PaperlayerView exact revision on arXiv

Source-linked AI summary

PRISM: A Dual View of LLM Reasoning through Semantic Flow and Latent Computation

Ruidi Chang, Jiawei Zhou, Hanjie Chen

arXiv:2603.22754v1cs.CL

TL;DR

Aggregate accuracy reveals whether an answer is correct, but not how reasoning progresses through semantic steps and internal computation. PRISM jointly models these dimensions and finds systematic failure patterns, including sticky verification and distinct ways of handling candidate answers. It also shows that prompting reshapes semantic transitions and internal computation, while remaining an analysis and diagnostic framework rather than a performance optimizer.

  • Problem

    Aggregate accuracy indicates whether a final answer is correct but offers little insight into how a model arrives there, while semantic reasoning trajectories and layer-wise hidden states provide complementary views that are not comprehensively characterized together.

    Source: §1 · p. 1

  • Method

    PRISM models reasoning steps as semantic category transitions, identifies latent computational regimes across transformer layers, and links consecutive steps through a bridge matrix.

    Source: §1 · pp. 1–2, §8 · p. 8

  • Results

    Across four models and four benchmarks, failed trajectories enter verification more readily but struggle to return to computation, show distinct patterns consistent with overthinking and premature commitment, and change in both semantic and computational patterns under prompting.

    Source: §1 · p. 2

  • Takeaways & Limitations

    PRISM provides a practical diagnostic framework for analyzing reasoning dynamics and connecting internal computation with external reasoning structure.

    Source: §8 · p. 8

    PRISM relies on external semantic labeling and unsupervised latent clustering that may introduce noise or ambiguity, and its evaluation covers a limited set of models and benchmarks.

    Source: Limitations · pp. 8–9

Abstract

from arXiv · show

Large language models (LLMs) solve complex problems by generating multi-step reasoning traces. Yet these traces are typically analyzed from only one of two perspectives: the sequence of tokens across different reasoning steps in the generated text, or the hidden-state vectors across model layers within one step. We introduce PRISM (Probabilistic Reasoning Inspection through Semantic and Implicit Modeling), a framework and diagnostic tool for jointly analyzing both levels, providing a unified view of how reasoning evolves across steps and layers. Across multiple reasoning models and benchmarks, PRISM uncovers systematic patterns in the reasoning process, showing that failed trajectories are more likely to become trapped in unproductive verification loops and further diverge into distinct modes such as overthinking and premature commitment, which behave differently once a candidate answer is reached. It further reveals how prompting reshapes reasoning behavior beyond aggregate accuracy by altering both semantic transitions and internal computational patterns. By modeling reasoning trajectories as structured processes, PRISM makes these behaviors observable and analyzable rather than relying solely on final-task accuracy. Taken together, these insights position PRISM as a practical tool for analyzing and diagnosing reasoning processes in LLMs.

1 Introduction

Aggregate accuracy reports whether an answer is correct but gives little insight into how reasoning unfolds, and existing analyses rarely integrate its semantic progression with internal computation. PRISM combines these views to analyze reasoning trajectories and reveals patterns in failures and prompting.

1 Introduction

  • PRISM models semantic category transitions, latent computational regimes within steps, and bridges linking consecutive steps into a unified analysis.The framework connects a step’s semantic role with the internal computational patterns that realize it.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Failed trajectories enter verification more readily but struggle to return to computation, while failure patterns also diverge into overthinking and premature commitment.These patterns differ in how candidate answers are reached and handled.Source: §1 · p. 2
  • Prompt changes reshape both semantic transitions and internal computational patterns, revealing reasoning behavior beyond aggregate accuracy.The paper reports that these changes also affect end-task accuracy.Source: §1 · p. 2

2 Methodology

PRISM models reasoning through semantic transitions, layer-wise latent computational regimes, and bridges between consecutive steps. Its method combines category-specific transition modeling with Gaussian mixture models and a two-phase training procedure.

2 Methodology

  • PRISM combines semantic category flows, category-specific Gaussian mixture regimes, and a cross-step bridge matrix, with the implicit stage and bridge jointly trained by two-phase EM.The bridge links category transitions with internal computational patterns between consecutive steps.Source: §2 · p. 2

2.2 Implicit Reasoning Regimes and Cross-Step Bridges

  • PRISM extracts first-token hidden states across layers for each step, PCA-projects them, and fits category-specific Gaussian mixtures to represent layer-wise latent-regime probabilities.The fitted clusters serve as latent reasoning regimes for the hidden-state geometry.Source: §2 · p. 2, §2.2 · p. 3

2.1 Explicit Reasoning Transitions

  • Each reasoning step receives one of four semantic roles plus an unknown fallback, and the category sequence is modeled as an m-th-order Markov chain.Higher orders capture context-dependent patterns; the first-order case is the standard category transition matrix.Source: §2.1 · pp. 2–3

2.2 Implicit Reasoning Regimes and Cross-Step Bridges

  • A bridge matrix links the preceding step’s exit regime to the next step’s entry regime jointly with the category transition.Its context-dependent entry distribution depends on the previous step’s exit regime and the category transition; joint EM follows a regime-only warm start.Source: §2.2 · pp. 3–4

3 Experimental Setup

The experiments span four open-source reasoning models and four benchmarks covering competition mathematics, graduate-level science, and diverse real-world problems.

3 Experimental Setup

  • Evaluation covers four open-source reasoning models spanning 1.7B–7B parameters and different architectures.The models are Bespoke-Stratos-7B, OpenThinker-7B, Qwen3-1.7B, and Llama-3.1-Nemotron-Nano-4B-v1.1.Source: §3 · p. 4
  • Bespoke-Stratos and OpenThinker are both distilled from DeepSeek-R1 on different datasets using Qwen2.5-7B-Instruct as the base.Source: §3 · p. 4
  • The four benchmarks are MATH-500 and AIME-2024 for competition mathematics, GPQA-Diamond for graduate-level science, and WebInstruct-Verified for diverse real-world problems.Source: §3 · p. 4

4 PRISM Components and Analysis

PRISM characterizes reasoning through semantic flows and latent computations, using these views to examine how steps transition and how internal regimes relate to those transitions.

4.1 Explicit Stage: Understanding Semantic Flow of Reasoning Trajectories

  • UV self-loops (pUV→UV = 0.438) and returns to AC (pUV→AC = 0.320); UV→AC→UV occurs with probability 27.1%.Source: §4.1 · p. 4
  • AC self-transitions dominate (pAC→AC = 0.715), while SR most often moves to AC (pSR→AC = 0.445), and AC has stationary probability πAC = 0.501.Source: §4.1 · p. 4, Table 1 · p. 5
  • Nearly all traces begin in SR (pstart = 0.982), and expected time to FA is approximately 22 steps from any non-FA category.Source: §4.1 · pp. 4–5, Table 1 · p. 5

4.2 Implicit Stage: Understanding Internal Computations

  • The PCA projection shows category activations in different regions, with UV centroid at (−7.66, −6.55).Source: Figure 2 · p. 5, §4.2 · p. 5
  • For Stratos on GPQA-Diamond, SR exits through Regime 1 remain SR more often than Regime 0 exits (p = 0.80 vs p = 0.35), while AC transitions are relatively stable across regimes.Source: Table 2 · p. 5, §4.2 · p. 5

5 Application with PRISM: Failure Patterns & Overthinking

PRISM distinguishes failed reasoning from correct reasoning by its semantic transitions and latent computational patterns. Failed traces tend to enter verification but struggle to return to computation, while long and short failures differ in how they reach and revisit candidate answers.

5 Application with PRISM: Failure Patterns & Overthinking

  • Failed traces enter Uncertainty & Verification more readily but return to Analysis & Computation less often, forming unproductive checking loops.Correct reasoning, by contrast, proceeds with more efficient computation.Source: §5 · pp. 5–6, Table 3 · p. 6
  • Long failures often revisit a candidate answer and continue deliberating, while short failures tend to reach an answer later and stop reasoning soon after.The analysis divides failures at 100 steps; long failures reach Final Answer at median relative position 0.28 and re-enter it 8 times, versus 0.58 and 3 times for short failures.Source: §5 · p. 6, Table 4 · p. 7
  • For incorrect Uncertainty & Verification steps, Regime 4 is more prevalent in the final layer, while Regime 0 appears more in earlier layers.The example is Bespoke-Stratos-7B on GPQA-Diamond; the reported final-layer difference for Regime 4 is approximately −0.24.Source: §5 · p. 6, Figure 3 · p. 6
  • Long and short failures show larger latent-regime divergence in Final Answer and Setup & Retrieval than in Analysis & Computation or Uncertainty & Verification.This suggests the failure modes engage qualitatively different computational regimes during problem setup and answer formulation.Source: §5 · pp. 6–7

6 Application with PRISM: Prompting Strategy Reshape Reasoning Flow

PRISM examines how concise and multi-path prompts reshape reasoning beyond final accuracy. Conciseness fragments computation into shorter segments alternating with verification, while multi-path exploration produces stronger cycling between those stages.

6 Application with PRISM: Prompting Strategy Reshape Reasoning Flow

  • The prompting comparison tests a baseline, a “Be concise” addition, and an instruction to explore separate reasoning paths before choosing an answer.The latter prompt explicitly asks the model to write multiple paths and then decide which option is best supported.Source: §6 · p. 7
  • P2’s conciseness instruction reduces the Analysis & Computation self-loop from 0.636 to 0.549 and raises transitions to Uncertainty & Verification from 0.205 to 0.284.The stationary probability of Analysis & Computation remains similar to P1, indicating shorter computation segments alternating more often with verification.Source: §6 · p. 7
  • P3’s multi-path instruction yields the highest transitions in both directions between Analysis & Computation and Uncertainty & Verification, with its lowest Uncertainty & Verification self-loop at 0.473.The reported transitions are 0.331 from verification to computation and 0.374 from computation to verification; the patterns indicate stronger bidirectional cycling.Source: §6 · p. 7

7 Related Work

Related work examines textual and structural properties of reasoning, internal mechanisms, and information-theoretic or trajectory-based accounts. PRISM is presented as a probabilistic framework connecting semantic transitions with latent computational patterns.

7 Related Work

  • Prior work studies textual content and structural patterns in reasoning, including findings that performance gains can persist despite flawed step content (Madaan and Yazdanbakhsh, 2022; Wang et al., 2022).Bogdan et al. (2025) report that a small set of “anchor” steps disproportionately shapes final outcomes.Source: §7 · p. 7
  • Mechanistic and information-theoretic studies probe reasoning through architectural components, hidden-space phenomena, entropy, mutual information, and encoded capacity.The section also surveys work on overthinking, token-level intervention points, latent-state MDPs, reasoning motifs, and links between trajectory structure and performance.Source: §7 · pp. 7–8
  • PRISM is positioned as complementary to these lines of work because it connects semantic transitions with latent computational patterns for structured trajectory analysis.The supplied prompting comparison describes transition matrices, regime-profile distances, hitting times, and performance statistics, but provides no outcome values here.Source: §7 · p. 8, Table 5 · p. 8

8 Conclusion

The conclusion presents PRISM as a framework that jointly analyzes semantic reasoning trajectories and internal computational structures. It reports diagnostic patterns in failures and prompt-driven changes, positioning the framework as a tool for studying reasoning dynamics beyond final accuracy.

8 Conclusion

  • PRISM jointly models semantic reasoning trajectories and internal computational structures, linking what a step does with how it is realized across transformer layers.Its explicit stage captures semantic category transitions, while its implicit stage identifies latent computational regimes.Source: §8 · p. 8
  • PRISM reveals systematic patterns across correct and incorrect trajectories, including distinct behaviors across failure modes and prompt-related shifts in transitions and internal computational patterns.The conclusion says these prompt variations reshape reasoning beyond final accuracy and influence end-task performance.Source: §8 · p. 8
  • The authors position PRISM as a probabilistic framework and practical diagnostic tool for analyzing and diagnosing language-model reasoning processes.The stated broader aim is to support more interpretable reasoning systems.Source: §8 · p. 8

Limitations

The limitations identify uncertainty in PRISM’s category and regime assignments, a restricted evaluation scope, and its diagnostic rather than performance-optimizing purpose. The appendix lists four open-source reasoning models used in the evaluation.

Limitations

  • External semantic labels and unsupervised latent clustering may make category assignments noisy or ambiguous, complicating regime interpretation.The authors also note that the evaluation covers a limited set of models and benchmarks, and that broader validation would strengthen generality.Source: Limitations · pp. 8–9
  • PRISM analyzes reasoning dynamics but does not directly optimize or improve reasoning performance.The authors say its insights may inform future optimization.Source: Limitations · p. 9

A.1 Models

  • The evaluated open-source models are Qwen3-1.7B, Bespoke-Stratos-7B, OpenThinker-7B, and Llama-3.1-Nemotron-Nano-4B-v1.1.Their listed configurations include temperature 0.6; Qwen3 uses thinking mode, while Bespoke-Stratos and OpenThinker use a reasoning-assistant system prompt.Source: §A.1 · p. 11

A.2 Datasets

The appendix specifies the datasets and prompts, then describes semantic labeling and the preprocessing, latent-regime fitting, decoding, and model-selection steps used in the analysis.

A.2 Datasets

  • The experiments span four benchmarks, with dataset splits, sizes, maximum tokens, and licenses listed in Table 6.WebInstruct-Verified is filtered by answer type and difficulty.Source: §A.2 · p. 11, Table 6 · p. 11

A.3 Prompts

  • GPQA uses a multiple-choice step-by-step prompt, while MATH-500, AIME-2024, and WebInstruct request step-by-step answers with a final-answer line.The appendix gives the exact output formats for both prompt groups.Source: §A.3 · p. 11

A.4 Semantic Category Classification

  • Each reasoning step is assigned a semantic category using Qwen2.5-32B-Instruct, and the prompt asks the model to classify a sentence into one of four tags, including setup_and_retrieval and analysis_and_computation.Source: §A.4 · p. 11

A.5 Preprocessing

  • Hidden states are mean-centered and RMS-normalized by layer, rescaled across layers within each step, then projected to 128 PCA dimensions.The explained-variance curve tends to flatten after 128 dimensions, with smaller gains from additional components.Source: §A.5 · p. 11, Figure 5 · p. 12

A.6 EM Step

  • EM estimates category-specific regime responsibilities and GMM parameters, while the bridge update accumulates joint posteriors across consecutive steps.The mixture weights use layers ℓ ≥ 2 because the bridge matrix governs the first-layer prior; the bridge links the preceding step’s exit posterior to the current step’s entry posterior.Source: §A.6 · p. 12

A.7 Decoding

  • The discrete regime for each layer is decoded using its maximum-posterior assignment.Source: §A.7 · p. 12

A.8 K Selection

  • The number of regimes is selected by sweeping candidate K values, scoring them with silhouette on a bridge-free EM fit, then retraining the selected K with the bridge.Source: §A.8 · p. 13

C Results

The results section reports semantic transition analyses across trajectory subsets, summary and failure-mode statistics, and an illustrative GPQA-Diamond example. The supplied passages identify the analyses and visuals but do not provide their numerical outcomes.

C Results

  • The appendix reports second-order semantic transition probabilities and first-order matrices for all, correct, and incorrect trajectories, plus the correct-minus-incorrect difference.Tables 8–12 present these transition analyses; the supplied passages do not state their cell values or comparative outcomes.Source: C Results · p. 14, Table 8 · p. 14, Table 9 · p. 14
  • The results also include overall trajectory statistics and an overthinking analysis.Tables 13 and 14 summarize these measures, with Table 14 distinguishing final-answer visit metrics from L2 values.Source: C Results · p. 14, Table 13 · p. 14, Table 14 · p. 17
  • A Bespoke-Stratos-7B example on GPQA-Diamond combines a category activation-space projection, posterior-probability differences, and explicit bridge matrices.The appendix identifies Figure 7, Figures 8–11, and Table 15 for these views.Source: C Results · p. 14, Figure 7 · p. 14, Figure 8 · p. 15
Loading 2603.22754v1…