Model Reasoning Inconsistency

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

Issue

The model produces different reasoning chains and different final conclusions when given logically identical inputs that differ only in superficial ways — order of options, phrasing, irrelevant surrounding text, or which call happens to sample a different token early in the chain of thought. An agent that relies on the model’s reasoning to make a consistent decision (approve/deny, rank A over B, classify as X) gets a decision that isn’t actually a function of the underlying facts, just of incidental surface variation.

Frequency: Common

Symptoms

  • Re-running the exact same decision prompt (same facts, reordered options) flips the conclusion a nonzero fraction of the time
  • The model gives contradictory answers when the same question is asked in two logically equivalent but differently-worded ways within the same session
  • Chain-of-thought traces for the same underlying problem take visibly different reasoning paths across calls, sometimes reaching different conclusions
  • Position/order bias: presenting option A before option B versus B before A changes which one the model prefers, independent of content
  • Downstream aggregate metrics (e.g. approval rate) shift measurably when only cosmetic prompt formatting changes, with no change to the underlying policy or facts

Root Cause

Autoregressive generation with sampling means the model’s reasoning is not a fixed computation over the input’s logical content — it is a probabilistic walk where each generated token depends on the exact preceding token sequence, so any surface-level difference (option order, phrasing, whitespace) changes the token sequence and can cascade into a different reasoning path. Even at low temperature, the model’s training didn’t optimize for invariance to logically-irrelevant surface variation; it optimized for plausible next-token prediction given the specific text seen, and many training examples do correlate position or phrasing with outcome (e.g. first-listed options being preferred in typical documents), so the model has learned superficial correlations that have nothing to do with the task’s actual logic. Chain-of-thought reasoning amplifies this: once the model commits to an early reasoning step that happens to differ across runs, subsequent steps condition on that divergent path, compounding a small initial difference into a materially different conclusion.

Example

A vendor-risk agent evaluates whether a supplier contract clause is
"acceptable" or "requires legal review," given identical clause text,
run twice with only the order of two reference example clauses in the
few-shot prompt swapped.

Run A (example order: lenient clause first, strict clause second):
"This clause is acceptable; the indemnification cap is standard for
this contract size." -> classified ACCEPTABLE

Run B (example order: strict clause first, lenient clause second):
"This clause's indemnification cap is below typical protection levels
and should be flagged." -> classified REQUIRES REVIEW

The underlying clause text was byte-for-byte identical in both runs;
only the order of unrelated few-shot examples changed, yet the agent's
downstream action (auto-approve vs. route to legal) flipped.

Statistics

FindingContext
Order-swapping options or few-shot examples in decision prompts typically flips the model’s conclusion in an estimated 5-20% of cases, depending on how close the underlying decision is to the model’s uncertainty boundaryTypical range observed in prompt-order-sensitivity evaluations
Rephrasing a logically equivalent question produces a different final answer in an estimated 10-15% of borderline cases, versus under 2% for clear-cut casesEstimated from paraphrase-consistency evaluations on classification tasks
Aggregating multiple independent reasoning samples and taking a majority vote (self-consistency) typically improves decision stability by a meaningful margin over single-sample reasoning, at the cost of multiple model callsTypical range reported across self-consistency technique evaluations

Mitigations

  1. Self-consistency sampling: For high-stakes decisions, run the same prompt multiple times (varying only sampling, not content) and take a majority vote or flag disagreement for human review, rather than trusting a single reasoning pass.
  2. Order-randomization testing: Before deploying a decision prompt, test it with option/example order systematically varied to measure and bound position-sensitivity before it reaches production.
  3. Canonicalize input presentation: Normalize option ordering (e.g. always alphabetical or by a fixed neutral rule) rather than passing through whatever incidental order the upstream system produced, removing one axis of irrelevant variation.
  4. Lower-temperature or deterministic decoding for decision-critical calls: Reduce (not eliminate, but reduce) run-to-run variance for calls where consistency matters more than creative diversity.
  5. Consistency monitoring in production: Periodically re-submit sampled historical decisions and compare current output to the original, tracking a consistency-rate metric over time rather than assuming it stays constant across model or prompt changes.

Production Signals

Key Metrics

MetricDescriptionAlert Threshold
decision_flip_rate_on_reorderRate at which identical-content, reordered prompts produce a different conclusionAlert if > 10% for high-stakes decision types
self_consistency_agreement_rateAgreement rate across multiple independent reasoning samples of the same promptAlert if < 80%
aggregate_outcome_drift_on_cosmetic_changeShift in aggregate decision-rate metrics attributable to formatting-only prompt changesAlert if drift exceeds 5 percentage points

Alerts

AlertConditionSeverityResponse
High flip rate on sampled re-testdecision_flip_rate_on_reorder exceeds threshold for a decision pipelineHighRoute affected decision type through self-consistency voting, add to human review queue
Consistency regression after changeself_consistency_agreement_rate drops after a prompt or model version changeMediumRoll back or investigate change, re-run consistency evaluation suite
  • Model Output Format Instability - both are forms of run-to-run nondeterminism, one in structure and one in substantive conclusions
  • Model Capacity Limits - reasoning inconsistency is more pronounced on complex, near-boundary decisions where capacity is already strained
  • Model Fairness Bias - a systematic, demographically-correlated subtype of the same underlying sensitivity to surface-level input variation