Squeeze Evolve: Confidence-Gated Routing & Evolutionary Aggregation

Concepts: CONCEPT:AU-ORCH.adapter.hot-cache-invalidation (Confidence-Gated Router), CONCEPT:AU-ORCH.adapter.hot-cache-invalidation (Evolutionary Aggregation Engine)

Based on: SQUEEZE EVOLVE: Adaptive Multi-Model Orchestration for Mathematical Reasoning (Maheswaran et al., 2026) — arXiv:2604.07725v2

Overview

Agent-utilities integrates two key innovations from the Squeeze Evolve paper to optimise model dispatch efficiency and reasoning quality:

1. CONCEPT:AU-ORCH.adapter.hot-cache-invalidation — Confidence-Gated Router: Adaptively selects cheaper or more expensive models based on runtime confidence signals from specialist consensus and historical performance.

2. CONCEPT:AU-ORCH.adapter.hot-cache-invalidation — Evolutionary Aggregation Engine: Groups specialist outputs, computes group-level fitness (confidence + diversity), and routes aggregation to the most cost-effective strategy.

Core Principle

Allocate model capability where it has the highest marginal utility.

Strong models should initialise solutions (where reasoning depth matters most), while cheaper models can handle aggregation when specialists already agree. This achieves significant cost reduction with equivalent or improved accuracy.

---

CONCEPT:AU-ORCH.adapter.hot-cache-invalidation: Confidence-Gated Model Router

How It Works

flowchart LR
    subgraph Signals
        WA["WorkspaceAttention<br/>CONCEPT:AU-ORCH.adapter.hot-cache-invalidation"]
        SM["SelfModel<br/>CONCEPT:AU-KG.memory.tiered-memory-caching"]
    end

    subgraph Routing
        WA -->|runtime score| BLEND["Confidence Blend<br/>70% runtime + 30% historical"]
        SM -->|proficiency| BLEND
        BLEND --> GATE{"confidence vs<br/>threshold"}
    end

    subgraph Dispatch
        GATE -->|high confidence| DOWN["Tier Down<br/>cheaper model"]
        GATE -->|neutral| KEEP["Keep Tier<br/>same model"]
        GATE -->|low confidence| UP["Tier Up<br/>stronger model"]
    end

    DOWN --> REG["ModelRegistry<br/>pick_for_task"]
    KEEP --> REG
    UP --> REG

Composition with CONCEPT:AU-OS.state.cognitive-scheduler-preemption (Homeostatic Downgrade)

The confidence-gated router composes with the existing budget-pressure downgrade (CONCEPT:AU-OS.state.cognitive-scheduler-preemption):

1. CONCEPT:AU-OS.state.cognitive-scheduler-preemption runs first: If the ResourceOptimizer detects budget pressure, it reduces the tier (e.g., heavy → medium).
2. CONCEPT:AU-ORCH.adapter.hot-cache-invalidation runs second: The confidence signal further adjusts within the budget-allowed range (e.g., medium → light if confidence is high).

This means budget pressure always takes precedence, but confidence can further optimise within the budget constraint.

Self-Model Integration (Soft Dependency)

CONCEPT:AU-ORCH.adapter.hot-cache-invalidation optionally uses historical proficiency from CONCEPT:AU-KG.memory.tiered-memory-caching (Persistent Self-Model):

# When SelfModel is available:
historical_confidence = self_model.tool_proficiency.get(specialist_id, 0.5)
effective = 0.7 * runtime_confidence + 0.3 * historical_confidence

# When no SelfModel (graceful degradation):
historical_confidence = 0.5  # Neutral — no bias
effective = 0.7 * runtime_confidence + 0.15  # Pure runtime signal

Configuration

Parameter	Location	Default	Description
routing_percentile	GraphDeps	50.0	Threshold (0–100). Lower = more aggressive downgrade.
ROUTING_PERCENTILE	Env var	50.0	Default for routing_percentile.

API Reference

from agent_utilities.models.model_registry import ModelRegistry

registry = ModelRegistry(models=[...])
chosen = registry.pick_for_task_adaptive(
    complexity="medium",
    confidence_signal=0.85,     # High confidence → downgrade
    routing_percentile=30.0,    # Aggressive: route 70% to cheap models
    required_tags=["code"],
)

---

CONCEPT:AU-ORCH.adapter.hot-cache-invalidation: Evolutionary Aggregation Engine

Group Fitness Signals

The engine computes two signals for each group of specialist proposals:

Signal	Formula	Description
Group Confidence (GC)	GC(g) = (1/K) × Σ C(τ)	Mean confidence across proposals
Group Diversity (D)	D(g) = \|{answer(τ) : τ ∈ g}\|	Unique answer cluster count

Three-Tier Aggregation

flowchart TB
    subgraph Input
        P1[ORCH-1.2: Proposal 1]
        P2[ORCH-1.2: Proposal 2]
        P3[ORCH-1.2: Proposal 3]
        P4[ORCH-1.2: Proposal 4]
    end

    P1 & P2 --> G1[ORCH-1.21: Group 1]
    P3 & P4 --> G2[ORCH-1.21: Group 2]

    G1 --> F1{"GC=0.9<br/>D=1"}
    G2 --> F2{"GC=0.3<br/>D=3"}

    F1 -->|"High GC + Low D"| MV["🆓 MAJORITY_VOTE<br/>No LLM call"]
    F2 -->|"Low GC + High D"| HM["🔴 HEAVY_MODEL<br/>Reasoning-tier aggregation"]

    style MV fill:#d4edda,stroke:#28a745
    style HM fill:#f8d7da,stroke:#dc3545

Strategy	When	Cost
MAJORITY_VOTE	High GC + Low D (specialists agree)	Free — no LLM call
LIGHT_MODEL	High GC + Some D (partial agreement)	Cheap — fast model synthesis
HEAVY_MODEL	Low GC + High D (specialists disagree)	Expensive — reasoning model

Convergence Monitor

The ConvergenceMonitor tracks diversity across evolutionary iterations and signals early stopping when diversity collapses:

from agent_utilities.graph.hierarchical_planner import ConvergenceMonitor

monitor = ConvergenceMonitor(convergence_threshold=0.1, patience=3)

for iteration in evolutionary_loop:
    diversity = compute_diversity(proposals)
    if monitor.update(diversity):
        break  # Diversity collapsed — stop early

Configuration

Parameter	Default	Description
population_size	4	Max proposals to consider (N)
group_size	2	Proposals per fitness group (K)
confidence_threshold	0.7	GC above this → lite aggregation
diversity_threshold	2	D above this + low GC → heavy aggregation
convergence_threshold	0.1	Diversity below this = converged
patience	3	Consecutive low-diversity readings before stopping

---

Architecture Integration

graph LR
    AU016["CONCEPT:AU-KG.memory.tiered-memory-caching<br/>Self-Model"] -->|historical proficiency| AU039
    AU017["CONCEPT:AU-ORCH.adapter.hot-cache-invalidation<br/>Workspace Attention"] -->|confidence scores| AU039["CONCEPT:AU-ORCH.adapter.hot-cache-invalidation<br/>Confidence Router"]
    AU017 -->|proposals| AU040["CONCEPT:AU-ORCH.adapter.hot-cache-invalidation<br/>Evolutionary Aggregation"]
    AU033["CONCEPT:AU-OS.state.cognitive-scheduler-preemption<br/>Homeostatic Downgrade"] -->|budget tier| AU039
    AU030["CONCEPT:AU-OS.state.cognitive-scheduler-preemption<br/>Cognitive Scheduler"] -->|hosts| CM["Convergence<br/>Monitor"]
    CM --> AU040
    AU039 -->|adaptive model| REG[AU-ECO.mcp.toolkit-live-discovery: Model Registry]

    style AU039 fill:#e6f2ff,stroke:#0066cc,stroke-width:2px
    style AU040 fill:#e6f2ff,stroke:#0066cc,stroke-width:2px

Related Concepts

Concept	Relationship
CONCEPT:AU-KG.memory.tiered-memory-caching Self-Model	Provides historical confidence priors
CONCEPT:AU-ORCH.adapter.hot-cache-invalidation Workspace Attention	Provides runtime scoring and proposals
CONCEPT:AU-OS.state.cognitive-scheduler-preemption Resource Optimization	Budget-aware model selection
CONCEPT:AU-OS.state.cognitive-scheduler-preemption Cognitive Scheduler	Hosts convergence monitor
CONCEPT:AU-OS.state.cognitive-scheduler-preemption Homeostatic Downgrade	Composes with confidence routing