Doubly-Efficient Debate

• Doubly-efficient debate is a class of multi-agent protocols that simultaneously reduces computational costs and maintains high accuracy via adaptive, early-stopping strategies.
• It employs dynamic routing, selective aggregation, and redundancy filtering to ensure efficient information exchange and minimize token usage while preserving reasoning quality.
• Empirical evaluations demonstrate significant token savings (up to 94.5%) and competitive performance across tasks like safety evaluation, bias detection, and factual consistency.

Doubly-Efficient Debate refers to a class of multi-agent protocols that achieve simultaneous gains in computational efficiency (reducing inference or communication cost) and performance efficiency (improving or preserving accuracy, robustness, or alignment quality) in debate-style reasoning or safety evaluation. These frameworks are motivated by the need to scale the oversight of complex AI systems and the evaluation of LLMs, while strictly controlling the otherwise prohibitive resource costs associated with naive multi-agent debate. Central to doubly-efficient debate are adaptive interaction topologies, dynamic routing based on disagreement or redundancy, and early-stopping or selective aggregation mechanisms—all designed to minimize token usage and computation without compromising substantive judgment or reasoning quality.

1. Formal Definitions and Theoretical Foundations

The formal structure of a doubly-efficient debate protocol arises from a three-party interaction: Honest Prover (A), Dishonest Prover (B), and an Efficient Verifier (V), with possible queries to an oracle 𝒪 representing human or reference judgments. In the seminal formulation, an (s, t, q)-debate protocol is defined by the runtime s(n) of the provers, runtime t(n) and maximum queries q(n) of the verifier, with the key property that both s(n) and t(n) are sublinear or polynomial in the intrinsic problem-complexity T(n), and q(n) is ideally constant or sublinear (Brown-Cohen et al., 2023).

Completeness and soundness are formalized as

$$\begin{aligned} &\text{For all }x \in L, \forall B',\;\Pr[V^{\mathcal O}(x, A, B') = 1] \geq c \ &\text{For all }x \notin L, \forall A',\;\Pr[V^{\mathcal O}(x, A', B) = 1] \leq s \end{aligned}$$

where c and s define the desired correctness guarantees. Deterministic constructions enable honest strategies to succeed by simulating executions in time O(T log T), while the verifier reads only O(1) or O(K^2) slices of the transcript, depending on whether the system is deterministic or stochastic (with K-Lipschitz continuity constraints). These protocols guarantee that even exponentially complex tasks can be adjudicated by efficient debates without direct full computation, provided task structure permits transcript-based breakdown (Brown-Cohen et al., 2023).

2. Adaptive Debate Topologies and Progressive Reasoning

Doubly-efficient debate systems adopt adaptive interaction patterns, with early-stopping and dynamic escalation conditioned on measured disagreement or convergence. Key frameworks include:

• Heterogeneous Consensus-Progressive Reasoning for Efficient Multi-Agent Debate (HCP-MAD) (Liu et al., 3 Apr 2026):
  • Stage 1: Heterogeneous Consensus Verification (HCV) employs two maximally diverse agents for single-shot response. If agreed (Φ_init = 1), the answer is immediately accepted, resolving ∼80% of queries (MMLU) at ∼980 tokens per query.
  • Stage 2: Heterogeneous Pair-Agent Debate (HPAD) retains the agent pair for up to T − 1 rounds, using adaptive counters (answer-exchange E_t, deadlock D_t) with thresholds η_e, η_d to guide escalation. About 14% of queries are resolved in this stage at ∼5,700 tokens.
  • Stage 3: Escalated Collective Voting (ECV) aggregates additional independent and contextual agents, with weighted majority favoring independent consensus. Only ∼5.6% of queries reach this phase, at ∼6,700 tokens.

Mathematically, the early-stop and escalation are governed by indicator criteria: $$\begin{aligned} \text{HCV:} &\quad \Phi_{\mathrm{init}} = \mathbb{1}(J(r_{1,0}) = J(r_{2,0})) = 1 \implies \text{stop} \ \text{HPAD:}\;\text{at round}\; t, &\quad \begin{cases} \hat{y}_{1,t} = \hat{y}_{2,t} \implies \text{stop} \ E_t \geq \eta_e \;\vee\; D_t \geq \eta_d \;\vee\; t = T-1 \implies \text{escalate} \end{cases} \end{aligned}$$ Enabling rapid culling of easy cases maximally concentrates resources on complex queries, aligning cost with inherent task difficulty (Liu et al., 3 Apr 2026).

3. Redundancy Filtering and Sparse Information Propagation

Token cost in naive multi-agent debate grows quadratically with the number of agents M and rounds T. Doubly-efficient protocols aggressively sparsify communication:

• S²-MAD (Sparsified and Selective MAD) (Zeng et al., 7 Feb 2025):
  • Implements a redundancy filter, where agent i only listens/responds if another agent’s previous viewpoint differs (similarity Sim < ϵ).
  • Participation in each round is conditioned on the indicator $P^t_i$, enabling “no_op” when redundant.
  • Token complexity reduces from $\mathcal{O}(M^2 T C)$ in MAD to $\mathcal{O}(MTQ + \sqrt{M^3T S} C P_{max})$ in S²-MAD, yielding empirical savings up to 94.5% in token cost with only a sub-2% accuracy drop.
• Diversity-Aware Retention (DAR) (Nguyen et al., 21 Mar 2026):
  • Broadcast in each round is limited to a minimal subset of maximally disagreeing agent messages, selected via index-based filtering maximizing a diversity objective on message embeddings or discrete answers.
  • Empirically, this approach improves final accuracy as the agent pool size grows (N = 4, 8), while simultaneously reducing communication and computation by a factor roughly $(1 - K/N)$, where K ≪ N is the number of retained messages.

These sparsification techniques ensure that only genuinely informative contradictions or disagreements are propagated, preventing context-window saturation and inference inefficiency that otherwise degrade both reasoning quality and cost (Zeng et al., 7 Feb 2025, Nguyen et al., 21 Mar 2026).

4. Protocol Instantiations and Empirical Performance

Doubly-efficient debate strategies have been operationalized in multiple frameworks:

• HCP-MAD (Liu et al., 3 Apr 2026):
  • Across six benchmarks (MMLU, CommonsenseQA, GPQA, MATH-500, GSM8K, AQuA): achieves average accuracy 82.46% vs. 80.09% for prior state-of-the-art, and average token cost of 2,137 per query (−19% vs. previous best).
  • Per-stage, HCV resolves 80.1% of queries (92.2% accuracy), HPAD 14.3% (64.3% accuracy), ECV 5.6% (62.2% accuracy).
• Efficient LLM Safety Evaluation through Multi-Agent Debate (Lin et al., 9 Nov 2025):
  • Three-agent protocol (Critic, Defender, Judge), running on Small LLMs (SLMs, e.g. Qwen3-14B), achieves Cohen’s κ of 0.7352 (vs. GPT-4o κ = 0.7627) at 46% inference cost (δ = 0.54), with three rounds maximizing cost-accuracy trade-off.
• S²-MAD (Zeng et al., 7 Feb 2025):
  • For GPT-4-0613 on GSM8K, token usage drops from 50.4k (MAD) to 2.78k (S²-MAD), accuracy 94.2% vs. 93.3%.
  • Ablation removing the redundancy filter increases cost by 183% for only a 2% accuracy gain.
• DAR (Nguyen et al., 21 Mar 2026):
  • On decentralized multi-agent debate (N = 4), accuracy on Qwen-3B increases from 61.37% (majority vote) to 64.02% (DAR).
  • As N increases, standard debate performance saturates/degrades, while DAR accuracy continues to climb; communication cost reduced proportionally with retention ratio.

These results establish that doubly-efficient systems deliver a strict improvement over “baseline” exhaustive debate without sacrificing accuracy, and are robust across reasoning, factual QA, mathematics, and high-stakes safety adjudication tasks.

5. Complexity, Trade-Offs, and Optimization Criteria

The cost-accuracy trade-off in doubly-efficient debate protocols is explicitly modeled. Cost C is often measured as total tokens consumed (proxy for inference cost), with debate complexity scaling as:

$$C_{\text{MAD}} \sim O(T \cdot N^2 \cdot L) \quad \text{(full agent broadcast)}$$

$$C_{\text{S}^2\text{-MAD}},\; C_{\text{DAR}} \sim O(T \cdot N \cdot K \cdot L),\;K \ll N$$

Optimal protocol hyperparameters (e.g., number of rounds n in (Lin et al., 9 Nov 2025)) are selected by maximizing objectives such as

$$\max_n\;\left[ A(n) - \lambda \, C(n) \right] \quad \text{or} \quad \max_n\,\frac{A(n)}{C(n)}$$

where A(n) is accuracy/Cohen’s κ. Empirical ablation confirms “diminishing returns” in additional debate rounds, with a small n (e.g. n = 3) being optimal.

Stagewise allocation and diversity-aware retention ensure that computation concentrates only where disagreement or ambiguity persists; for most instances, early termination mechanisms yield savings without loss of verdict quality.

6. Broader Applications and Limitations

Beyond core reasoning or QA, doubly-efficient debate has been applied to scalable safety evaluation (jailbreak detection, risk assessment (Lin et al., 9 Nov 2025)), factual consistency, bias assessment, and toxic content moderation. The frameworks are generic: by swapping in domain-appropriate evaluation modules or value-alignment topics, the same infrastructure can scale to diverse, subjective, or high-stakes AI oversight settings.

Principal limitations remain around:

• The need for tasks to admit a compact, transcript-checkable debate structure.
• The requirement for provers (agents) to simulate human-judgment oracles with sufficient fidelity.
• Degradation when a nontrivial subset of oracle queries is adversarially noisy or if full cross-examination is impractical (Brown-Cohen et al., 2023).

Potential extensions include dynamic thresholding for redundancy filters, per-task adaptive grouping, richer neural disagreement metrics, deeper integration with chain-of-thought/or human-in-the-loop protocols, and robust protocols against adversarial oracle errors (Zeng et al., 7 Feb 2025).

---

In summary, doubly-efficient debate embodies a principled shift in multi-agent AI oversight: from exhaustive, communication-heavy protocols to lean, disagreement-driven, and adaptively routed debate architectures. Such protocols are grounded in precise theoretical analysis and deliver empirically validated order-of-magnitude savings in resource consumption while either matching or exceeding the substantive evaluation quality of prior approaches (Brown-Cohen et al., 2023, Zeng et al., 7 Feb 2025, Lin et al., 9 Nov 2025, Nguyen et al., 21 Mar 2026, Liu et al., 3 Apr 2026).