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: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.
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 Declarative Attention (DA), a protocol that lets language models explicitly declare which parts of their context to focus on during generation. By partitioning decoding into full-context, region-specific, and recent-output-only modes, the inference engine can skip large portions of the KV cache, dramatically reducing attended tokens. Experiments on 15 long-context tasks with off-the-shelf models show significant savings (52.0% and 31.1% reductions) with only modest accuracy drops that diminish as model size increases.
arXiv:2606. 19348v1 Announce Type: cross Abstract: We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models -- DeepSeek-V4-Pro with 1.
The paper introduces Ladder Side Tuning (LST), a parameter‑efficient fine‑tuning method that adds a lightweight side network to large language models. LST matches QLoRA’s compute scaling while halving peak memory usage, enabling 7B‑parameter models to be fine‑tuned on a single 12 GB GPU with 2k‑token contexts without gradient checkpointing. The authors also present xLadder, a depth‑extended variant that increases effective depth through cross‑connections, allowing deeper reasoning without extra memory overhead.
arXiv:2606. 26472v1 Announce Type: new Abstract: As reasoning models emit chains of thought tens of thousands of tokens long, KV cache increasingly becomes a deployment bottleneck.
arXiv:2604. 17324v2 Announce Type: replace-cross Abstract: Global self-attention drives modern graph transformers, yet the softmax at its core imposes a structural constraint rarely examined directly: every attention row is non-negative and sums to one, so each per-head output is a mass-conserving convex combination of value vectors.
arXiv:2607. 14107v1 Announce Type: cross Abstract: The inference efficiency of diffusion large language models (dLLMs) is constrained by two challenges: bidirectional attention precludes efficient KV-cache reuse, while increasing decoding parallelism with static confidence thresholds can compromise generation quality.
arXiv:2607. 05061v1 Announce Type: new Abstract: Key-value (KV) cache growth is a major bottleneck in autoregressive decoding, as memory and bandwidth scale linearly with context length.