--- title: Two-Stage Abstention Paradigm url: https://www.emergentmind.com/topics/two-stage-abstention-paradigm type: topic --- # Two-Stage Abstention Paradigm Searching arXiv for the cited papers and related abstention literature. The Two-Stage Abstention Paradigm is a class of decision procedures in which abstention is not treated as a single terminal refusal, but as the outcome of an upstream stage that estimates answerability, evidential adequacy, confidence, representational scope, or participation warrant, followed by a downstream stage that either answers, defers, halts, or abstains. Across the recent abstention literature, this pattern appears in large language model reasoning, clinical decision support, selective question answering, multimodal systems, multi-class classification, policy learning, active learning, and even costly voting models. The common structural claim is that abstention quality improves when refusal is conditioned on a separately modeled intermediate object—such as a knowledge graph evidence pool, a latent unanswerability signal, an explicit answerability judgment, a calibrated confidence score, a disagreement region, or a perceived pivotality threshold—rather than on a single undifferentiated confidence output alone [2407.18418]. ## 1. Conceptual definition and canonical decomposition In the survey formulation, abstention is the model’s refusal to answer a query when proceeding would be unsafe, misleading, or unreliable, and a practical system organizes abstention in two stages: Stage 1 is a gate that decides answer versus abstain, and Stage 2 is an action stage that either proceeds with answering or returns a refusal, clarification request, disclaimer, escalation, or related deferral behavior [2407.18418]. This gate can depend on query answerability $a(x)$, model confidence $c(x,y)$, and human-value alignment $h(x), h(x,y)$, with refusal function $$ r(x,y) = 1 \text{ if } a(x) < \theta_a \text{ or } c(x,y) < \theta_c \text{ or } h(x,y) < \theta_h; \text{ otherwise } r(x,y) = 0. $$ The same survey places this within cost-sensitive abstention and selective prediction, with expected risk $$ R = \mathbb{E}[ \ell(f(x), y) \cdot \mathbb{1}\{g(x)=1\} ] + c \cdot \mathbb{E}[ \mathbb{1}\{g(x)=0\} ]. $$ Subsequent papers instantiate this decomposition in more specialized ways. In clinical reasoning, KnowGuard explicitly describes an “investigate-before-abstain” design in which Stage 1 discovers evidence and Stage 2 evaluates and reprioritizes it before an LLM doctor decides whether to abstain or diagnose [2509.24816]. In reasoning models under insufficient information, Judge-Then-Solve casts abstention as a control decision: Stage 1 produces an explicit answerability commitment, and Stage 2 either terminates immediately with abstention or proceeds to solution generation [2605.28070]. In prompt-only factual QA, I-CALM defines Stage 1 as confidence elicitation and calibration and Stage 2 as a reward-aware abstain-or-answer decision [2604.03904]. In the residual-adequacy architecture, Stage 1 attempts clean emission with disciplined escalation, while Stage 2 returns a typed, witnessed abstention when representational warrant fails [2605.24999]. A recurring misconception is that “two-stage” merely means “answer first, abstain later.” The cited papers do not define it that way. In the strongest formulations, the first stage is not a preliminary answer draft but a distinct estimation or control layer: evidence discovery, cognitive monitoring, answerability judgment, value estimation, disagreement localization, or regime adequacy assessment. This suggests that the paradigm is best understood as a separation between *warrant formation* and *terminal action*, not simply as a two-turn dialogue template. ## 2. Formal structures and decision rules The paradigm is formalized in several mathematically distinct but structurally similar ways. In KnowGuard, abstention in multi-round clinical reasoning is modeled as a binary decision at each round with patient knowledge $K=\{k_0,k_1,\dots,k_n\}$, round-specific context $K_t$, and a shared contextualized evidence pool $B_t$. The abstention function is $$ A_t: K_t \to \{0,1\}, $$ where $A_t=0$ means abstain and seek more information, and $A_t=1$ means diagnose. The implementation conditions the doctor agent on both patient context and the top-ranked evidence pool, $$ A_t = \mathrm{LLM}_{doctor}(K_t, B_t, \{x_{text}, x_{img}\}), $$ while an inferred formalization states: abstain if $\max_{e \in B_t} p_{final}(e) < \tau_{evidence}$ [2509.24816]. In I-CALM, the second-stage rule is explicitly decision-theoretic. If $p$ is the estimated probability of correctness, then $$ U_{answer}(p) = p \cdot r_{correct} + (1-p)\cdot r_{incorrect}, \qquad U_{abstain} = r_{abstain}, $$ so answering is optimal when $$ p \ge \frac{r_{abstain} - r_{incorrect}}{r_{correct} - r_{incorrect}}. $$ With $r_{correct}=R>0$, $r_{incorrect}=-\beta$, and $r_{abstain}=\gamma$, the threshold becomes $$ p \ge \tau := \frac{\gamma+\beta}{R+\beta}. $$ Under Scheme B $(+1,-1,+0.4)$, $\tau=0.7$ [2604.03904]. In Judge-Then-Solve, the state variable is an explicit answerability commitment $z \in \{\text{Answerable}, \text{Unanswerable}\}$. If $z=\text{Unanswerable}$, the trajectory halts immediately and emits abstention; if $z=\text{Answerable}$, reasoning continues. The paper evaluates the conversion from internal detection to actual abstention through Detection Rate, Overall Abstention Rate, and Abstention@Detection, $$ \mathrm{A@D} = P(A=1 \mid D=1), $$ with the detection-to-abstention gap defined either as $1-\mathrm{A@D}$ or as $\mathrm{DR}-\mathrm{OAR}$ [2605.28070]. In the residual-adequacy architecture, the central quantity is the residual against representational scope. A regime $R_i$ activates when $$ \rho_i(c) = S_i c - P_i(S_i c), \qquad \rho_i(c) \le \theta_i \text{ and } S_i c \ge \phi_i. $$ The active-set residual is $$ r(c) = \sum_{k \in \mathcal{D}} e_k(c), $$ and clean emission requires $r(c)\le \theta_r$, no HALT action, and no conflicts. Otherwise Stage 2 returns one of three typed terminals: `freeze_halt`, `freeze_time`, or `freeze_resid`, each with witness $$ W=(\mathrm{Regime}_{on}(c), \mathrm{Act}_{on}(c), X) $$ that records the cause of abstention “by construction” [2605.24999]. These formulations differ in ontology—evidence sufficiency, expected utility, answerability commitment, residual adequacy—but they share the same architectural separation. Stage 1 computes or updates a latent decision object; Stage 2 maps that object into answer, abstention, or further control flow. ## 3. Major instantiations across domains In clinical reasoning, KnowGuard uses a single shared, contextualized evidence pool and repeats two stages across rounds. Stage 1 performs graph expansion and direct retrieval from a multi-modal medical knowledge graph built from 300+ WHO guidelines with 22k entities and 100k+ relationships. Stage 2 scores candidate evidence through dual relevance assessment, graph coherence prioritization, demographic-guided weighting, and temporal decay, with final priority $$ p_{final}(h,r,t) = (w_{sim}\cdot s_{sim} + w_{rel}\cdot s_{rel} + w_{coh}\cdot s_{coh}) \cdot s_{pop}. $$ The design claim is that the system investigates the medical knowledge space first, then decides whether evidence suffices for diagnosis or whether more information must be sought [2509.24816]. In large reasoning models on unanswerable mathematics, the two-stage method in “Answering the Unanswerable Is to Err Knowingly” couples cognitive monitoring with inference-time intervention. Stage 1 segments the reasoning trace at cues such as “wait” or paragraph boundaries, then applies a linear probe to hidden activations $$ x^c_l = \sum_{h=1}^{H} Q_l^h \, \mathrm{Att}_l^h(P_l^h x_l), \qquad p_\theta(x^c_l) = \sigma(\langle \theta, x^c_l \rangle), $$ to detect unanswerability. Stage 2 appends instructional guidance that restates the possibility that the question is unanswerable, discourages fabricated assumptions, and supports early exit once abstention is established [2508.18760]. Judge-Then-Solve advances a stricter trajectory-level control formulation. The response must begin with an `` block and conclude either “Conclusion: ANSWERABLE” or “Conclusion: UNANSWERABLE.” If the conclusion is UNANSWERABLE, the model closes `` immediately and emits an abstention explanation; if ANSWERABLE, it continues solution generation. The paper explicitly states that abstention is not “another style of final answer” but a policy decision over the trajectory [2605.28070]. Dynamic abstention for chain-of-thought reasoning introduces a token-level variant. Stage 1 monitors the value of continuation during generation and abstains at the first token position where the estimated value falls below the abstention reward $r_\perp$. For $\beta=0$ and binary rewards, the value simplifies to $$ V_0(x,y_{1:t};T)=P_T(r(x,y)=1 \mid x,y_{1:t}), $$ and the rule is to stop when $\hat V_t < r_\perp$. Stage 2 then applies a final selective prediction or safety filter to responses that survive Stage 1 and reach `eos` [2604.18419]. In prompt-only factual QA, I-CALM realizes the paradigm without retraining. Stage 1 elicits verbal confidence and permits a first-round “I don’t know,” followed by a forced “Best Guess” and “Best Guess Confidence” for evaluation. Stage 2 announces explicit reward schemes and humility-oriented norms that reshape the answer-versus-abstain choice [2604.03904]. In scientific reasoning, the framework of “Knowing When Not to Answer: Abstention-Aware Scientific Reasoning” separates claim decomposition and evidence auditing from the final confidence-aware decision. Inputs are decomposed into minimal conditions, each condition is audited against evidence with NLI, and then a claim-level confidence $$ \alpha = \max_{c_i \in C_{crit}} |m_i| $$ is compared against threshold $\tau$ to determine support, refute, or abstain [2602.14189]. These instances show that the paradigm is not tied to one mechanism. The first stage may be retrieval, probing, decomposition, judgment, or calibration; the second stage may be diagnosis, refusal, selective answering, or trajectory halting. ## 4. Theoretical interpretations and guarantees Several papers interpret the two-stage design as a way to move abstention from heuristic confidence reporting toward explicit control, calibration, or certified decision-making. The survey frames the general problem through Chow’s reject option and selective prediction, including coverage $\phi(g)=\mathbb{P}(g(x)=1)$ and selective risk $$ SR(f,g)=\mathbb{E}[\ell(f(x),y)\mid g(x)=1], $$ with risk–coverage curves indexed by a threshold $\tau$ [2407.18418]. The strongest finite-sample guarantees in the provided material appear in “Geometry-Calibrated Conformal Abstention for Language Models.” There, Stage 1 is a conformal abstention rule based on an uncertainty score $$ U_i \coloneqq r(\widehat{Y}_i \mid X_i), $$ with threshold $$ \tau := \operatorname{Quantile}\!\left(\left\lceil(1-\alpha)(n+1)\right\rceil/n,\{U_i\}_{i=1}^n\right). $$ The kept region is $\mathcal K=\{x:r(f(x)\mid x)\le \tau\}$, and the participation guarantee is $$ \mathbb{P}(X_{n+1}\in\mathcal K)\in\left[1-\alpha,\;1-\alpha+\frac{1}{n+1}\right). $$ The same paper derives a conditional correctness lower bound $$ \mathbb{P}(J_{n+1}=1 \mid X_{n+1}\in\mathcal K) \ge \frac{1-\beta}{1-\alpha+\frac{1}{n+1}}\cdot \frac{c}{n}, $$ then uses Stage 2 geometry calibration to align confidence with knowledge involvement in the model’s internal representation geometry [2604.27914]. Dynamic abstention is derived within a KL-regularized RL framework that treats abstention as a terminal action with reward $r_\perp$. The abstention-augmented policy $a(T)$ abstains exactly when the base policy value falls below $r_\perp$: $$ a(T)(y_t \mid x,y_{1:t-1})= \begin{cases} b, & y_t=L\\ (1-b)T(y_t \mid x,y_{1:t-1}), & \text{otherwise}, \end{cases} $$ where $b=\mathbb{I}\{V_\beta(x,y_{1:t-1};T)