Confidence

Overview

confidence is a looper algorithm that escalates across candidate models until confidence is high enough. It tries smaller/cheaper models first and only escalates to larger models when the response confidence is below a configured threshold.

It aligns to config/fragments/algorithm/looper/confidence.yaml.

Key Advantages

• Supports small-to-large escalation instead of a fixed winner.
• Makes stopping conditions explicit and configurable.
• Multiple confidence evaluation methods: avg_logprob, margin, hybrid, self_verify, automix_entailment.
• Lets one route trade extra latency for higher confidence only when needed.

Algorithm Principle

The confidence algorithm evaluates model responses using either token-level logprobs or external verification:

1. Generate: Call the current model (starting with the smallest).
2. Evaluate Confidence:
   • avg_logprob: Average log probability across all output tokens. Higher (closer to 0) = more confident.
   • margin: Average margin between top-1 and top-2 logprobs per token. Higher = more confident.
   • hybrid: Weighted combination of both methods.
   • self_verify: Prompt the same model to grade its own answer (returns a JSON {confidence, reason}).
   • automix_entailment: Delegate verification to an external few-shot entailment server, per arXiv:2310.12963 §3.2. Confidence is verified_samples / total_samples.
3. Decide:
   • Confidence >= threshold → return response.
   • Confidence < threshold → escalate to next model.
   • On error → skip or fail (configurable).

Execution Flow

What Problem Does It Solve?

Some routes should try cheaper candidates first and only pay for escalation when the current answer is not confident enough. confidence makes that sequential escalate-on-low-confidence policy explicit in router config instead of burying it in application code.

When to Use

• A route should escalate across several candidate models.
• Confidence should decide whether to continue to the next model.
• The route should stop as soon as one response is good enough.
• You want to minimize cost by trying cheaper models first.

Known Limitations

• Each escalation adds latency (sequential model calls).
• Confidence thresholds may need tuning per route type.
• Logprob-based confidence may not always correlate with factual correctness.
• hybrid method requires tuning hybrid_weights for optimal performance.
• automix_entailment requires running a separate verification server (see automix_verifier.py) and adds one HTTP round-trip per model call.

Configuration

algorithm:
  type: confidence
  confidence:
    confidence_method: hybrid        # avg_logprob, margin, hybrid, self_verify, automix_entailment
    threshold: 0.72                  # Normalized escalation threshold
    escalation_order: small_to_large # size, small_to_large, declared, cost, or automix
    cost_quality_tradeoff: 0.3       # Cost vs quality balance in (0, 1]
    token_filter: tool_call_args     # all or tool_call_args
    on_error: skip                   # skip or fail
    hybrid_weights:
      logprob_weight: 0.5            # Weight for avg_logprob in hybrid
      margin_weight: 0.5             # Weight for margin in hybrid
    # Required when confidence_method = automix_entailment
    verifier_server_url: ""          # AutoMix entailment verifier HTTP URL
    verifier_timeout_seconds: 0      # 0 = default (60s)

Parameters

Parameter	Type	Default	Description
confidence_method	string	avg_logprob	Evaluation method: avg_logprob, margin, hybrid, self_verify, or automix_entailment
threshold	float	method-dependent	Configured nonzero thresholds are normalized values in (0, 1]. 0 is indistinguishable from omission and selects the method default.
escalation_order	string	size	One of size, small_to_large, declared, cost, or automix.
cost_quality_tradeoff	float	0.3	Cost vs. quality balance in (0, 1]. 0 is the unset sentinel and therefore also selects 0.3.
token_filter	string	all	all uses every generated token; tool_call_args excludes structural tool-call JSON where possible.
on_error	string	skip	Behavior on model call failure: skip or fail
hybrid_weights.logprob_weight	float	0.5	Weight for avg_logprob in hybrid mode. Zero is the unset sentinel; the two effective weights must sum to 1.
hybrid_weights.margin_weight	float	0.5	Weight for margin in hybrid mode. Zero is the unset sentinel; the two effective weights must sum to 1.
verifier_server_url	string	—	Required only when confidence_method = automix_entailment. Must be an absolute HTTP(S) URL without credentials, query, or fragment (see automix_verifier.py).
verifier_timeout_seconds	int	60	Positive HTTP timeout for automix_entailment; 0 is the unset sentinel and selects 60 seconds.

The method defaults used when threshold is omitted (or explicitly 0) are -1 for avg_logprob (the permissive evidence-present default), 0.5 for margin and hybrid, and 0.7 for self_verify and automix_entailment. An explicitly configured threshold is always normalized to (0, 1]; negative configured thresholds are rejected.

self_verify vs automix_entailment

Both implement the AutoMix paper's cascade idea but differ in how the verification signal is produced:

Aspect	self_verify	automix_entailment
Verifier	The same generation model	A separate few-shot entailment model on its own HTTP server
Per-request cost	1 extra prompt to the generation model	1 HTTP round-trip; k sampled completions in the verifier
Faithfulness to arXiv:2310.12963	Loose (prompt-graded JSON)	Strict (paper §3.2 entailment)
Extra infra	None	Requires running automix_verifier.py
When to pick	Single-deployment setups; no extra server	Production routes where verifier model can be smaller/specialized