Winner-Identity Probing

• Winner-identity probing is a framework that identifies the optimal entity among alternatives using formal rules and latent criteria across diverse domains.
• It leverages techniques like dynamic programming, adaptive sampling, and linear decoders to address challenges in voting, dueling bandits, neural interpretability, and swarm robotics.
• The approach highlights computational complexity boundaries and informs practical designs for scalable winner verification, active learning, and system explainability.

Winner-identity probing encompasses a family of algorithmic, statistical, and representational methods for identifying which alternative, option, or agent—among a set—can be regarded as the “winner” under some formal rule or latent criterion. Applications span computational social choice (committee and election winners), machine learning (Copeland winner in dueling bandits), LLM internals (symbol-binding in MCQA), swarm robotics (leader identification), and more. This article systematizes the theoretical principles, technical problems, computational complexity, algorithms, and probing methodologies across these contexts.

1. Definitions and Contexts

Winner-identity probing denotes procedures or decision problems whose goal is to determine or certify the identity of a "winner"—or an entity belonging to some optimal winner set—according to a specified rule, structure, or underlying process. Mechanisms and formalizations vary by field:

• Voting/Committee selection: Given an election with $n$ voters and $m$ candidates (possibly structured into districts or subject to multiwinner rules), probing takes the form of verifying if a candidate $c$ (or a committee $S$) is optimal, or if $c$ can be part of some optimal solution under, e.g., Chamberlin–Courant, Monroe, or Copeland rules (Sonar et al., 2020, Bengs et al., 2023).
• Bandits and partial-feedback models: Winner-identity probing entails adaptively querying the system (e.g., through pairwise duels) to identify arms (options) of maximal score (e.g., Copeland winner), especially under indifference or stochastic transitivity (Bengs et al., 2023).
• Model representations: In neural models solving multiple-choice QA, "winner-identity probes" are linear decoders trained on residual activations to extract when (and where) internal state represents the correct answer, distinguishing between content-based decision and symbol-binding stages (Wong et al., 7 Jan 2026).
• Swarm and multi-agent systems: The winner-identity problem becomes identifying a hidden "leader" in a swarm, typically via active interaction and probing policies rather than static observation (Bachoumas et al., 20 Dec 2025).

Despite disparate concrete instantiations, a common thread is that the "winner identity" is typically not directly observable and must be inferred, verified, or extracted from complex, often noisy or partial processes.

2. Formal Problem Statements

The exact formalism varies. Representative canonical formulations include:

• Candidate Winner (CW): "Does there exist an optimal solution containing a distinguished candidate $c$?" For Chamberlin–Courant (CC) and Monroe (M) multiwinner rules:

$\begin{align*} \text{CC--CW} &= \{\, (E,k,c)\mid \exists S\subseteq C,\,|S|=k,\,c\in S,\,\score_{CC}(E,S)=\OPT_{CC}(E,k)\,\} \ \text{M--CW} &= \{\, (E,k,c)\mid \exists S\subseteq C,\,|S|=k,\,c\in S,\,\score_{M}(E,S)=\OPT_{M}(E,k)\,\} \end{align*}$

• Winner Verification (WV): "Is a given solution $S$ optimal?" Formally:

$\text{R--WV} = \{\, (E, k, S) \mid |S| = k,\, \score_{R}(E, S) = \OPT_{R}(E, k) \,\}$

• Bandit/Copeland Winner: Observing stochastic pairwise outcomes, the winner-identity probe aims to output $\hat{\imath}$ such that $\Pr[\hat{\imath}\notin \mathcal{C}] \leq \delta$ where $\mathcal{C}$ is the Copeland set (Bengs et al., 2023).
• Residual-State Winner Probes in Models: At layer $\ell$ and position $p$, the probe maximizes the cross-entropy between its predictions $f^{(\ell)}(r^{(\ell)})$ and the model’s answer index $y$, with top-$1$ accuracy as the main metric (Wong et al., 7 Jan 2026).

Other scenarios—liquid democracy, data stream processing, district-based elections—offer analogous, though domain-tailored, winner-identity queries (Bentert et al., 2022, Bhattacharyya et al., 2015, Dey et al., 2022).

3. Computational Complexity and Hardness Results

Winner-identity probing problems exhibit significant computational boundaries:

• Committee Winner Verification (WV) for CC and Monroe rules is coNP-complete: Given $(E, k, S)$, deciding whether $S$ is optimal is as hard as any problem in coNP (Sonar et al., 2020).
• Candidate Winner (CW) for CC and Monroe is $\Theta_2^{P}$-complete (also $P^{NP}_{\|}$): Deciding if a candidate $c$ can belong to some optimal committee is strictly harder than NP-complete problems under standard complexity assumptions (Sonar et al., 2020). This result holds both for ranking-based and approval-based ballots.
• Dueling Bandits Copeland Winner Identification: Any $(1-\delta)$-correct algorithm requires instance-specific sample complexity scaling as $\Omega(n^2 \Delta^{-2} \ln(1/\delta))$ in general, and $O(n\ln n \Delta^{-2} \ln(1/\delta))$ under certain transitivity assumptions (Bengs et al., 2023).
• Liquid Democracy One-Winner Determination: Under Plurality, the ONE variant is NP-complete even for three alternatives and bounded out-degree graphs (Bentert et al., 2022).
• Data-Stream Models: Computing the exact winner in a one-pass setting requires memory linear in the input; finding an $\varepsilon$-winner is feasible in sublinear space, with optimal bounds derived for many rules (Bhattacharyya et al., 2015).

Only under highly structured constraints (e.g., single-peaked preferences) do islands of tractability appear; for CC and Monroe, dynamic programming yields polynomial-time algorithms in such cases (Sonar et al., 2020).

4. Winner-Identity Probing: Algorithmic Approaches

Winner-identity probing leverages several sophisticated algorithmic frameworks, determined by the informational, computational, or physical interface with the system:

• Parallel Oracle Queries and DP: For CC–CW on single-peaked profiles, a dynamic program combined with dummy-voter injections or axis-based state restriction delivers an efficient test for candidate membership in optimal committees (Sonar et al., 2020).
• Adaptive Sampling and Bandit Algorithms: POCOWISTA maintains potential Copeland scores and iteratively selects duel pairs to resolve uncertainty, attaining instance-optimal sample complexity (Bengs et al., 2023). Under stochastic transitivity, logical inference on transitive closures further reduces required queries.
• Sampling for Winner Prediction: In elections, winner prediction with $(\epsilon, \delta)$ confidence relies on drawing random vote samples whose number is tailored to the margin of victory and the specifics of the voting rule (Bhattacharyya et al., 2015, Dey et al., 2022).
• Data-Stream Model: Reservoir sampling combined with heavy-hitter sketches (Misra–Gries, Count-Min) supports approximate winner identification, subject to provable memory lower and upper bounds for various voting rules (Bhattacharyya et al., 2015).
• Active Probing in Multi-Agent Systems: In swarm leader identification, winner-identity probing is enacted by an adversarial agent using deep reinforcement learning, with policy architectures built from graph transformers (TGR) and structured state-space models (S5), reinforced by Bayesian post-processing for confidence calibration (Bachoumas et al., 20 Dec 2025).
• Model Probing via Linear Decoders: In transformer-based MCQA models, winner-identity is tracked by multinomial probes across the residual stream; content-decisions are visible mid-computation while symbol-binding (the mapping from chosen content to the answer label) is only decodable at emission (Wong et al., 7 Jan 2026).

5. Statistical and Sample Complexity Results

Winner-identity probing, especially in prediction and learning settings, is governed by sample complexity trade-offs:

Setting	Rule/Class	Sample Complexity (asymptotic)
Plurality, k-Approval, Runoff	Single-district, or k=o(m)	$\Theta(\frac{1}{\epsilon^2} \log\frac{1}{\delta})$ (Bhattacharyya et al., 2015)
Approval, Scoring, Maximin, Bucklin	Single-district	$\Theta(\frac{\log(m/\delta)}{\epsilon^2})$ (Bhattacharyya et al., 2015)
Copeland$^\alpha$	Single-district	$O(\frac{\log^3(m/\delta)}{\epsilon^2})$ (Bhattacharyya et al., 2015)
STV	Single-district	$O(\frac{m^2 (m + \log(1/\delta))}{\epsilon^2})$ (Bhattacharyya et al., 2015)
District Election	Any $r$	$$O(\chi_r(m,\epsilon,\delta)\, \epsilon^{-2} \log(1/\delta))$$ (Dey et al., 2022)
Dueling Bandits (Copeland)	$n$ arms	$\Omega(n^2 \Delta^{-2} \log(1/\delta))$ (Bengs et al., 2023)

Marginality of the victory and the structure of underlying preferences critically influence the sample requirements. For $\varepsilon$-winners in data streams, space complexity scales with $1/\varepsilon$ and $\log m$, while tournament structures or bandit settings pay in $n^2$ or $n \log n$ rounds depending on potential logical inferences.

6. Probing in Neural Models and Symbol-Binding

Winner-identity probing in neural LLMs uncovers a two-stage internal process for multiple-choice answering (Wong et al., 7 Jan 2026):

• Stage 1 (Content Decision): At the end of the options, residual stream activations encode the index of the winning option in a linearly decodable fashion. Probe accuracy rises from chance to high values already in mid-layers, indicating an early "commitment" to a choice.
• Stage 2 (Symbol Binding): The association between the selected content and its corresponding output symbol (A/B/C/D) only becomes linearly decodable at positions immediately preceding token emission, and only in the final layers. Symbol-permutation and content-permutation interventions dissociate the stages, confirming the conceptual split.
• Causal Interventions: Probe-aligned activation patching demonstrates that manipulating activations along the winner-identity direction only influences the output at the binding step.

This mechanistic separation clarifies phenomena such as symbol biases and slippage between model reasoning and response. Probing at intermediate states (not just at answer output) differentiates reasoning failures from symbol-binding errors.

7. Practical Implications and Design Guidelines

The proliferation of winner-identity probing methodologies across domains provides actionable frameworks for practitioners:

• For Computational Social Choice: Hardness results delimit the tractability of winner verification and candidate inclusion; islands of tractability (e.g., single-peakedness) are algorithmically exploitable (Sonar et al., 2020). Data-stream and sample-based algorithms facilitate scalable, approximate winner detection in streaming applications (Bhattacharyya et al., 2015, Dey et al., 2022).
• For Bandit and Active Learning Settings: Information-theoretic lower bounds guide the design of adaptive algorithms. Exploiting structure (transitivity, indifference) improves sample and computational efficiency (Bengs et al., 2023).
• For Interpretability Research in ML: The two-stage winner-identity representation in transformers suggests precise targets for mechanistic interpretability, error analysis, and prompt design. Symbol-content disentanglement enables diagnostic and remedial methods for MCQA (Wong et al., 7 Jan 2026).
• For Robotics and Swarm Systems: Probing policies trained via DRL, with graph-based encoders and state-space sequence models, effectively uncover latent leaders in complex systems, generalize across configurations, and are robust to dynamic changes (Bachoumas et al., 20 Dec 2025).

Universal lessons include modeling unknown environments as POMDPs, designing probes that can exploit and actively create information, leveraging permutation-invariant neural architectures, and integrating Bayesian post-processing for calibrated confidence estimates.

---

References:

• "On the complexity of Winner Verification and Candidate Winner for Multiwinner Voting Rules" (Sonar et al., 2020)
• "Identifying Copeland Winners in Dueling Bandits with Indifferences" (Bengs et al., 2023)
• "When Models Decide and When They Bind: A Two-Stage Computation for Multiple-Choice Question-Answering" (Wong et al., 7 Jan 2026)
• "On Swarm Leader Identification using Probing Policies" (Bachoumas et al., 20 Dec 2025)
• "Sample Complexity for Winner Prediction in Elections" (Bhattacharyya et al., 2015)
• "Fishing out Winners from Vote Streams" (Bhattacharyya et al., 2015)
• "Sampling-Based Winner Prediction in District-Based Elections" (Dey et al., 2022)
• "Who won? Winner Determination and Robustness in Liquid Democracy" (Bentert et al., 2022)