Skip to content

φ-Attention: Head-dimension structure complementary to MoBA's position-level gating #41

Description

@wangjun112233

Complementarity, not competition

MoBA solves which positions to attend to (block-sparse gating on the sequence dimension). φ-Attention solves how attention heads should talk to each other (C5-cycle coupling on the head dimension). These two dimensions are orthogonal — combining them should be strictly additive.

The core mathematical insight

A 5-cycle adjacency matrix with cos72° weights has a unique spectral property:

  • Maximum eigenvalue = φ (golden ratio, 1.618034..., verified to 6.7e-16 precision)
  • This is the only weight that achieves this — it's algebraically necessary, not an engineering choice
  • Spectral gap: C5 gives 0.0707 vs standard attention's 0.0034 (20× improvement)

In MoBA's terms: you gate which blocks each query sees (position-level sparsity). We gate which heads share information via a fixed C5 graph (head-level coupling). The position graph and the head graph are independent design spaces.

Why this matters for long-context

MoBA's block-sparse gating already reduces position-level compute. But standard multi-head attention still treats heads as independent — each head sees only its own features. C5 coupling creates an information highway between heads with zero additional parameters:

Metric Standard 5-head C5-coupled 5-head
Max eigenvalue 1.000 φ = 1.618
Spectral gap 0.0034 0.0707 (20×)
Mixing steps to steady state 25 11 (2.3× faster)
DOF (effective params) 5.0 4.50
DC concentration 0.198 0.439 (2.2×)
Retention @2048 0.606 0.622

The DC concentration result is the key one for long-context: more signal energy in the DC component means slower decay, which means better information retention over long distances.

The pharmacological chain (all code-verified)

C5 adjacency → DFT spectral filter (λ₀=φ amplifies DC, λ₂=0.5 compresses harmonics)
  → DC concentration (φ5=0.439 vs std5=0.198)
    → Redundancy (DOF 5→2.2, 56% reduction)
      → Slower decay (α=0.000216 < 0.000249)
        → Better long-range retention (ret@2048=0.622 > 0.606)

All 7 causal links pass, all experiments zero-fitting (no free parameters tuned to data).

Concrete combination scenario

MoBA: select relevant KV blocks → sparse position-level attention
  ↓ (output per head: mixed position information)
φ-Attention: C5 coupling between heads → mixed head information
  ↓ (output: both position-mixed AND head-mixed)
  → Potentially: fewer heads needed for same capacity, or same heads with better long-range signal

The combination is architecturally simple: after MoBA's block-sparse attention, apply the C5 mixing matrix across heads. No new parameters, no new learnable components.

Full code and detailed results

Repo: https://github.com/wangjun112233/phi-attention

The README follows a five-step "alchemical" framework (furnace → pharmacology → fire → refinement → mechanism) with all experiments reproducible.

Open question

Has the MoBA team explored any form of head-level structure? The current design treats heads independently — if there's interest in testing C5 coupling on top of MoBA's block-sparse attention, we'd be happy to collaborate on benchmarks.


Note: I'm the author of φ-Attention, an independent researcher. This is a genuine research observation about complementarity, not a product pitch.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions