---
title: 'PFRB in LLMs: Measuring Reasoning Boundaries'
url: https://www.emergentmind.com/topics/partially-feasible-reasoning-boundary-pfrb
type: topic
---

# PFRB in LLMs: Measuring Reasoning Boundaries

A partially feasible reasoning boundary (PFRB) is a technical concept in the quantitative analysis of large language model (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 [2505.13307].

## 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 $0.10 < \text{Acc} < 0.90$, i.e., $\mathcal{B}_{0.10 < \text{Acc} < 0.90}$.

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 $T_1, \ldots, T_n$, 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,
\[
\mathcal{B}(T_1 \oplus \cdots \oplus T_n) \approx \frac{1}{\sum_{i=1}^n [1 / \mathcal{B}(T_i)]}.
\]
More generally, allowing per-task scale $N_i$ and offset $b_i$,
\[
\mathcal{B}(T_1, \ldots, T_n) \approx \frac{1}{\sum_{i=1}^n \left[ \frac{N_i}{\mathcal{B}(T_i) - b_i} \right]}.
\]
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 [2505.13307]. 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 $z_i$:
\[
\mathcal{B}(T_1, \ldots, T_n) \approx \frac{1}{\sum_{i=1}^j \frac{N_i}{\mathcal{B}(T_i)-b_i} + \sum_{i=j+1}^N z_i}.
\]
$z_i$ 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 $\mathcal{B}_u$ (e.g., vertical-domain reasoning) is still too coarse, RBF++ proposes a division mechanism:
\[
\mathcal{B}_u = \frac{1}{1/\mathcal{B}_{v_1} + 1/\mathcal{B}_{v_2}},
\]
for instance, decomposing $\mathcal{B}_u$ into domain knowledge ($\mathcal{B}_k$) and multimodal perception ($\mathcal{B}_{mm}$):
\[
\mathcal{B}_u = \frac{1}{1/\mathcal{B}_k + 1/\mathcal{B}_{mm}},
\]
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, $\mathcal{B}_{90\%} \approx 2.2 \times 10^5$, $\mathcal{B}_{10\%} \approx 2 \times 10^6$.
- Step-planning RB: $\mathcal{B}_{90\%} \approx 2$ steps, $\mathcal{B}_{10\%} \approx 4$ steps.
- BigGSM (GPT-3.5-Turbo): CoT $57.00\%$, Tool Usage (TU) $71.64\%$, Program-of-Thought (PoT) $78.25\%$.
- In PFRB, self-consistency voting boosts accuracy from $\sim60\%$ to $\sim75\%$; in CFRB, zero-shot CoT rationales increase correctness $+29.2\%$ over PFRB/CIRB; in CIRB, ensemble techniques yield no tangible gain (always $\sim10\%$) [2505.13307].
- Synthetic-CoT prompts localize $>\!65\%$ of samples into CFRB, demonstrating models' self-awareness of their RB.
- In multimodal contexts (M3CoT), direct-prompt measurable $z_1$ and the constant-augmented combination law locate distinct 90%/10% RBs, with similar three-zone structure.
- Open-source models often have $\mathcal{B}_{90\%} = 0$ 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 ($\mathcal{B}(c) \rightarrow \infty$), PoT (raises $\mathcal{B}(p)$), MARP (caps per-step operations).
- **Domain-knowledge RB ($\mathcal{B}_k$)**: Context injection, retrieval, expert-curated exemplars.
- **Perceptual RB ($\mathcal{B}_{mm}$)**: Attention-focused prompting, object cropping, perceptual tool integration.
- **Optimization in practice**: MARP++ (explicit multimodal/perception/knowledge constraints) raises accuracy to $73.77\%$, outperforming both standard MARP ($68.68\%$) and baseline CoT [2505.13307].

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 [2505.13307], is:
1. Identify measurable and unmeasurable subtasks, and their respective difficulty axes.
2. For measurable branches, empirically estimate $\mathcal{B}_1, \ldots, \mathcal{B}_j$ by analyzing accuracy vs. difficulty at thresholds $K_1=0.90,\ 0.10$.
3. For unmeasurable components, instantiate constants $z_{j+1}, \ldots, z_N$ using direct accuracy in non-CoT settings.
4. Decompose coarse unmeasurable RBs into knowledge ($\mathcal{B}_k$) and perception ($\mathcal{B}_{mm}$), 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 $\mathcal{B}$ 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 [2505.13307].

Source: https://www.emergentmind.com/topics/partially-feasible-reasoning-boundary-pfrb