and -1.4 points and LongBench v2 by -1.4; the 95 % intervals include zero but do not rule out losses of a few points. On Qwen3.8-Flash-Next in vLLM it is 1.50 to 1.60x faster than exact at 9K to 32K tokens, with RULER's needle tasks at 99.8 against 100.0 and perplexity on novel tokens 2.65 % higher [-0.02 %, +4.93 %]. On three 8 to 9B reference models, at the same split and tail as the identity fill, the fitted map retrieves 49.1 and 49.6 more filled-region RULER points on the two dense models (at that split Llama's loss is worse), and its long-document loss is +/-0.03 nats of the released SwiftKV checkpoint's.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Dylan Johnston, Robert P. Smith. 2026-11-06. RidgeFill: Training-Free Late-Layer Cache Fill for Faster Prefill. https://doi.org/10.5281/zenodo.23890246

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

A Theoretical Bridge Between Long-Tailed Recognition and Continual Learning

We theoretically and empirically establish a previously unexplored connection between Long-Tailed Recognition (LTR) and Continual Learning (CL). Specifically, we show that training on a long-tailed dataset with imbalance factor $\mathrm{IF}$ drives model parameters into an $O(1/\sqrt{\mathrm{IF}})$ neighborhood of the dominant-class solution, under both uniform and exponentially decaying cardinality distributions. Building on this result, we establish an objective-level reformulation of LTR as a sequential learning problem and prove that the CL objective upper-bounds the balanced LTR loss up to a constant determined by the residual Head loss. Motivated by this connection, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled framework that leverages standard off-the-shelf CL methods to sequentially learn Head and Tail classes while mitigating catastrophic forgetting. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions and demonstrate strong performance across diverse LTR benchmarks. In the foundation-model regime, CLTR with DualPrompt on a CLIP backbone outperforms standard adaptation baselines and is competitive with specialized methods using external semantic supervision. Our work bridges LTR and CL both theoretically and empirically, providing a principled approach for addressing long-tailed learning using standard CL strategies.

cs.LG↗

TRAM: Test-Time Risk Adaptation with Mixture of Agents

Deployed reinforcement learning agents often face safety requirements that are specified only after training: new hazard maps, revised risk thresholds, or behavioral alignment constraints. We study zero-update deployment-time adaptation in which a fixed library of risk-neutral source policies must be reused under a newly specified reward-risk tradeoff. We propose TRAM (Test-time Risk Adaptation with Mixture of Agents), a source-scored composition rule that evaluates each source under the target reward and an occupancy-based deployment risk, then selects actions using risk-adjusted source scores. Unlike training-time risk-sensitive methods tied to a fixed surrogate such as return variance, TRAM supports spatial barrier exposure, divergence to a reference behavior, and local volatility risks specified at test time. We make the surrogate nature of the method explicit: TRAM is not claimed to solve the full occupancy-control problem of the stitched policy, but it admits a measurable source-hull mismatch term connecting source-scored risk to realized risk. Experiments in gridworlds, MuJoCo Reacher, Safety-Gymnasium, and an inference-time LLM alignment setting show that TRAM improves deployment risk while preserving reward and requires no parameter updates at test time.

cs.LG↗

Towards the Worst-case Robustness of Large Language Models

Recent studies have revealed the vulnerability of large language models to adversarial attacks, where adversaries craft specific inputs to induce incorrect or harmful outputs. Although various empirical defenses have been proposed, their worst-case robustness remains poorly understood, raising concerns about how they would behave against stronger future attacks. Here, we study the worst-case robustness of LLMs from both empirical and theoretical perspectives. First, we provide tighter empirical upper bounds for the tested deterministic defenses using stronger white-box attacks, under which most tested configurations achieve nearly 0\% robustness. Second, we formulate randomized-smoothing certification for bounded and binary base functions as fractional and 0-1 knapsack problems, respectively, and derive grouped solvers for two token kernels. For example, on AdvBench, we certify a GPT-4o safety detector smoothed by a uniform kernel to an average certified $\ell_0$ radius of 2.02. In our stated suffix-attack setting, the same kernel certifies a Llama-3-8B safety detector to an average certified radius of 6.41.

cs.LG↗