# Multi-State Linear Attention with Generalized Second-Order Routing

Source: [arXiv](https://arxiv.org/abs/2610.37645v1)  
Feed7 permalink: https://feed7.dev/p/2610-37645v1-12yb9up  
Published: 2026-10-27T17:55:49.000Z  
Trust: Needs Review (needs_review)

## Why Included

MERLIN adds mean and variance estimates to how multi-state linear attention routes information from long-context segments. It improves reported RULER and LongBench results on two model variants, including against baselines using four times more memory.

## Source Summary

MERLIN tracks the first and second moments of each segment’s keys, letting a query estimate both mean and variance in its alignment with that segment. On **Gated DeltaNet-1.3B** and **Mamba-2-780M**, it improves reported RULER and LongBench results, including over some baselines using **4× more memory**.

## Practical Implication

Builders exploring long-context models can watch this approach as a way to improve how compressed segment memories are read. It suggests routing quality may matter alongside the amount of memory available.

## Agent-Ready Context

MERLIN tracks the first and second moments of each segment’s keys, letting a query estimate both mean and variance in its alignment with that segment. On **Gated DeltaNet-1.3B** and **Mamba-2-780M**, it improves reported RULER and LongBench results, including over some baselines using **4× more memory**.

Builders exploring long-context models can watch this approach as a way to improve how compressed segment memories are read. It suggests routing quality may matter alongside the amount of memory available.

These are results on multi-state variants and specific long-context evaluations. The material does not say how the method behaves in deployed agent workloads or across other model families.

## Connected Context

Feed7 judgment across 944 accumulated Signals:

MERLIN adds first- and second-moment routing to compressed segment memories and reports gains on two long-context model families. It suggests that how a model reads stored segments can matter alongside memory capacity, while leaving performance in deployed agent workloads unestablished.

- [ConvMem: Convolutional Memory for Long-Context Reasoning](https://feed7.dev/p/2609-10441v1-066ygq8) — Both change how long inputs are represented and accessed: ConvMem builds hierarchical summaries, while MERLIN routes queries using segment key moments. Comparing them could clarify the tradeoff between memory organization and readout quality; neither supplied summary establishes coding-agent value.
- [Structured Memory for Edge Language Models: Persistent Context and Corpus Retrieval via O(1) SSM State Injection](https://feed7.dev/p/2608-02560v1-1u8ls14) — PRECOG injects reusable corpus states into compatible SSMs to reduce repeated prefill, while MERLIN targets query routing over compressed segment memories and includes Mamba-2 results. They address distinct parts of SSM context handling, and neither summary establishes RAG-equivalent quality or deployed-agent performance.

## Context Map

- Layer: model
- Domains: research
- Topics: open-models, context-engineering, retrieval

## Uncertainty

- These are results on multi-state variants and specific long-context evaluations. The material does not say how the method behaves in deployed agent workloads or across other model families.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.