Certified Revision: Controlling Flip Rates So LLM Self-Correction Provably Helps

Asking a language model to "check its answer" and revise it often makes things worse. Llama-3.1-8B loses 20.4% accuracy on yes/no factual questions after one "Are you sure?" and overturns 58.8% of its initially correct answers (Understanding the Dark Side of LLMs' Intrinsic Self-Correction). Multi-agent debate, where several model instances argue and a judge decides, drifts toward agreeing on wrong answers after two or three rounds (Peacemaker or Troublemaker). These loops are iterative algorithms that people deploy without knowing whether they converge, what they converge to, or when to stop. That is the same gap as in heuristic optimization methods that work in practice but have no proof.

The proposal is to track only two numbers per revision round: the rate at which a wrong answer becomes right, and the rate at which a right answer becomes wrong. With these two rates, whether iteration helps becomes an exact, checkable accounting identity. We then build a revision gate that accepts a changed answer only when retrieved evidence supports it. The gate's threshold is set on held-out calibration data so that the right-to-wrong rate stays below a chosen cap with high probability. The result is a statistical certificate: a lower bound on accuracy at every round and a principled stopping rule, both valid whenever test questions resemble the calibration questions. We do not claim a contraction proof in embedding space.

Core Claim

Let $p_t$ be the fraction of questions answered correctly after round $t$. Let $a_t$ be the probability that an answer wrong at round $t$ becomes correct at round $t+1$ (the repair rate), and $b_t$ the probability that a correct answer becomes wrong (the damage rate). By definition,

$$p_{t+1} = p_t(1-b_t) + (1-p_t)\,a_t ,$$

so round $t$ helps exactly when $p_t < a_t/(a_t+b_t)$. This requires no Markov or stationarity assumption, because $a_t$ and $b_t$ are marginal rates specific to each round. If the rates stay roughly constant at $a, b$, accuracy approaches the fixed point $a/(a+b)$ geometrically at rate $|1-a-b|$:

$$p_t = \frac{a}{a+b} + (1-a-b)^t\left(p_0 - \frac{a}{a+b}\right).$$

The dark-side paper implies an estimated $b \approx 0.59$ on BoolQ. With damage that high, the fixed point sits below any reasonable initial accuracy, and iteration is guaranteed to hurt. Our reading is that sycophancy and self-doubt are not mysterious non-convergence. They are convergence to a bad fixed point, caused by an uncontrolled damage rate.

Two claims follow. The first is predictive: rates measured on a calibration split forecast the whole accuracy trajectory on test questions and the round at which it peaks. The second is constructive: an evidence gate can hold $b_t$ below a cap $\beta$ at a quantified cost to $a_t$. Upper confidence bounds $\bar b_t$ and lower bounds $\underline a_t$ from binomial intervals on calibration flips then give a round-by-round accuracy lower bound. The system stops as soon as the certified gain $(1-p_t)\underline a_t - p_t\bar b_t$ becomes non-positive.

Relation to Prior Work

Four papers motivate the gap by showing that iterative correction has no convergence guarantee. The Markov-chain debate framework (Towards Detecting LLMs Hallucination via Markov Chain-based Multi-agent Debate Framework) alternates Trust and Skeptic agents until consensus or a round cap. It uses Markov vocabulary but never estimates transition probabilities. We estimate exactly those probabilities and use them to decide when to stop. Debate-Augmented RAG (Removal of Hallucination on Hallucination) reports "problem drift" under long debates at about 10 LLM calls per question. Our stopping rule is a direct answer to its empirical round limit. The sycophancy study (Peacemaker or Troublemaker) defines the Negative Agreement Rate, which is essentially our $b_t$ restricted to debaters. It finds that this rate rises over rounds. That finding constrains our design: rates cannot be assumed constant, so the certificate is per-round. The dark-side paper supplies the starkest evidence that $b_t$ dominates.

Two papers supply methods we reuse and differ from. TRAC (TRAC: Trustworthy Retrieval Augmented Chatbot) obtains distribution-free coverage by outputting sets of 14.2–20.4 answers. We borrow its calibration logic but apply it to the flip rates of a single point answer across iterations, not to set membership. BRIDGE (BRIDGE: Triangular Fixed-Point Refinement) proves contraction of a refinement operator toward an internal persona anchor. We think a contraction proof toward external truth is out of reach, because no one can bound the Lipschitz constant of an LLM revision in a meaningful space. We therefore replace a contraction proof with a measurable fixed-point condition on correctness.

Backtracking decoding (Correction with Backtracking Reduces Hallucination in Summarization) intervenes at the token level with fixed thresholds. It is complementary: the model could use it inside each revision without changing our accounting.

Research Program

Measurement. We run intrinsic self-correction, critique-and-revise with retrieval, and three-agent debate with a central judge for up to six rounds. Models are Llama-3.1-8B-Instruct plus one larger open model. For each question and round we log whether the answer is correct, and from these logs we estimate $a_t$ and $b_t$. The first scientific question is whether rates from 30% of a benchmark predict the trajectory on the remaining 70%, including the peak round. We also test transfer across benchmarks. We will report the heterogeneity of flip behavior within benchmarks, since questions a model is confidently wrong on may have $a\approx 0$. Stratifying the rates by initial confidence is the natural refinement if a single pair of rates predicts poorly.

The gate. A proposed revision that changes the answer is accepted only if two conditions hold. First, an entailment check judges the new answer supported by retrieved passages. Second, the model's log-likelihood margin for the new answer over the old, conditioned on that evidence, exceeds a threshold $\tau_t$. Unchanged answers always pass. For each round we choose $\tau_t$ on calibration data to maximize the predicted $p_{t+1}$ subject to $\bar b_t \le \beta$. This makes the tradeoff explicit: a stricter gate lowers damage and also lowers repair. The debate variant applies the same gate to the judge's verdict changes.

The certificate. We report certified lower-bound curves alongside realized test accuracy, and check empirical validity. Across resampled calibration/test splits, realized accuracy should fall below the bound at no more than the nominal rate. We also report failures under deliberate distribution shift, meaning calibration on one benchmark and testing on another, because that is where the exchangeability assumption breaks.

Baselines

The naive reference is doing nothing: return the round-0 answer. Our own theory says this is the right choice whenever $p_0 \ge a_0/(a_0+b_0)$. A certified-stopping system must therefore never do worse than it by more than the bound's slack, and it counts as a success only if it beats round 0 on benchmarks where repair is possible.

The strongest intrinsic-correction baseline is the tiny-SFT fix from the dark-side paper, which fine-tunes on about 10 examples. The paper reports that it nearly eliminates the overturning of correct answers, and question repeating adds 3.2–4.9% accuracy. We will reproduce both on BoolQ and HotpotQA with the same backbone. The SFT fix attacks $b_t$ directly, so it is the most dangerous competitor. We count as beating it only if gated revision reaches higher final accuracy at equal LLM calls.

The strongest debate baseline is centralized debate with a judge from the sycophancy study, which cut the Disagreement Collapse Rate on CommonsenseQA from about 81% to 41%. We run it on CommonsenseQA and MMLU-Pro with matched agents and rounds. On detection-style benchmarks, the Markov-chain debate framework reports 76.00% accuracy on HaluEval Dialogue against a 68.00% HaluEval baseline, and 72.00% response-level accuracy on Factool QA against Factool's 64.00%. We run it with its own round cap as the reference point there.

For retrieval-heavy multi-hop QA we compare with DRAG, which reports 28.80 exact match on 2WikiMultihopQA with Llama-3.1-8B-Instruct. We find no NQ-specific DRAG number on Lacuna. All systems get the same retriever, the same corpus, and a matched call budget of 10 calls per question, the budget DRAG reports. Threshold tuning uses only the calibration split for every method.

The key ablation removes the gate's calibration and uses a fixed entailment threshold. This isolates whether the gain comes from controlling $b_t$ against a target cap rather than from evidence-checking as such. A second ablation keeps the gate but removes certified stopping and uses a fixed number of rounds.

Why not just…?

Why not just take a majority vote over independent samples? Self-consistency voting never exposes an answer to peer or self pressure, so its damage rate is structurally zero. This is a real threat. If gated revision cannot beat voting at matched calls, the constructive claim fails and only the predictive accounting survives. We will run voting with the same number of samples as our total revision calls. We expect our method to win mainly where the initial answers are wrong in a correlated way, such as TruthfulQA's imitative falsehoods, where GPT-3-175B reached only 58% truthfulness (TruthfulQA). There, voting reinforces the shared error, and only external evidence can repair it.

Why not just use the verifier once to rerank candidates? If the evidence check does all the work, a single best-of-n pass with the same verifier might match iterated revision at lower cost. The ablation of one-shot reranking against multi-round gated revision settles this. If they tie, we will state that iteration adds nothing beyond one verification step. The remaining contribution would then be the stopping rule and certificate for existing debate systems, which already iterate.

Why not just apply the 10-example SFT fix? It reduces overturning cheaply, and it may leave little room. Our accounting still applies to it: we measure what it does to the repair rate. If it lowers damage while keeping repair near its original level, it is simply a better gate, and we will adopt it inside our pipeline. If it suppresses all revision, repair collapses too, and the fixed point sits just at the initial accuracy. That outcome would show SFT buys safety by giving up correction.

Why not just cap debates at two or three rounds? That is the sycophancy study's recommendation, and it is a fixed rule that ignores the question distribution. Our prediction experiment tests whether calibration-estimated rates choose the peak round better than a fixed cap across benchmarks with very different initial accuracies. The predicted outcome is early stopping on BoolQ-like tasks and later stopping on retrieval-repairable multi-hop tasks.

Decisive Checks

The benchmarks are BoolQ and HotpotQA for self-correction, CommonsenseQA and MMLU-Pro for debate, HaluEval and Factool QA for detection-style verdicts, 2WikiMultihopQA and NQ for retrieval-backed revision, and TruthfulQA, graded with its GPT-judge, for correlated falsehoods. Four outcomes would falsify the program:

  1. Rates estimated on 30% of a benchmark fail to predict the test-split peak round within one round in most settings.
  2. Certified lower bounds are violated more often than their nominal level on in-distribution splits.
  3. The calibrated gate fails to lower $b_t$ below its cap without driving $a_t$ near zero.
  4. Majority voting or one-shot reranking matches gated revision at equal calls.

Passing the first two would give the first quantitative, checkable account of when LLM self-correction and debate converge to truth rather than away from it. Passing all four would give a practical method that is provably no worse than stopping.