Trendoday

LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time

Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across t

  • Upvotes27
–Rank now
#6Peak rank
3Days on list
361Score · 7d
Nov 9, 2026First seen
Rank · 7 days

About this paper

A method compresses KV caches by predicting values from keys within each attention head and applying cross-layer low-rank key compression, enabling higher decoding throughput and composability with quantization and eviction.

Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or lower numerical precision. We identify a complementary axis: within each attention head, values are largely predictable from their corresponding keys. We exploit this with value cache amortisation, which fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, giving 4times near-lossless value compression. The predictor can be initialised by meta-learning for substantially reduced prefill…

autoregressive decodingKV cachecontext lengthpost-training methodsnumerical precisionattention headvalue cache amortisationper-head predictorresidualsvalue compression
Upvotes
27
Authors
5
Published
Nov 4, 2026

Summary and abstract excerpt via Hugging Face Papers ↗ · arXiv.

Trending history

LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time was first seen on Hugging Face · Daily papers on Nov 9, 2026. Since then it has appeared in 18 checks on 3 different days. Its best position was #6 of 30, reached on Nov 10, 01:28 UTC. It dropped off the list after Nov 11, 01:28 UTC.

Rank history

Position on Hugging Face · Daily papers at every check (lower is better).

#1#11#20#30 15:2823:5808:2816:5801:28 Nov 9, 15:28 UTC · #8Nov 9, 17:28 UTC · #7Nov 9, 19:28 UTC · #7Nov 9, 21:28 UTC · #7Nov 9, 23:28 UTC · #7Nov 10, 01:28 UTC · #6Nov 10, 03:28 UTC · #6Nov 10, 05:28 UTC · #6Nov 10, 07:28 UTC · #8Nov 10, 09:28 UTC · #10Nov 10, 11:28 UTC · #9Nov 10, 13:28 UTC · #11Nov 10, 15:28 UTC · #9Nov 10, 17:28 UTC · #8Nov 10, 19:28 UTC · #7Nov 10, 21:28 UTC · #7Nov 10, 23:28 UTC · #7Nov 11, 01:28 UTC · #7

Last 14 days

Best rank per day (UTC).

Nov 3today

Metric history

Platform metric as reported by Hugging Face.

0112234 15:2823:5808:2816:5801:28 Nov 9, 15:28 UTC · 27Nov 9, 17:28 UTC · 29Nov 9, 19:28 UTC · 29Nov 9, 21:28 UTC · 29Nov 9, 23:28 UTC · 30Nov 10, 01:28 UTC · 30Nov 10, 03:28 UTC · 30Nov 10, 05:28 UTC · 30Nov 10, 07:28 UTC · 30Nov 10, 09:28 UTC · 30Nov 10, 11:28 UTC · 30Nov 10, 13:28 UTC · 30Nov 10, 15:28 UTC · 30Nov 10, 17:28 UTC · 31Nov 10, 19:28 UTC · 34Nov 10, 21:28 UTC · 34Nov 10, 23:28 UTC · 34Nov 11, 01:28 UTC · 34

Questions

Is LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time trending on HF Daily Papers right now?

Not at the moment. It was last on the list on Nov 11, 01:28 UTC, after 3 days of trending.

What is the highest rank LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time reached?

#6 on Hugging Face · Daily papers, on Nov 10, 01:28 UTC. Trendoday checks this list every 120 minutes.

How long has LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time been trending?

It was first seen on Nov 9, 2026 and has been on the list on 3 days (18 checks) since.