Binary Candidate-included Question (BCQ) Overview

• The paper demonstrates that decomposing multi-choice questions into candidate-conditioned binary decisions leads to improved performance over traditional softmax approaches.
• BCQ is defined as a framework that employs binary decisions for candidate inclusion in answer selection, retrieval indexing, and prompt-based reinforcement learning.
• Empirical results highlight trade-offs such as candidate similarity and positional bias, stressing the importance of hard-negative sampling and ranking-aware methods.

Searching arXiv for papers on "Binary Candidate-included Question" and closely related uses of BCQ. Binary Candidate-included Question (BCQ) denotes a family of formulations in which a task is expressed as a binary decision over an explicit candidate, but the term is used in distinct ways across the literature. In question answering and reasoning, BCQ commonly refers to representing each candidate answer as a separate binary decision such as “Is this candidate correct?”, “Is this candidate reasonable?”, or “Is this candidate feasible?” (Ghosal et al., 2022). In retrieval, the term can be mapped to a binary-encoded query whose Hamming neighborhood determines candidate inclusion, as in the interpretation of Binary Passage Retriever (BPR) for open-domain question answering (Yamada et al., 2021). In prompt-based reinforcement learning, BCQ appears as a prompt reformulation that presents one correct teacher response and one incorrect student response as anonymized candidates for a hard question (Lee et al., 16 Jun 2026). Outside NLP, the acronym “BCQ” has unrelated meanings, including “basic constraint qualification” in variational analysis (Huang et al., 2015) and “block clustered quantization” in low-bit LLM inference (Elangovan et al., 7 Feb 2025). The encyclopedic sense of BCQ therefore requires domain-specific disambiguation.

1. Terminological scope and major usages

In the question answering literature, BCQ is most directly associated with recasting a task into binary decisions over candidate-conditioned inputs. "Two is Better than Many? Binary Classification as an Effective Approach to Multi-Choice Question Answering" treats every $(q,c,a_i)$ triple as a separate instance, with a positive label if $a_i$ is correct and a negative label otherwise (Ghosal et al., 2022). "BUCA: A Binary Classification Approach to Unsupervised Commonsense Question Answering" uses the same structural idea, but its binary label is “reasonable” versus “unreasonable” for a $(q,a_i)$ pair rather than “correct” versus “incorrect” in a supervised MCQA setting (He et al., 2023). "John is 50 years old, can his son be 65?" Evaluating NLP Models' Understanding of Feasibility" explicitly describes its binary classification questions as what the query calls Binary Candidate-included Questions, where the model decides whether a candidate statement is feasible given a context (Gupta et al., 2022).

A different but related use appears in retrieval. "Efficient Passage Retrieval with Hashing for Open-domain Question Answering" does not use the term BCQ, but its details explicitly map BPR’s Stage 1 retrieval to a BCQ-style mechanism: the question is binary-encoded as $\mathbf{h}_q \in \{-1,1\}^d$, and passages are included as candidates according to Hamming-distance proximity in binary space (Yamada et al., 2021). In that interpretation, BCQ refers not to binary label prediction over answer options, but to binary candidate inclusion induced by learned hash codes.

The term also appears in recent prompt-based RL. "Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients" defines a Binary Candidate-included Question as a prompt-level reformulation of a hard question in which two anonymized candidate responses are appended, exactly one correct and one incorrect, and the student must solve the original question again using them as references (Lee et al., 16 Jun 2026). Here BCQ is neither classical binary classification nor retrieval indexing; it is a structured prompt augmentation for on-policy RL.

Because “BCQ” is overloaded, the surrounding task definition is essential.

Usage	Core object	Representative paper
MCQA reformulation	Binary correctness decision for each candidate $(q,c,a_i)$	(Ghosal et al., 2022)
Unsupervised commonsense QA	Binary reasonableness decision for each $(q,a_i)$	(He et al., 2023)
Feasibility reasoning	Binary feasibility decision for a candidate statement	(Gupta et al., 2022)
Retrieval interpretation	Binary query code induces candidate inclusion in Hamming space	(Yamada et al., 2021)
Prompt-based RL	Two candidate responses, one correct and one wrong, embedded in prompt	(Lee et al., 16 Jun 2026)

This suggests that BCQ is best understood as an architectural pattern rather than a single standardized benchmark task.

2. Binary candidate conditioning in multi-choice and commonsense QA

The most explicit BCQ formalization in QA is the binary refactoring of multi-choice question answering. In (Ghosal et al., 2022), standard MCQA is described as scoring answer choices with an $n$-way softmax, whereas the proposed classification-based method asks, for each candidate, whether it is correct. For each answer choice $a_i$, the model outputs two logits $s_i^-$ and $s_i^+$, with per-choice softmax

$a_i$0

and training uses the summed binary objective

$a_i$1

Inference chooses

$a_i$2

This is a canonical BCQ formulation because each input explicitly includes one candidate answer, and the model decides whether that candidate belongs to the positive class (Ghosal et al., 2022).

The same paper reports that this refactoring is often stronger than the conventional $a_i$3-class formulation across abductive reasoning, commonsense question answering, science question answering, and sentence completion. On validation with DeBERTa Large, the classification-based method improves ANLI from 89.75 to 92.23, CQA2 from 66.63 to 69.57, QASC-IR from 89.31 to 91.09, and HellaSwag from 94.67 to 95.47, while underperforming on SIQA, PIQA, and CICERO v1 (Ghosal et al., 2022). The authors relate these task differences to lexical and semantic similarity among answer options: when options are highly similar, the standard score-based softmax can remain preferable (Ghosal et al., 2022).

BUCA uses the same binary candidate-conditioned pattern in an unsupervised regime. It converts each downstream multiple-choice instance into a set of binary decisions by asking whether the sentence formed by combining the question and candidate answer is reasonable, with the final prediction

$a_i$4

Training data are synthesized from knowledge graph triples verbalized into QA pairs, with positive instances derived from real triples and negative instances created by random tail replacement (He et al., 2023). BUCA uses a two-node classification head and explores both a traditional binary loss and a margin ranking loss

$a_i$5

with $a_i$6, plus supervised contrastive learning (He et al., 2023).

The significance of these works is methodological. They show that a multi-choice task need not be treated as a single multinomial decision; it can instead be decomposed into candidate-conditioned binary judgments. This suggests a general BCQ design principle: if the downstream task can be written as ranking or selecting among explicit candidates, a binary formulation can be used to exploit richer negative supervision per question.

3. BCQ as candidate reasonableness, feasibility, and yes/no support

Beyond MCQA reformulation, BCQ also covers candidate-conditioned binary judgments in tasks where the positive class is not simply “correct answer.” In FeasibilityQA, each BCQ instance is a pair $a_i$7 where $a_i$8 is a context and $a_i$9 is a candidate statement such as “The price of the book could be 55 dollars.” The output is True if the candidate is feasible and False if it is infeasible (Gupta et al., 2022). Each context yields two BCQs, one feasible and one infeasible, and the dataset contains 3072 BCQ instances with balanced labels, 1536 True and 1536 False (Gupta et al., 2022). The paper reports that GPT-3 reaches 62.96% zero-shot without knowledge and 69.11% zero-shot with knowledge on BCQ, while GPT-2 large remains at about 50.00% and T5-11B stays around 50–51% (Gupta et al., 2022). Pairwise accuracy, which requires both BCQs for a context to be correct, is substantially lower, indicating consistency failures even when single-candidate judgments are partially correct (Gupta et al., 2022).

A related but more naturalistic binary setting is BoolQ. There the instance is $(q,a_i)$0, with passage $(q,a_i)$1, natural-language yes/no question $(q,a_i)$2, and binary label $(q,a_i)$3, modeled as

$(q,a_i)$4

Its core semantics are entailment-like rather than candidate-ranking, but the details explicitly note that this formalization is exactly the same as a BCQ setting where one must determine whether a candidate proposition is supported or contradicted by the context (Clark et al., 2019). BoolQ is noteworthy because it isolates naturally occurring yes/no questions and shows that transfer from entailment data, especially MultiNLI, is more effective than transfer from paraphrase or extractive QA. BERT-large reaches 76.70% test accuracy, while BERT-large plus MultiNLI reaches 80.43%, compared to 90% human accuracy and a 62.31% majority baseline (Clark et al., 2019). The paper argues that “no” answers often require contradiction detection rather than simple absence-of-evidence heuristics (Clark et al., 2019).

These examples illustrate two distinct BCQ semantics. FeasibilityQA uses BCQ to test whether a candidate world state is possible under commonsense constraints, whereas BoolQ uses binary support judgments over naturally occurring propositions. A plausible implication is that BCQ is best seen as a generic interface for proposition evaluation under context, with the meaning of the positive class determined by the task: correctness, reasonableness, feasibility, or entailment.

4. Binary candidate inclusion in retrieval systems

In retrieval-oriented work, BCQ can be interpreted at the indexing level rather than at the answer-classification level. BPR extends Dense Passage Retriever by adding a hashing mechanism that maps questions and passages to binary codes

$(q,a_i)$5

with $(q,a_i)$6, and uses Hamming distance for Stage 1 candidate generation and continuous inner-product reranking for Stage 2 (Yamada et al., 2021). The details explicitly map this to a BCQ-style mechanism: the binary query representation $(q,a_i)$7 determines which passages are likely candidates, and candidate inclusion is defined by proximity in Hamming space (Yamada et al., 2021). In that formulation,

$(q,a_i)$8

The model uses a multi-task objective

$(q,a_i)$9

where candidate generation is trained with a margin-based ranking loss over binary proxies

$\mathbf{h}_q \in \{-1,1\}^d$0

with $\mathbf{h}_q \in \{-1,1\}^d$1, and reranking uses a DPR-style softmax objective with continuous question embeddings and binary passage proxies (Yamada et al., 2021). The practical effect is large: for 21M Wikipedia passages, DPR’s index is 64.6 GB, while BPR’s binary index is 2.0 GB for linear scan and 2.2 GB with a hash table; query time falls from 456.9 ms/query for DPR to 38.1 ms/query for BPR hash-table search at $\mathbf{h}_q \in \{-1,1\}^d$2, without loss of QA accuracy on Natural Questions and TriviaQA (Yamada et al., 2021).

This retrieval interpretation broadens BCQ beyond classification. The question representation itself contains candidate-inclusion information, and candidate selection becomes a geometric property of learned binary neighborhoods rather than an explicit label prediction. This suggests a unifying perspective: whether BCQ is applied to answer options or document retrieval, the common structure is a question-conditioned binary mechanism that separates plausible from implausible candidates before downstream reasoning or reading.

5. Prompt-level BCQ in reinforcement learning

A recent extension of BCQ appears in post-training for small student models. ZPPO defines a Binary Candidate-included Question as a prompt-level reformulation of a hard question $\mathbf{h}_q \in \{-1,1\}^d$3 in which two anonymized candidate responses are appended, one correct from a frozen teacher and one wrong from the student, in shuffled order under identical <candidate> tags (Lee et al., 16 Jun 2026). The BCQ prompt is

$\mathbf{h}_q \in \{-1,1\}^d$4

where the teacher-correct and student-wrong responses are compressed before insertion (Lee et al., 16 Jun 2026).

The defining property is that the teacher stays inside the prompt rather than the gradient. The student generates new rollouts from $\mathbf{h}_q \in \{-1,1\}^d$5, and all policy-gradient tokens remain student-generated. The RL objective is unchanged from the underlying GRPO-style PPO formulation; BCQ only changes the prompt context (Lee et al., 16 Jun 2026). The mechanism is designed for hard questions defined by $\mathbf{h}_q \in \{-1,1\}^d$6, especially all-wrong groups where vanilla RL yields zero advantage and no gradient (Lee et al., 16 Jun 2026).

Empirically, BCQ contributes materially to ZPPO’s gains at small scales. At 0.8B, the VLM average rises from 45.4 for GRPO$\mathbf{h}_q \in \{-1,1\}^d$7 to 48.6 for GRPO$\mathbf{h}_q \in \{-1,1\}^d$8+BCQ, while full ZPPO, which combines BCQ, NCQ, and replay, reaches 50.3 (Lee et al., 16 Jun 2026). The paper argues that BCQ works because it turns a hard all-wrong question into a prompt-level contrastive discrimination problem while preserving on-policy training (Lee et al., 16 Jun 2026).

This use of BCQ differs sharply from standard binary classification. The binary structure is still present, but the model is not asked to output “candidate 1” or “candidate 2.” Instead, the contrastive pair reshapes the state distribution from which the student must solve the original task. A plausible implication is that BCQ can function either as an output-layer formulation or as an input-layer scaffolding device.

6. Positional bias, uncertainty, and evaluation issues

Binary candidate-conditioned evaluation introduces an additional methodological issue: position effects. "Positional Bias in Binary Question Answering: How Uncertainty Shapes Model Preferences" studies binary question answering where two explicit candidate answers are presented and the preferred option is systematically flipped between first and second position (Labruna et al., 30 Jun 2025). It defines two metrics: Preference Fairness and Position Consistency. Position Consistency is

$\mathbf{h}_q \in \{-1,1\}^d$9

measuring whether the same substantive choice is made under order reversal (Labruna et al., 30 Jun 2025). Preference Fairness is based on a normalized raw score

$(q,c,a_i)$0

and the paper reports $(q,c,a_i)$1 as a measure of directional positional bias (Labruna et al., 30 Jun 2025).

Across SQuAD-it-2 Low, Medium, High, WebGPT, and Winning Arguments, the main finding is that positional bias is nearly absent under low uncertainty but grows rapidly when the task becomes uncertain (Labruna et al., 30 Jun 2025). For example, on SQuAD-it-2 Low, Gemini-2 achieves 0.930 accuracy when the correct option is first and 0.913 when it is second, whereas on SQuAD-it-2 High the same model obtains 0.256 versus 0.522, revealing a strong reversal under uncertainty (Labruna et al., 30 Jun 2025). On Winning Arguments, all evaluated models perform much better when the more persuasive message is in second position, indicating a systematic recency bias (Labruna et al., 30 Jun 2025).

This matters for BCQ because many candidate-conditioned systems implicitly assume order invariance. The evidence indicates that such invariance can fail badly when the semantic signal is weak. Therefore, two-pass evaluation with flipped order and metrics such as PF and PC are relevant for any BCQ-style benchmark in which candidates are presented explicitly (Labruna et al., 30 Jun 2025).

A common misconception is that binary candidate reformulations eliminate ranking artifacts because only two options are shown at a time. The positional-bias results indicate the opposite: pairwise presentation can amplify structural prompt biases unless order is randomized or explicitly audited.

7. Conceptual extensions and non-NLP meanings

Some papers use BCQ in ways that are conceptually broader or entirely unrelated to candidate-answer QA. "The Mathematics of Questions" develops a formal algebra of binary questions, decomposing a proposition $(q,c,a_i)$2 into a pure question $(q,c,a_i)$3, an askable question $(q,c,a_i)$4, and an answer (O'Flanagan, 23 Mar 2025). Raising a pure question $(q,c,a_i)$5 acts on probabilities as

$(q,c,a_i)$6

while giving a proposition sets $(q,c,a_i)$7 (O'Flanagan, 23 Mar 2025). The paper also introduces a non-trivial informational relation $(q,c,a_i)$8 defined by

$(q,c,a_i)$9

with

$(q,a_i)$0

and argues that the simplest geometric orientation of askable questions has the same structure as the two-state quantum system (O'Flanagan, 23 Mar 2025). This is not a benchmark or architecture paper, but it provides a formal way to think about binary questions as objects with algebraic and geometric structure.

A still different use appears in quantum-like cognition. "A dynamical approach to compatible and incompatible questions" represents a binary question as a Hermitian operator with eigenvalues $(q,a_i)$1 on $(q,a_i)$2, and interprets BCQs as yes/no questions about specific options or propositions (Bagarello, 2019). Compatibility is defined by commutation,

$(q,a_i)$3

while incompatibility induces order effects and uncertainty relations (Bagarello, 2019). This is conceptually related to binary candidate questioning, especially where order and context matter, but belongs to a distinct formal tradition.

Two important non-NLP senses of BCQ should also be separated. In variational analysis, BCQ means “basic constraint qualification,” with formulations such as

$(q,a_i)$4

and strong BCQ variants for generalized equations (Huang et al., 2015). In LLM compression, BCQ means “block clustered quantization,” a post-training quantization scheme for W4A4 inference with effective bitwidth

$(q,a_i)$5

used to quantize weights and activations via clustered block-wise codebooks (Elangovan et al., 7 Feb 2025). These works are unrelated to candidate-included questioning except for the acronym.

8. Limitations, trade-offs, and open directions

Across its NLP usages, BCQ exposes a recurrent trade-off between local binary supervision and global candidate interaction. The MCQA reformulation in (Ghosal et al., 2022) shows consistent gains on many datasets, but it underperforms on tasks with highly similar answer options such as PIQA and CICERO v1. The authors attribute this to the fact that the binary classifier evaluates each candidate independently, whereas some tasks require finer relative ranking among near-paraphrases (Ghosal et al., 2022). BUCA similarly observes that random-tail negatives can be too easy, causing the model to label many human-generated distractors as reasonable at test time (He et al., 2023). This suggests that BCQ systems often benefit from hard-negative design and ranking-aware objectives.

Retrieval-based BCQ interpretations face a different limit. BPR’s binary-only retrieval without reranking degrades Top-1, Top-20, and Top-100 recall by about 2–3 points, showing that binary candidate generation is efficient but not fully sufficient for high-precision retrieval without a continuous second stage (Yamada et al., 2021). In ZPPO, BCQ depends on teacher competence: if the teacher cannot produce a correct response for a hard question, no BCQ instance can be constructed for that question (Lee et al., 16 Jun 2026). Positional-bias work adds that pairwise BCQ evaluation itself can become unreliable when uncertainty is high unless option order is systematically controlled (Labruna et al., 30 Jun 2025).

Several open directions recur across the papers. (Ghosal et al., 2022) suggests using binary candidate scoring as a bridge between MCQA and open-ended QA reranking. (He et al., 2023) points toward harder negative sampling, richer knowledge integration, and joint candidate reasoning. (Yamada et al., 2021) leaves code length and richer binary indexing structures unexplored. (Lee et al., 16 Jun 2026) suggests that prompt-level BCQ is most useful within the student’s “zone of proximal development,” which raises the broader question of how candidate-based scaffolding should be scheduled across training difficulty. (Labruna et al., 30 Jun 2025) implies that any future BCQ benchmark should report position-sensitive metrics in addition to raw accuracy.

Taken together, the literature supports a stable core meaning. A Binary Candidate-included Question is a formulation in which the question is evaluated through explicit candidate-conditioned binary structure, whether at the output level as a classifier over $(q,a_i)$6, at the retrieval level as a binary inclusion rule over indexed candidates, or at the prompt level as a contrastive pair used to guide generation. The main attractions of BCQ are modularity, strong negative supervision, and compatibility with ranking and retrieval pipelines. Its main complications are candidate similarity, positional bias, dependence on high-quality negatives or teachers, and terminological ambiguity with unrelated uses of the acronym BCQ.