PFRB in LLMs: Measuring Reasoning Boundaries

• PFRB is a formal concept that delineates an LLM's transition zone between complete feasibility (CFRB) and infeasibility (CIRB) in chain-of-thought reasoning.
• The framework employs harmonic-mean laws to combine measurable subtasks and uses constants for unmeasurable components, enabling precise estimation of reasoning limits.
• Empirical findings across benchmarks validate PFRB’s utility for targeted model optimization, enhancing understanding of LLM performance under varying task complexities.

A partially feasible reasoning boundary (PFRB) is a technical concept in the quantitative analysis of LLM reasoning capabilities, formalized within the Reasoning Boundary Framework++ (RBF++). PFRB specifies the band of task difficulty within which an LLM’s accuracy transitions between complete feasibility and infeasibility for chain-of-thought (CoT) reasoning. In RBF++, this notion provides a rigorous and actionable partitioning of models’ reasoning limits across both measurable and unmeasurable cognitive dimensions, supporting targeted optimization and theory-grounded benchmarking (Chen et al., 19 May 2025).

1. Mathematical Formalization of Reasoning Boundaries

Let $M$ be a fixed LLM, $T$ a reasoning task, and $d \in \mathbb{D}$ a scalar quantifier of task difficulty (e.g., number of arithmetic steps, plan depth, multi-hop count). For accuracy threshold $K_1 \in [0,1]$, the reasoning boundary is

$$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$

Here, $\text{Acc}(T \mid d, M)$ is model accuracy for task $T$ at difficulty $d$. Typically, three regions are delineated:

• CFRB ($\mathcal{B}_{\text{Acc}\ge0.90}$): completely feasible region (accuracy at least 90%).
• CIRB ($\mathcal{B}_{\text{Acc}\le0.10}$): completely infeasible region (accuracy at most 10%).
• PFRB: the intermediate band $T$0, i.e., $T$1.

This partition allows precise localization of a model's chain-of-thought capability threshold as a function of task complexity.

2. Combination Law for Measurable Subtasks

Complex reasoning tasks are typically decomposed into subtasks $T$2, each exhibiting discrete reasoning boundaries. RBF++ demonstrates that, under mild independence and smoothness assumptions, the combined RB is governed by a harmonic-mean law. In the normalized case,

$T$3

More generally, allowing per-task scale $T$4 and offset $T$5,

$T$6

Empirically, this law accurately predicts RBs in GSM8K multi-step mathematics (90% and 10% contours), HotpotQA multi-hop QA (global planning and entity knowledge RBs), and other tasks (Chen et al., 19 May 2025). The combination law enables quantitative dissection of complex multi-component reasoning and provides actionable compositional guidance.

3. Handling Unmeasurable Reasoning Boundaries: Constant Assumption and Division

In many real-world or multimodal tasks, some sub-boundaries—such as domain knowledge breadth or perception ability—are not experimentally variable. RBF++ replaces each such unmeasurable sub-RB with a scenario-specific constant $T$7: $T$8 $T$9 is computed by evaluating non-CoT direct accuracy for the corresponding sub-domain and solving for the effective RB denominator, enabling continuity of the combination-law machinery when unmeasurable factors are present.

Where such an unmeasurable RB $d \in \mathbb{D}$0 (e.g., vertical-domain reasoning) is still too coarse, RBF++ proposes a division mechanism: $d \in \mathbb{D}$1 for instance, decomposing $d \in \mathbb{D}$2 into domain knowledge ($d \in \mathbb{D}$3) and multimodal perception ($d \in \mathbb{D}$4): $d \in \mathbb{D}$5 with further constants used to fix perception complexity when invariant.

4. Empirical Findings: PFRB Bandwidth and Model Behavior

Extensive experiments validate the PFRB formulation, using 38 models (27 text LLMs and 5 multimodal LLMs) across 13 benchmarks. Quantitative highlights include:

• For multiplication, $d \in \mathbb{D}$6, $d \in \mathbb{D}$7.
• Step-planning RB: $d \in \mathbb{D}$8 steps, $d \in \mathbb{D}$9 steps.
• BigGSM (GPT-3.5-Turbo): CoT $K_1 \in [0,1]$0, Tool Usage (TU) $K_1 \in [0,1]$1, Program-of-Thought (PoT) $K_1 \in [0,1]$2.
• In PFRB, self-consistency voting boosts accuracy from $K_1 \in [0,1]$3 to $K_1 \in [0,1]$4; in CFRB, zero-shot CoT rationales increase correctness $K_1 \in [0,1]$5 over PFRB/CIRB; in CIRB, ensemble techniques yield no tangible gain (always $K_1 \in [0,1]$6) (Chen et al., 19 May 2025).
• Synthetic-CoT prompts localize $K_1 \in [0,1]$7 of samples into CFRB, demonstrating models' self-awareness of their RB.
• In multimodal contexts (M3CoT), direct-prompt measurable $K_1 \in [0,1]$8 and the constant-augmented combination law locate distinct 90%/10% RBs, with similar three-zone structure.
• Open-source models often have $K_1 \in [0,1]$9 in CFRB, indicating significant headroom.

5. Strategies for Optimizing the Partially Feasible Region

PFRB can be deliberately manipulated by targeting its constituent sub-boundaries:

• Measurable boundaries: Tool Usage ($$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$0), PoT (raises $$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$1), MARP (caps per-step operations).
• Domain-knowledge RB ($$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$2): Context injection, retrieval, expert-curated exemplars.
• Perceptual RB ($$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$3): Attention-focused prompting, object cropping, perceptual tool integration.
• Optimization in practice: MARP++ (explicit multimodal/perception/knowledge constraints) raises accuracy to $$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$4, outperforming both standard MARP ($$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$5) and baseline CoT (Chen et al., 19 May 2025).

Self-consistency and rational prompt design shift more tasks into CFRB, while over-fragmentation (e.g., excessive least-to-most division, complex-CoT) can degrade performance if demonstrations become too granular.

6. Workflow for PFRB Localization and Improvement

The RBF++ recipe for PFRB assessment and enhancement, as detailed in (Chen et al., 19 May 2025), is:

1. Identify measurable and unmeasurable subtasks, and their respective difficulty axes.
2. For measurable branches, empirically estimate $$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$6 by analyzing accuracy vs. difficulty at thresholds $$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$7.
3. For unmeasurable components, instantiate constants $$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$8 using direct accuracy in non-CoT settings.
4. Decompose coarse unmeasurable RBs into knowledge ($$\mathcal{B}_{\text{Acc}=K_1}(T \mid M) \coloneqq \sup \{ d \mid \text{Acc}(T \mid d, M) \ge K_1 \}.$$9) and perception ($\text{Acc}(T \mid d, M)$0), measuring each as possible or holding the other fixed.
5. Assemble the full RB using the harmonic-mean forms, including all constants and per-branch measurements.
6. Apply targeted interventions to raise specific sub-boundaries and contract the PFRB.
7. Re-evaluate the model, seeking rightward (more difficult) movement of the 90%/10% RB contours and a reduced PFRB gap.

This closed-loop process rigorously quantifies and advances LLM CoT performance beyond empirical status-quo.

7. Theoretical and Practical Significance

The PFRB, as formalized by RBF++, bridges the gap between largely qualitative assessments of LLM reasoning and rigorous, model-agnostic quantification of cognitive performance ceilings. The framework’s harmonic-mean combination law and constant-division mechanisms provide a uniquely compositional approach to understanding both measurable and unmeasurable task structures. Experimental results establish scaling relationships between $\text{Acc}(T \mid d, M)$1 and benchmark accuracy, validating the central theoretical insight that PFRBs delimit the regimes of partial capability—and thus, optimization focus—in real-world modeling. This framework enables both interpretability and actionable model improvement by rendering the boundaries of reasoning competence both measurable and mutable (Chen et al., 19 May 2025).