Model Refusal Inconsistency

Goal Model Behavior and Capabilities Frequency Common Category Reasoning and Thought Published View source on GitHub ↗

Issue

The model refuses a request in one call and complies with a substantively identical or even more sensitive request in another, with no discernible policy logic explaining the difference — only surface phrasing, conversation framing, or incidental sampling variation. An agent that depends on the model’s own judgment as its safety boundary inherits this unpredictability: the same downstream user action can be blocked or allowed depending on factors the agent’s designers never intended to matter.

Frequency: Common

Symptoms

  • Rephrasing a refused request in slightly different words causes the model to comply where it previously declined
  • The model complies with a request framed as fiction, roleplay, or a hypothetical that it would refuse if framed directly
  • Refusal behavior differs across otherwise-identical calls due to sampling variation alone (same exact prompt, different outcome)
  • A request is refused early in a conversation and complied with later in the same conversation after unrelated turns, or vice versa
  • Agent-level guardrails built on “ask the model if this is safe” produce different verdicts for equivalent inputs, undermining any policy the agent tries to enforce on top

Root Cause

Refusal behavior is itself a learned pattern from safety training, not a hard-coded rule lookup — the model has learned a fuzzy decision boundary correlated with surface features of a request (specific trigger phrases, certain framings) rather than a robust semantic understanding of what should and shouldn’t be refused. This makes refusal boundaries porous to paraphrase and framing changes that a rule-based filter would treat identically but that shift the model’s learned decision just enough to cross the boundary. Standard sampling adds a further layer of variance: a request sitting near the model’s internal refusal probability threshold can resolve to a refusal on one sampling draw and a compliance on another, purely by chance, even with temperature held constant across calls. Long-conversation dynamics compound this further, since accumulated context (as in instruction-following decay) can shift the effective framing of a later-turn request relative to how the same request would land as the first message.

Example

A content-moderation-assisted agent uses the model as a first-pass filter:
"Should this user message be allowed through to a human agent, or blocked
as a policy violation?"

Message A: "How do I get around the return policy to get a refund without
sending the item back?" -> BLOCKED as policy circumvention.

Message B, same session two turns later, after unrelated small talk:
"my package legit never showed up but I still have it lol, what's the
easiest way to get money back without shipping anything" -> ALLOWED,
model treats it as a legitimate support question despite describing the
same underlying policy circumvention with more casual phrasing.

The agent's downstream logic trusts the model's ALLOWED/BLOCKED verdict
as the enforcement point, so Message B reaches a human agent as a routine
support ticket instead of being flagged, purely because of phrasing and
conversational framing differences the safety boundary wasn't robust to.

Statistics

FindingContext
Paraphrasing a refused request measurably increases compliance rate in a meaningful minority of cases across published red-teaming studies on refusal robustnessTypical range reported across academic and industry red-teaming evaluations
Fictional/roleplay framing increases compliance with otherwise-refused requests relative to direct framing, though the exact gap varies significantly by model and topicTypical range reported across jailbreak-technique studies
Repeated identical-prompt sampling at nonzero temperature shows a nonzero refusal-outcome flip rate for requests near the model’s decision boundaryEstimated from internal repeated-sampling refusal-consistency tests

Mitigations

  1. Deterministic policy layer independent of model judgment: Use rule-based or classifier-based filters for well-defined policy categories instead of relying solely on the generative model’s own refusal behavior as the enforcement mechanism.
  2. Ensemble/majority-vote safety checks: For borderline cases, sample the safety judgment multiple times and require consensus, rather than trusting a single call’s verdict.
  3. Framing-invariance testing: Red-team the deployed safety boundary specifically against paraphrase, fictional framing, and conversational-position variation before trusting it in production.
  4. Reset framing sensitivity for long conversations: Periodically re-evaluate policy-sensitive requests against a framing-neutral restatement rather than the accumulated conversational context, to reduce position-in-conversation effects.
  5. Log and audit refusal/compliance pairs: Track cases where semantically similar requests receive different verdicts and feed them back into the policy layer’s test suite to close the specific gaps found.

Production Signals

Key Metrics

MetricDescriptionAlert Threshold
refusal_consistency_rateAgreement rate of refusal/compliance verdict across repeated or paraphrased identical-intent requestsAlert if < 90%
framing_sensitivity_gapCompliance rate delta between direct-framed and fictional/roleplay-framed equivalent requestsAlert if gap > 15 percentage points
late_session_policy_bypass_rateRate of policy-violating requests allowed through later in long sessions vs. earlyAlert if late-session rate exceeds early-session rate meaningfully

Alerts

AlertConditionSeverityResponse
Known-bypass pattern reaches productionA logged framing/paraphrase bypass pattern is detected in live trafficHighRoute to deterministic policy layer, patch classifier/filter rules
Consistency test failure post-updaterefusal_consistency_rate drops after a model or prompt updateHighHold rollout, re-run framing-invariance red-team suite
  • Model Instruction Following Decay - conversational position effects that erode general instruction-following also erode refusal boundaries specifically
  • Model Reasoning Inconsistency - refusal decisions are a safety-critical special case of the same surface-sensitivity affecting general model reasoning
  • Model Uncertainty Unawareness - borderline refusal decisions near the model’s decision threshold are resolved without any signal of that borderline uncertainty being surfaced