Soundness-Aware Level: A Microscopic Signature that Predicts LLM Reasoning Potential

Xuansheng Wu, Xiaoman Pan, Wenlin Yao, Jianshu ChenSubmission to ICLR 2026 · Rejected

TL;DR

Why do some pre-trained large language models flourish under reinforcement learning with verifiable rewards (RLVR) while others stagnate? This paper reveals that high-potential base models possess an intrinsic, microscopic "soundness awareness"—their internal feature transitions clearly distinguish between logically strict deductions and noisy heuristics. By measuring this separation via the novel Soundness-Aware Level (SAL) metric, the authors can accurately forecast post-RLVR reasoning performance without requiring any downstream labeled data.

What Problem Does This Paper Address?

Reinforcement learning with verifiable rewards (RLVR)—the core training engine behind models like DeepSeek-R1—can unlock impressive mathematical and logical reasoning. However, applying identical RLVR recipes to different pre-trained base models yields wildly inconsistent results: some scale gracefully into capable reasoners, while others plateau.

Prior attempts to explain these disparities have focused on macroscopic outputs, such as looking for reasoning tokens ("wait", "therefore") or measuring output entropy. This paper digs into the model's microscopic internals to answer: What fundamental representation property inside a base model determines its latent capacity to reason?

Methods

The authors build a mechanistic framework to extract and evaluate internal "logic rules":

  1. Feature Extraction with Cross-Layer SAEs: They train cross-layer Sparse Autoencoders (SAEs) across 8 evenly spaced layers on the model's residual stream. Crosscoders reconstruct hidden states using a shared dictionary of sparse, interpretable features across layers.
  2. Implicit Logic Extraction (Horn Clauses): The authors model internal reasoning as forward-chaining Horn clauses ($P_1 \wedge P_2 \rightarrow Q$). Instead of using expensive causal ablations, they track layer-wise co-occurrence frequencies across thousands of reasoning traces to calculate empirical conditional probabilities: $\hat{p}(Q \mid P)$.
  3. Semantic Soundness Annotation: Using DeepSeek-R1 as an automated judge, extracted features and transition rules are categorized into three soundness classes based on semantic descriptions:
  4. Strict: Necessary, deductive truths (e.g., mathematical theorems or definitions).
  5. Plausible: Helpful heuristics or problem-solving intuitions (e.g., isolating a variable).
  6. Noise: Spurious, non-causal correlations.
  7. Soundness-Aware Level (SAL): The authors plot the probability distributions assigned by the model to Strict, Plausible, and Noise rules, computing the multi-distribution Jensen-Shannon Divergence (JSD): $$\text{SAL} = \frac{1}{|Y|} \sum_{y \in Y} \text{KL}(\rho_y \parallel m)$$ A higher SAL indicates that the model assigns distinct confidence distributions to sound vs. unsound rules.

Datasets

  • SAE Training & Rule Extraction Corpus: Curated 128K unique math questions collected from MATH, Open Reasoner Zero, GSM8K, and subsets of NuminaMath (AMC, AIME, Olympiad). Each candidate LLM generates "thinking"-style rollouts for these questions to create model-specific activation corpora (no ground-truth labels used). A 3,267-sample subset of MATH covering 7 categories is used for scalable rule extraction.
  • Evaluation Benchmarks: Post-RLVR downstream reasoning evaluation is compared across MATH500, GSM8K, AMC23, AIME24, Minerva, and Olympiad benchmarks.

Key Insights and Ideas

  • Soundness-Aware vs. Soundness-Agnostic: Strong base models (like Qwen-2.5-7B) assign high probabilities to strict rules and low probabilities to noise rules. Weaker models (like Llama-3.1-8B) suffer from a collapse where all rules—strict, plausible, and spurious—share an overlapping, right-skewed probability profile.
  • Probabilistic, Not Formal Logic: Even the "strictest" extracted internal rules are contextual deductions (e.g., equating variable patterns to algebraic expressions) rather than rigid axiomatic proofs. High-potential models succeed because they maintain a distinct probabilistic hierarchy among rules.
  • Scale Helps, but Diminishes: Within the Qwen family, SAL scales steeply from 0.5B to 1.5B parameters, then begins to plateau toward 14B, indicating that sheer parameter scale alone hits diminishing returns in refining internal logic boundaries.
  • Architecture and Pre-training Matter Immensely: At the 7B scale, models show massive variance. Qwen-2.5-7B (~0.16) and DeepSeek-Math-7B (~0.11) achieve far higher SAL than generalist counterparts Mistral-7B (~0.06) and Llama-3.1-8B (~0.06), showing that math-heavy pre-training corpora fundamentally reshape internal rule boundaries.

Key Results

  • Predictive Power via Power Law: Post-RLVR error rate ($\epsilon$) can be modeled as an exponential power law of the base model's soundness-aware level ($s$): $$\epsilon = \exp(-\alpha \cdot s^\beta)$$ The model fits observed data with $R^2 = 0.985$ and achieves an $R^2 = 0.872$ on held-out models under leave-one-out cross-validation.

  • Benchmark Correlation: SAL achieves an average Spearman rank correlation of 96.43% with post-RLVR performance across AMC23, MATH500, Minerva, Olympiad, and AIME24, outperforming behavioral metrics like verification, backtracking, and subgoal decomposition.

  • Zero-Label Advantage: Unlike evaluation metrics that require large test sets with ground-truth answers, SAL is computed purely from unlabeled activation dynamics and LLM-assisted feature descriptions.

Limitations and Future Work

  • Correlation, Not Causality: The authors establish a predictive relationship between SAL and RLVR reasoning potential, but do not perform causal interventions showing that modifying internal rule separation directly changes downstream RL outcomes.
  • Domain Scope: The study primarily evaluates mathematical reasoning; testing whether SAL generalizes to code generation, formal logic, or open-ended reasoning remains open.
  • Labeling Noise: Rule soundness categorization relies on DeepSeek-R1 prompts, which achieve a moderate 0.566 agreement rate with human annotators (mostly confusing adjacent "Strict" vs. "Plausible" boundaries).

Summary updated

Abstract

Reinforcement learning with verifiable rewards (RLVR) can elicit strong reasoning in large language models (LLMs), while their performance after RLVR varies dramatically across different base models. This raises a fundamental question: what microscopic property of pre-trained models leads to this variation? To investigate, we formalize reasoning as chains of Horn clauses ("if-then" rules) built from features extracted from the LLM's latent space via cross-layer sparse autoencoders (SAEs). We estimate the transition probabilities between its features, and further categorize each rule by its semantic soundness level (e.g., strict, plausible, noisy) with an LLM. Our key discovery is that high-potential models are inherently soundness-aware: their internal probability distributions systematically shift across rules' soundness levels, becoming highly distinct for "strict" versus "noisy" rules. In contrast, weaker models are soundness-agnostic, collapsing to one distribution regardless of soundness levels. To quantify this, we introduce the Soundness-Aware Level (SAL), a microscopic metric using the Jensen-Shannon Divergence to measure the separation between these distributions. We show that SAL's predictions of post-RLVR reasoning performance follow a precise empirical law ($R^2=0.87$) across diverse model families (Qwen, Mistral, Llama, DeepSeek) and scales (0.5B–14B). This reveals that a model's reasoning potential is tied to its intrinsic, pre-trained ability to distinguish sound knowledge from unsound ones. These findings underscore the critical role of model pre-training in shaping reasoning and offer a practical metric grounded in the model's internal mechanisms for selecting/designing stronger base models.
BibTeX
@misc{wu2025soundness,
  author = {Xuansheng Wu and Xiaoman Pan and Wenlin Yao and Jianshu Chen},
  title = {{Soundness-Aware Level: A Microscopic Signature that Predicts LLM Reasoning Potential}},
  year = {2025},
  eprint = {2510.15216},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.15216},
}
Figure