Correct-Detect Trade-Off

• CORRECT–DETECT trade-off is a design challenge where systems balance accurate task performance with detecting uncertainty and errors, using metrics like EIR and ECR.
• The quantitative formulation employs a closed-loop control model and Markov chain analysis to determine when iterative self-correction benefits overall accuracy.
• Empirical studies across LLM self-correction, coreference resolution, and robotic control show that maintaining near-zero error introduction rates is key to preventing degradation of correct outputs.

The CORRECT–DETECT trade-off denotes a recurring design tension in which a system must balance a capability to act correctly with a capability to detect uncertainty, error, drift, or anomaly without introducing new failure modes or excessive cost. In its most explicit recent formulation for iterative LLM self-correction, the trade-off is between how often a model corrects previous errors and how often it detects something and changes a previously correct answer into an error; closely related formulations appear in coreference ambiguity handling, action-chunked robotic control, blind code detection, watermarking, compression-based anomaly detection, and other detection-limited systems (Liu et al., 24 Apr 2026).

1. Control-theoretic formulation in iterative LLM self-correction

In the self-correction setting, iterative refinement is framed as a closed-loop control system in which the same LLM serves as both controller and plant. For each problem $q_i$ with ground-truth answer $a_i^*$, the baseline and refined responses are

$$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$

Correctness is represented by a two-state Markov chain over $\{\text{Correct}, \text{Incorrect}\}$, with indicator

$$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$

and batch accuracy

$\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$

The central transition rates are the Error Introduction Rate

$$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$

and the Error Correction Rate

$$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$

These define the transition matrix

$$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$

This formulation makes the CORRECT–DETECT trade-off quantitative. ECR measures successful detect-and-correct behavior; EIR measures the destabilizing case in which the model “detects something” and edits an answer that was already correct. Accuracy evolves as

$$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$

so the net benefit of one more refinement step is

$a_i^*$0

The resulting deployment diagnostic is

$a_i^*$1

equivalently, iterate only when

$a_i^*$2

In this view, EIR functions as a stability margin and prompting functions as lightweight controller design (Liu et al., 24 Apr 2026).

2. Empirical threshold, instability, and prompt-level interventions

Across 7 models and 3 datasets—GSM8K, MATH, and StrategyQA—with 4 iterations of self-correction per problem, the empirical pattern is that a near-zero EIR threshold separates beneficial from harmful refinement. On GSM8K, the paper reports the following trajectory summaries and iteration-$a_i^*$3 transition rates (Liu et al., 24 Apr 2026).

Model	GSM8K accuracy $a_i^*$4	EIR / ECR $a_i^*$5
GPT-4o-mini	$a_i^*$6 $a_i^*$7	$a_i^*$8
GPT-5	$a_i^*$9 $$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$0	$$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$1
Claude Sonnet 4	$$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$2 $$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$3	$$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$4
Claude Opus 4.6	$$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$5 $$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$6	$$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$7
o3-mini	$$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$8 $$r_i^{(0)} = \mathcal{M}(q_i), \qquad r_i^{(k)} = \mathcal{M}(q_i,\, r_i^{(k-1)},\, p_{\text{refine}}).$$9	$\{\text{Correct}, \text{Incorrect}\}$0

The practical threshold is

$\{\text{Correct}, \text{Incorrect}\}$1

Models that benefit or at least do not degrade—o3-mini, Claude Opus 4.6, and o4-mini—have near-zero EIR; models that degrade have EIR in the $\{\text{Correct}, \text{Incorrect}\}$2 to $\{\text{Correct}, \text{Incorrect}\}$3 range. The mechanism is a pool-size asymmetry: for strong models, the correct pool is much larger than the incorrect pool, so even tiny EIR can dominate moderate ECR. This is the paper’s accuracy–correction paradox.

Two interventions make the trade-off actionable. First, a verify-first prompt instructs the model to solve the problem again from scratch, compare against the previous answer, and only change the answer if it identifies a clear error. On GPT-4o-mini for GSM8K, verify-first changes accuracy from $\{\text{Correct}, \text{Incorrect}\}$4 under standard refinement to $\{\text{Correct}, \text{Incorrect}\}$5, reduces average EIR from about $\{\text{Correct}, \text{Incorrect}\}$6 to $\{\text{Correct}, \text{Incorrect}\}$7, and yields a paired McNemar result of $\{\text{Correct}, \text{Incorrect}\}$8 with $\{\text{Correct}, \text{Incorrect}\}$9Acc $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$0 and $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$1 CI $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$2. Second, Adaptive Self-Correction (ASC) halts refinement when confidence exceeds a threshold or when a running estimate satisfies $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$3. On GPT-4o-mini, ASC correctly halts harmful refinement at iteration 0, but the confidence-elicitation step itself reduces baseline accuracy from $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$4 to $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$5, a 3.8 pp confidence-elicitation cost. Together, these results argue that self-correction is a control decision rather than a default behavior (Liu et al., 24 Apr 2026).

3. Coreference resolution and ambiguity detection

In coreference resolution, the term is used more literally as a trade-off between resolving a pronoun when humans see a strong preferred referent and detecting when the pronoun is genuinely ambiguous and should be left unresolved. On the AmbiCoref benchmark, CORRECT is measured by Correct-Unamb on unambiguous items, while DETECT is measured by Detect-Ambig on ambiguous items. The central finding is that LLMs can do well on one behavior or the other, but cannot do both at the same time (Shore et al., 17 Sep 2025).

For GPT-4o, a minimal Basic prompt yields Correct-Unamb = 87.70% but Detect-Ambig = 22.86%. Ambiguity-sensitive prompts move the model in the opposite direction: Ambi-Ask gives $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$6 detection and $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$7 correctness, Ambi-Stop gives $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$8 detection and $$c_i^{(k)} = \mathbb{1}[\text{extract}(r_i^{(k)}) = a_i^*],$$9 correctness, Ambi-Wait gives $\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$0 detection and $\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$1 correctness, and Ambi-CoT gives $\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$2 detection and $\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$3 correctness. For Llama 3.1, Basic gives Correct-Unamb = 90.33% and Detect-Ambig = 3.72%, while Ambi-CoT yields 75.17% detection and 42.90% correctness. Human performance occupies a different regime: Detect = 78.47% and Correct = 76.77%.

The resulting Pareto-like pattern is central. Prompting that emphasizes ambiguity causes over-ambiguity in clear cases; prompting that emphasizes resolution causes under-detection in ambiguous cases. The paper therefore characterizes current LLM behavior as a systematic CORRECT-DETECT trade-off: they can be steered toward high disambiguation or high ambiguity detection, but not toward the human-like upper-right regime that combines both (Shore et al., 17 Sep 2025).

4. Event-triggered correction in embodied control

In action-chunked Vision-Language-Action control, an analogous trade-off arises between how aggressively the system detects drift and how often it corrects by interrupting execution and replanning. A standard action-chunked VLA policy predicts a chunk

$\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$4

but deployment executes only a fixed action horizon

$\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$5

open-loop, with no further policy calls until all $\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$6 queued actions are executed. This creates a robustness–efficiency trade-off: smaller $\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$7 gives more frequent replanning and better closed-loop responsiveness, while larger $\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$8 reduces policy-call frequency but sacrifices reactivity (Pan et al., 2 Jul 2026).

VLA-Corrector decouples monitoring from correction. A Latent-space Vision Monitor (LVM) predicts expected latent visual dynamics with a learned residual model,

$\text{Acc}(k) = \frac{1}{N}\sum_{i=1}^N c_i^{(k)}.$9

compares them to actual latent evolution, and computes a visual-dynamics inconsistency score

$$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$0

A robust sliding-window rule with median, MAD, dual ON/OFF thresholds, and a persistence counter triggers an interrupt when deviation is persistent. The remaining stale actions are discarded, so the realized horizon becomes

$$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$1

A corrective replan is then guided by Online Gradient Guidance (OGG), which aligns the predicted latent effect of the next action with a corrective latent direction.

The system is presented as mitigating the trade-off imposed by static horizons between execution robustness and policy-call frequency. For $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$2 on MetaWorld, the static baseline shows success $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$3, calls $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$4 at $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$5 and success $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$6, calls $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$7 at $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$8. At $$\text{EIR}(k) = P(c_i^{(k+1)} = 0 \mid c_i^{(k)} = 1)$$9, VLA-Corrector yields success $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$0, calls $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$1, corresponding to a success-per-call efficiency gain of +24.6%. In component ablations, the baseline is 48.7%, +Truncation only is 60.35%, and +Truncation + OGG is 64.35%. On real-world disturbance tasks with an AgileX PiPER arm, average success rises from 40.0% to 68.3%. Here the CORRECT–DETECT balance is operationalized as continuous, cheap detection with infrequent, event-triggered correction (Pan et al., 2 Jul 2026).

5. Cross-domain variants and mathematical analogues

The term and its close analogues recur across several research areas, but the objective pair changes with the system under study.

Domain	CORRECT side	DETECT side or limiting factor
Blind detection of polar codes	low MDR, keep decodable block in top-$$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$2	Stage-1 effort $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$3 or $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$4, and candidate pruning
Multi-bit LLM watermarking	high TPR, robust payload decoding	strict FPR control under Strict-Safe or calibrated FPR–TPR under FPR-Calibrated
Compression for anomaly detection	low distortion $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$5, rate $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$6	distinguishability $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$7 or $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$8 after compression
Streaming imbalanced classification	minority detection rate $$\text{ECR}(k) = P(c_i^{(k+1)} = 1 \mid c_i^{(k)} = 0).$$9	limited inspection capacity $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$0
Invariant representation learning	retain dependence on $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$1	suppress detectability of $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$2
Local discrimination of entangled states	success probability $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$3	post-measurement fidelity $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$4

In blind detection of polar-coded control messages, the two-stage detector–decoder pipeline explicitly studies a trade-off between missed-detection rate (MDR) and computational effort in the detection stage. Complexity is controlled by the number of BP iterations $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$5 or the number of Fast-SSC leaf nodes $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$6. At the operating point where CRC-aided SCL with $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$7 has BLER $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$8, the paper shows that, for Method 2, $$T(k) = \begin{pmatrix} 1-\text{EIR}(k) & \text{EIR}(k) \ \text{ECR}(k) & 1-\text{ECR}(k) \end{pmatrix}.$$9 is enough to reach MDR $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$0 while passing no more than $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$1 candidates, and $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$2 allows no more than $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$3 of the candidates to be kept at the same MDR; for the Fast-SSC detector without SPC updates, visiting $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$4 nodes allows 40 of 44 candidates to be dismissed while maintaining MDR $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$5 (Giard et al., 2018).

In multi-bit LLM watermarking, CORE-BREW turns the trade-off into an explicit detector-design choice. A Constant-hit-Rate embedding targets a fixed $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$6, yielding per-token LLRs

$$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$7

with erasures mapped to $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$8. Strict-Safe preserves the bounded-distance designated-codeword acceptance region and the corresponding combinatorial FPR guarantees, while FPR-Calibrated uses score thresholds and lightweight list decoding to characterize the FPR–TPR trade-off. In the clean setting with $$\text{Acc}(k+1)=\text{Acc}(k)(1-\text{EIR}(k)) + (1-\text{Acc}(k))\text{ECR}(k),$$9, CORE-BREW-Cal at $a_i^*$00 has FPR $a_i^*$01–0.002 and TPR $a_i^*$02; under 10% deletion, it achieves TPR $a_i^*$03 with FPR $a_i^*$04 (Kim et al., 23 Jun 2026).

In compression-based anomaly detection, the RDD framework formalizes a three-way rate–distortion–distinguishability trade-off. The rate is

$a_i^*$05

distortion is

$a_i^*$06

and distinguishability is measured by either the anomaly-agnostic metric

$a_i^*$07

or the anomaly-aware Jeffreys divergence

$a_i^*$08

The optimization

$a_i^*$09

shows that rate–distortion optimal compression can be suboptimal for detection because low-variance directions that are dispensable for reconstruction may be crucial for anomaly discrimination (Enttsel et al., 29 Sep 2025).

Further formal analogues reinforce the breadth of the idea. In invariant representation learning, the trade-off functional

$a_i^*$10

balances predictive utility against detectability of a protected or semantic attribute, and the optimal encoders are obtained from a generalized eigenproblem (Sadeghi et al., 2021). In streaming binary classification, the minority detection rate $a_i^*$11 is traded against inspection capacity $a_i^*$12, with an end-to-end upper bound

$a_i^*$13

and dynamic thresholds nearly match the batch-processing upper bound (Hassanzadeh et al., 2021). In local discrimination of entangled states, the combined score

$a_i^*$14

obeys

$a_i^*$15

so local information gain is fundamentally limited by state disturbance, especially for maximally entangled states (Lim et al., 2023).

6. Conceptual synthesis, misconceptions, and limits

Several common themes emerge. First, the trade-off is rarely about a single scalar notion of “performance.” In self-correction, the relevant asymmetry is between correcting existing errors and damaging the larger pool of currently correct outputs (Liu et al., 24 Apr 2026). In coreference, it is between resolving and abstaining when semantics underdetermine the referent (Shore et al., 17 Sep 2025). In VLA control, it is between closed-loop robustness and policy-call efficiency (Pan et al., 2 Jul 2026). This suggests that the CORRECT–DETECT trade-off is best treated as a family of quantitatively specified operating-point problems rather than as a single universal metric.

A common misconception is that “more detection” or “more reflection” is automatically beneficial. The cited work repeatedly rejects that default assumption. Naive LLM refinement can degrade strong models when EIR is not near zero; ambiguity-sensitive prompts can make a model over-abstain even in clear coreference cases; aggressive drift detection in robotics can increase interrupts and computational overhead; watermark detectors that enlarge the acceptance region without calibration lose the strict false-positive guarantees that designated-codeword architectures were designed to provide (Liu et al., 24 Apr 2026). A plausible implication is that detection mechanisms must usually be paired with either explicit thresholds, budget constraints, or stability criteria.

Another important limit is domain specificity. The quantities that mediate the trade-off differ sharply: $a_i^*$16 versus $a_i^*$17 in self-correction, Correct-Unamb versus Detect-Ambig in coreference, score thresholds and list-decoding budgets in watermarking, or $a_i^*$18, $a_i^*$19, and $a_i^*$20 in compression. The literature therefore does not support a single, domain-independent law of CORRECT–DETECT behavior. What it does support is a more general pattern: systems that act on uncertain internal or external detection signals require explicit accounting of how often those signals produce net correction versus false or costly intervention. In that sense, the CORRECT–DETECT trade-off has become a useful organizing concept for turning vague intuitions about “better checking” into measurable error dynamics, stopping rules, and operating curves (Shore et al., 17 Sep 2025).