IG-PRM: Multi-Agent LLM Optimization

• IG-PRM is a principled framework for multi-agent LLM reasoning that operationalizes proposal diversity, precise feedback, and iterative synthesis to improve problem-solving.
• It decomposes performance improvements into exploration gains, information quality, and aggregation efficiency, offering theoretical guarantees and empirical validation.
• Empirical benchmarks show IG-PRM achieving state-of-the-art results with significant accuracy gains and enhanced compute efficiency compared to single-agent methods.

The Information-Gain Propose–Review–Integrate Synthesis (IG-PRM) framework is a principled methodology for @@@@1@@@@ with LLMs, grounded in a theoretical gain decomposition that quantifies the distinct advantages of exploration (proposal diversity), information (feedback fidelity), and aggregation (consensus formation) in collaborative problem-solving. IG-PRM not only formalizes the sources of performance improvement in multi-agent LLM systems but operationalizes these insights into an iterative workflow combining structured role diversity, high-fidelity execution feedback, rigorous cross-review, and closed-loop synthesis, leading to state-of-the-art results on mathematical, coding, and function-calling benchmarks (Yang et al., 9 Feb 2026).

1. Theoretical Foundations and Gain Decomposition

IG-PRM models the multi-agent reasoning process using a tuple $(\mathcal X, \mathcal T, Q)$ for the task and $(K, \{P_k\}, \mathcal E, f)$ for the multi-agent system, where $K$ is the number of agents, $\{P_k\}$ are agent policies, $\mathcal E$ is the executor (feedback provider), and $f$ is the aggregation function producing the final solution $\tau^{\mathrm{MAS}}$. The single-agent baseline is defined by success rate $p = P(Q(\tau)=1)$.

The gain decomposition (Theorem 3.1) yields: $$\mathbb{E}[Q(\tau^{\mathrm{MAS}})] = C_K \times \eta(f,s),$$ where:

• $C_K = P(\bigvee_{k=1}^K Q(\tau^{(k)})=1)$ is coverage probability (exploration gain).
• $$\eta(f,s) = P(Q(\tau^{\mathrm{MAS}})=1 | \exists k: Q(\tau^{(k)})=1)$$ is selection accuracy given the feedback signal $s$ and aggregator $f$.

The performance gain over the single agent baseline decomposes as: $$\mathbb{E}[Q(\tau^{\mathrm{MAS}})] - p \leq \mathcal{G}_{\mathrm{explore}} + \mathcal{G}_{\mathrm{info}} + \mathcal{G}_{\mathrm{aggr}},$$ where:

• $\mathcal{G}_{\mathrm{explore}} = C_K - p$
• $$\mathcal{G}_{\mathrm{info}} = C_K[\eta^*(e) - \eta^*(\sigma)]$$, with $\eta^*(s)$ the Bayes-optimal selection accuracy for signal $s$.
• $$\mathcal{G}_{\mathrm{aggr}} = C_K[\eta(f,s) - \eta(f_{\mathrm{base}}, s)]$$

The sum is strictly subadditive due to multiplicative interactions between dimensions.

2. Information Gain and Feedback Signal Hierarchy

The central insight of the IG-PRM framework is that performance hinges not only on proposal diversity but critically on the fidelity of feedback used during review and aggregation. The mutual information between the ground truth $Q$ and feedback signals determines maximal achievable selection accuracy (Proposition 3.3):

• $e$ = execution feedback (deterministic for code, maximal information).
• $\sigma$ = textual self-evaluation (limited fidelity).
• $\sigma_v$ = model-based pseudo-verification (intermediate).

Formally, $I(Q;e) = H(Q)$ (deterministic, maximal), whereas $I(Q;\sigma) < H(Q)$, and

$$\mathcal{G}_{\mathrm{info}} = C_K[\eta^*(e) - \eta^*(\sigma)].$$

Fano’s inequality implies higher $I(Q;s)$ directly lifts selection accuracy, and code execution settings permit maximal information extraction. In domains without deterministic feedback, pseudo-verification approximates this benefit.

3. IG-PRM Workflow: Propose–Review–Integrate Cycle

The IG-PRM procedure executes as a series of structured phases:

1. Propose (Exploration): $K$ proposers, each assigned a distinct role $R_k$ (e.g., Minimalist, Explorer), generate solutions $\tau^{(k)}$. Negative correlation between roles ($\rho_{ij} < 0$) enhances proposal diversity.
2. Execute (Information): Each proposal is evaluated via executor $\mathcal{E}$, yielding $e^{(k)}$ (pass/fail signal, tests, error trace), or via a pseudo-verifier $\sigma_v^{(k)}$ if deterministic execution is unavailable.
3. Review (Enhanced Information): Each proposal is cross-reviewed by other agents using structured signals, providing judgments $v_j^{(k)} = \mathcal{R}_j(\tau^{(k)}, e^{(k)})$. Under bounded error $\epsilon_0$ per review and independence, aggregation achieves efficiency $\eta(f_{\mathrm{PRM}}) \geq 1 - \epsilon_0^{K-1}$ (Proposition 3.7b).
4. Integrate (Aggregation): A synthesizer $\mathcal S$ iteratively refines solutions, leveraging cross-review and execution feedback. Synthesis proceeds via iterative closed-loop validation: an initial draft is executed, and failures drive further refinement up to $S$ iterations. The theoretical bound (Theorem 3.8c) is: $$\mathbb{E}[Q(\tau^{\mathrm{PRM}})] \geq [1-(1-p)^K][1-\epsilon_0^{K-1+S}]$$ Pseudocode, as provided, reflects this loop.

4. Role of Information-Gain Signals in Driving the Cycle

Execution-grounded or model-verified feedback fundamentally anchors the review and integration steps. During review, high-fidelity signals ensure each judgment’s error is bounded, and evidence-based cross-evaluation amplifies robustness against incorrect proposals. Aggregation maximizes $C_K \eta(f,s)$, and synthesis can be formulated as a potential game whose equilibria yield consensus solutions.

The explicit decomposition of the task into exploration, high-fidelity information extraction, and precise aggregation enables efficient and reliable multi-agent consensus, which is unattainable with purely heuristic aggregation or low-fidelity feedback.

5. Empirical Validation and Performance Analysis

Benchmarks across arithmetic reasoning, code generation, and function calling (GSM8K, AIME-2025, MBPP, BFCL-SP) demonstrate consistent empirical advantages for IG-PRM over established baselines:

Benchmark	IG-PRM Accuracy (%)	Best Baseline (%)	Single Agent (%)
GSM8K	91.1	87.1 (MoA)	83.6
AIME-2025	93.3	86.7	70.0
MBPP	84.6	78.0 (Self-Cons.)	76.0
BFCL-SP	92.3	88.8 (Two Heads)	81.8

Ablations on each gain dimension indicate additive improvements (MBPP: +3.6 pp from K proposal, +2.6 pp reviewers, +1.0 pp synthesis iterations), and full IG-PRM achieves 8.6 pp over the single agent with a synergy coefficient of 0.88, confirming subadditivity arising from interdimensional interactions.

Compute-efficiency analysis on MBPP shows IG-PRM surpasses majority-of-agents (MoA) ceilings at a fraction of the compute, and scales accuracy further with higher budget, remaining Pareto-optimal.

Code-execution benchmarks (MBPP, BFCL-SP) benefit maximally from the information gain provided by deterministic feedback, while tasks with only textual verification (GSM8K) achieve more modest gains due to lower feedback fidelity.

6. Diagnostic Insights, Limitations, and Future Directions

Empirical diagnostics attribute the largest gains on code and function-calling tasks to information gain; in contrast, tasks with high single-agent baseline require robust aggregation design. Naive voting is vulnerable when per-agent accuracy is low, while IG-PRM’s evidence-based synthesis maintains strong performance.

Identified limitations include:

• Dependence on the availability and fidelity of execution feedback; deterministic execution yields maximal gain.
• Heuristic step segmentation; richer structured representation could further enhance review accuracy.
• Theoretical MCMC proposals are approximated via LLM iterator steps, and future work may address this via better probabilistic modeling or variational filtering.
• The potential for hierarchical or multi-level multi-agent arrangements for scaling to complex domains.

IG-PRM thus provides actionable design principles for optimizing multi-agent LLM systems and sets the current state-of-the-art for both compute efficiency and solution quality in diverse reasoning benchmarks (Yang et al., 9 Feb 2026).