Learned KV-cache eviction often faces a soft-to-hard mismatch: during training, differentiable gates typically attenuate token contributions, whereas inference saves memory only when KV entries are ph...
The paper examines why learned gates in sparse attention models offer little advantage over random gates when jointly trained with the transformer. Through experiments on a 31M-parameter transformer, the authors attribute this to routing absorption, where the model’s representations adapt to the imposed mask, diminishing the benefit of learned routing. They also explore hard masking, stochastic mask training, and the impact of trainable attention layers on gate performance, concluding that freezing the model stabilizes routing targets for effective post‑hoc sparsification.
arXiv:2610.14754v1 Announce Type: new Abstract: Hybrid language models interleave Gated DeltaNet (GDN) layers, whose recurrent memory has a fixed size, with full-attention layers, whose key-value (KV...
ValueDiff introduces a value‑geometric KV cache eviction strategy for large language models that suppress attention sinks. It ranks tokens by the L2 deviation of their value vectors from the cache mean, a score that aligns with minimal‑disturbance eviction under a max‑entropy assumption. Across several benchmarks—RULER, LongBench, and MATH‑500—ValueDiff consistently retains a higher proportion of useful tokens than prior methods, especially under tight cache budgets.
Elastic Threshold Attention (ETA) is a trainable attention mechanism that dynamically predicts contextual thresholds from query representations, enabling selective pruning of KV cache tokens during long‑context decoding. By multiplicatively suppressing sub‑threshold logits during training, ETA avoids representation collapse and eliminates localized attention sinks, allowing a 1.45B model to match dense attention performance at roughly 85% training sparsity and 38% active decode density. At inference, a custom Triton kernel achieves up to 2.5× faster decoding on sequences up to 512K tokens, and an offline calibration step can further reduce compute by 27% by freezing per‑head thresholds.
arXiv:2606. 09079v1 Announce Type: cross Abstract: Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving.
The paper investigates using short polynomial approximations to accelerate special‑function operations in large language models on NVIDIA Blackwell GPUs. By replacing native sigmoid, tanh, and SiLU with degree‑3 or degree‑4 bfloat16 programs, the authors achieve up to 2.19× speed‑ups in isolated FP16 benchmarks and modest training‑step throughput gains (2.7–8.0%) across four integration tasks. The study also evaluates model behavior, finding negligible training‑loss differences within 100 billion tokens.
The paper introduces GatedV, an attention variant that applies the SiLU activation only to the value matrix before aggregation, adding no extra trainable parameters. Experiments on a decoder‑only Transformer show that, with matched parameter counts and training budgets, GatedV consistently lowers held‑out perplexity—by 2.0% at 2.33M parameters and 2.4% at 26.16M parameters—across multiple seeds. Ablations indicate that the performance gains are not fully captured by simple affine rescaling of values, and a benchmark on NVIDIA DGX Spark reports a modest 0.7% increase in full‑context forward latency for the larger model.
The paper introduces Universal Attention, a transformer architecture that unifies and extends decay-based KV‑cache pruning mechanisms while preserving RoPE embeddings and Softmax attention. It achieves highly expressive, end‑to‑end trainable pruning that adaptively removes low‑contribution tokens, yielding up to 10× compression on standard datasets and 25× compression at 16k context length without sacrificing downstream performance. The approach outperforms existing state‑of‑the‑art baselines and unpruned models in both compression and generalization.
arXiv:2606. 17872v1 Announce Type: cross Abstract: Large language models (LLMs) outperform earlier architectures on generative inference and long-context tasks, but their large size introduces significant challenges in memory usage, energy cost, and on-device deployment.
The paper reports that chain‑of‑thought (CoT) supervised fine‑tuning (SFT) improves reasoning but systematically harms long‑context recall in hybrid linear‑attention models such as HypeNet and Jet‑Nemotron. Retrieval performance on the Needle‑In‑A‑Haystack benchmark drops dramatically after CoT‑SFT, especially with harder settings and longer contexts. The authors introduce QK‑Restore, a training‑free method that reinstates the query‑key projection matrices from the pre‑SFT checkpoint, which recovers long‑range recall while preserving reasoning gains.
The paper introduces First Passage Attention (FPA), a new attention mechanism that replaces the traditional summation of decay contributions with a path‑extrema approach, lowering a key’s logit based on the deepest dip of a learned scalar field over the intervening span. FPA generalises the Forgetting Transformer (FoX) and, when the field and levels share a learned walk, naturally forms a stack of alive keys. Experiments show that a constant‑dimension FPA head can realise any subset of keys left alive by subsequent dips, outperforming softmax heads and FoX baselines in synthetic tasks, and matching FoX perplexity while achieving perfect passkey retrieval at eight times the training length in a 110M‑parameter language model.