LLMs

SelKV tackles the linear memory growth of LLM KV caches by selectively merging or dropping tokens based on value-vector similarity. The method uses a soft cosine gate to modulate these decisions, preventing the representation degradation and attention sag common in indiscriminate aggregation approaches. This training-free framework aims to maintain model fidelity while significantly reducing the memory footprint during autoregressive generation.

• Eliminates need for retraining or fine-tuning via a training-free approach.
• Reduces KV cache memory usage by merging similar tokens and dropping dissimilar ones.
• Mitigates attention sag by compensating for softmax mass distribution mismatches.
• Uses soft cosine gating to make adaptive, per-token merge-or-drop decisions.

A new arXiv paper identifies a structured confound in RLHF where pairwise preference labels reflect the annotator's mental state rather than just output quality. Under stress, raters' preferences shift in ways that create systematic bias in reward models. This state-dependent error propagates through policy optimization, distinct from ordinary disagreement or random noise. The authors propose an audit framework to detect and mitigate this specific source of bias.

• RLHF preference data may encode annotator stress levels, not just response quality.
• State-dependent bias is systematic and propagates through reward modeling.
• Distinguish rater state shifts from random label noise or normal disagreement.
• New audit framework proposed to test for and mitigate this structured confound.

The paper identifies that self-generated rollouts in RL with verifiable rewards often suffer from semantic redundancy, causing models to converge on erroneous reasoning basins with low reward contrast. To fix this, the authors propose W2SPO, an off-policy method that leverages a weaker, computationally efficient auxiliary model to inform the stronger policy's exploration. This weak-to-strong paradigm helps break out of local optima by introducing diverse trajectories that the primary model might miss.

• Solves semantic redundancy in self-generated rollouts for reasoning tasks.
• Uses a weaker auxiliary model to guide stronger model exploration.
• Off-policy RL approach avoids converging on erroneous reasoning basins.
• Improves reward contrast for more effective policy updates.

LaCache accelerates Diffusion-based Large Language Models by eliminating operator-level redundancy in Semi-Autoregressive decoding. The framework employs Lossless State Memoization to cache embedding outputs, RoPE pre-attention states, and FlashAttention softmax statistics. By reusing these invariant components during denoising, it achieves training-free acceleration without sacrificing generation quality.

• Targets Diffusion LLMs using Semi-Autoregressive decoding for parallel text generation.
• Caches EmbedCache, RoPECache, and FACache to avoid recomputing invariant sequence blocks.
• Training-free approach requiring no model retraining or architectural changes.
• Reduces redundancy by recognizing prefix and masked suffix invariance within denoising steps.

SpecLA introduces a speculative decoding runtime tailored for stateful linear-attention models, which use recurrent states instead of growing KV caches. Existing speculative systems fail here because verification must respect recurrent dependencies across chains and branches. The approach ensures acceptance updates only the accepted state trajectory and prevents the drafter from wasting work on invalid candidates.

• Linear-attention models use recurrent states, not KV caches, requiring new verification logic.
• SpecLA handles topology-aware verification for chains and trees in stateful targets.
• Acceptance logic updates only the accepted state trajectory to maintain correctness.
• Drafter constraints prevent submitting candidates that waste stateful verification work.