AuraTracer智迹闻
中文

EVENT DOSSIER

What Attention Recalls and Recurrence Controls in Hybrid Language Models

2026-09-07 12:00 Models 🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
2mentions
SummaryAI generated

To address the unclear mechanisms of action of the attention mechanism and cyclic states in hybrid language models, the study proposed two caching-level intervention methods: Split-prefill and State-swap. Experimental验证 on Qwen3.5 and Falcon-H1 showed that precise retrieval mainly depends on attention persistence (accounting for 64%-98%), while the output language and persona patterns are primarily determined by cyclic states (with accuracy improvements of 70%-80% and 3-5 times). State-swap confirmed the causal relationship, that is, the answer content originates from the KV cache side, while the language style comes from the cyclic state side. Additionally, it was found that cyclic generation can also understand vocabulary shared with context, and the attention mechanism is responsible for retrieving previously spoken content.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
Falcon-H1Qwen3.5

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
Falcon-H1 × Qwen3.51

SignalsSIGNALS

Keyword heat
  • Qwen3.51
  • Falcon-H11

All reports (1)SOURCES

A arXiv cs.CL en 2026-09-07 12:00

What Attention Recalls and Recurrence Controls in Hybrid Language Models

The mechanisms of attention and cyclic states in hybrid language models remain unclear. Two caching-level interventions were proposed: Split-prefill, which only retains the KV cache or generates answers using cyclic states; State-swap, which pairs KV caches from different contexts with cyclic states for a single forward propagation. Verification on Qwen3.5 and Falcon-H1 showed that precise retrieval is maintained only through attention (accounting for 64-98% of total accuracy), while cyclic processes result in zero accuracy; the output language and character settings differ, with both methods maintaining accuracy through cyclic processes (70-80% and 3-5 times higher accuracy), while the KV cache only achieves about 1% language accuracy. State-swap confirmed the causal relationship: answer values come from the KV side, while language style comes from the cyclic side. Additionally, cyclic generation can also accept meaning shared with the context or partial vocabulary, attention provides a search for previously spoken content, and cyclic states shape how the model expresses itself.