Probing the Trajectories of Reasoning Traces in Large Language Models

TL;DR

This study introduces a trajectory-probing protocol to dissect how reasoning traces in large language models influence final answers. By generating full traces, slicing them into token-based deciles, and injecting partial traces to measure subsequent answer distributions, the authors quantify accuracy gains, commitment dynamics, and information contributed by reasoning content. They show that gains are largely driven by instance-specific semantic signals rather than mere length or generic reasoning style, and that stronger models can rescue incorrect traces—especially with continuation—though some traces lead to anchoring. The work provides practical diagnostics for compute-efficient and safer deployment of reasoning systems and highlights design choices for trace handling and monitoring without assuming intermediate tokens are faithful explanations. Overall, trajectory-based analysis offers a nuanced view of how reasoning traces shape outcomes and how cross-model interactions can be leveraged or mitigated in real-world pipelines.

Abstract

Large language models (LLMs) increasingly solve difficult problems by producing "reasoning traces" before emitting a final response. However, it remains unclear how accuracy and decision commitment evolve along a reasoning trajectory, and whether intermediate trace segments provide answer-relevant information beyond generic length or stylistic effects. Here, we propose a protocol to systematically probe the trajectories of reasoning traces in LLMs by 1) generating a model's reasoning trace, 2) truncating it at fixed token-percentiles, and 3) injecting each partial trace back into the model (or a different model) to measure the induced distribution over answer choices via next-token probabilities. We apply this protocol to the open-source Qwen3-4B/-8B/-14B and gpt-oss-20b/-120b models across the multiple-choice GPQA Diamond and MMLU-Pro benchmarks. We find that accuracy and decision commitment consistently increase as the percentage of provided reasoning tokens grows. These gains are primarily driven by relevant content in the model generation rather than context length or generic "reasoning style" effects. Stronger models often backtrack successfully from incorrect partial traces, but immediate answers often remain anchored in the weaker model's incorrect response. More broadly, we show that trajectory probing provides diagnostics for efficient and safer deployment of reasoning models as the measurements can inform practical trace-handling and monitoring policies that improve reliability without assuming intermediate tokens are inherently faithful explanations.

Figures (20)

• Figure 1: Overview of our protocol, which probes the trajectory of LLM reasoning traces when solving the GPQA Diamond and MMLU-Pro benchmarks. The probing protocol consists of three steps: 1., Generate the full reasoning trace of an LLM. 2., Slice each reasoning trace in deciles with respect to the total token length. 3., Construct a probing prompt that includes the system instruction, original question, injected reasoning slice, and an early stopping suffix (e.g., </think>). We analyze properties of the reasoning dynamics and the probed final responses, and we control for the effect of length, trace-form and token-identity. Finally, we investigate whether a target model (stronger) can rescue a reasoning trace that led to an incorrect answer in the initial model by either responding immediately or after being allowed to continue and complete the reasoning.
• Figure 2: Accuracy and decision commitment increase overall with reasoning depth. Accuracy and decision dynamics for Qwen3-4B (dark blue), Qwen3-8B (light blue), Qwen3-14B (green), gpt-oss-20b (orange), and gpt-oss-120b (red) on GPQA Diamond ($n{=}198$ questions) and MMLU-Pro ($n{=}12{,}032$ questions), averaged over 3 independent runs and stratified by reasoning decile (0--100%). a,b, Accuracy increases overall with reasoning decile, with steeper gains in later deciles (especially for the gpt-oss family). The exception is Qwen3-8B, which shows a substantial decline going from decile 90 to 100 (see \ref{['fig:appendix_non_choice']} and \ref{['tab:boxed-collapse-full']}). c,d, Probability assigned to the eventual final response rises throughout, reflecting growing decision commitment as reasoning progresses. e,f, Non-choice probability (mass on tokens other than answer letters) declines with decile for the Qwen3 family (and is always low for the gpt-oss family), indicating models increasingly commit to valid answer choices. The exception is Qwen3-8B, which shows elevated non-choice probability going from decile 90 to 100 due to frequent \\ boxed{} formatting in its traces (see \ref{['fig:appendix_non_choice']} and \ref{['tab:boxed-collapse-full']}). g,h, Flip rate (probability of switching the argmax answer relative to the previous decile) generally decreases with reasoning depth for MMLU-Pro and remains mostly stable (except for deciles 10 and 100) in GPQA Diamond.
• Figure 3: Instance-specific semantic reasoning drives accuracy gains. Comparison of original reasoning traces with three length-matched controls for Qwen3-4B (dark blue), Qwen3-8B (light blue), Qwen3-14B (green), gpt-oss-20b (orange), and gpt-oss-120b (red), on GPQA Diamond ($n{=}198$ questions) and MMLU-Pro ($n{=}12{,}032$ questions), averaged over 3 independent runs. Dots mark baseline accuracy at decile 0 (left) and original full-trace accuracy at decile 100 (right). a,b, Original traces: accuracy rises progressively with reasoning depth (+18.1--33.1% gain over baseline), with larger models achieving higher final accuracy. c,d, Random control (length-matched random token sequences): accuracy stays near baseline ($-$3.9% to +4.2% vs. d=0), confirming gains are not attributable to reasoning length alone. e,f, Swap control (another question's reasoning of matched length to control for trace-form): accuracy remains flat or declines ($-$18.4% to +3.0% vs. d=0), demonstrating that misaligned reasoning provides no benefit and can actively mislead. g,h, Shuffle control (same tokens in randomized order to control for token-identities): accuracy shows modest gains above baseline (+0.3--10.8% vs. d=0), indicating that while coherent sequential structure is critical, some lexical signal persists in the bag-of-tokens. Note that in \ref{['tab:summary-statistics']}, we show that all the accuracy differences at decile 100 between the original and control reasoning traces are statistically significantly different at $p < 0.001$ following the McNemar's test.
• Figure 4: Stronger models can rescue incorrect reasoning, and free continuation substantially improves recovery. Rescue rate (probability that a target model answers correctly when given an initial model's partial trace that led the initial model to an incorrect response) across deciles 20%, 40%, 60%, and 80%, averaged over 3 runs for GPQA Diamond and a single run for MMLU-Pro. Rows show initial models (Qwen3-4B, Qwen3-8B, Qwen3-14B, gpt-oss-20b) and columns show target models (same plus gpt-oss-120b). a, In base mode (answer now) for GPQA Diamond, we see that rescue rates in decile 20 range from 20.8--34.1%, while from 16.4--24.6% in decile 80, indicating that models generally have difficulty to recover from incorrect traces in base mode, especially when reasoning traces get longer. b, In free mode (allowing continuation) for GPQA Diamond, we see that rescue rates in decile 20 range from 42.9--68.1%, while from 25.5--50.3% in decile 80, indicating that models are able to substantially rescue incorrect reasoning with free continuation, especially when reasoning traces are still comparatively short. c, In base mode for MMLU-Pro, we see that rescue rates in decile 20 range from 13.9--29.5%, while from 9.2--23.9% in decile 80, confirming the findings in GPQA Diamond. d, In free mode for MMLU-Pro, we see that rescue rates in decile 20 range from 46.4--69.4%, while from 21.2--45.4% in decile 80, confirming the findings in GPQA Diamond. Across both benchmarks, stronger target models consistently achieve higher rescue rates, and free-mode continuation provides substantial benefits. Note that in \ref{['tab:rescue-summary']}, we show that the (pooled) average differences across all deciles between base and free rescue rates are statistically significantly different from each other at $p < 0.001$ (following McNemar's test) for all model pairs.
• Figure A1: Alternative early-stopping prompt for Qwen3 family yields consistent results. Replication of main analysis (\ref{['fig:main_2']}) using an alternative early-stopping suffix ("\\ n</think>\\ n\\ n") for Qwen3-4B (dark blue), Qwen3-8B (light blue), and Qwen3-14B (green) on GPQA Diamond and MMLU-Pro, averaged over 3 runs. a,b, Accuracy by decile. c,d, Probability on final response (decision commitment). e,f, Non-choice probability. g,h, Flip rate. All metrics exhibit the same qualitative patterns as the main analysis, confirming that results are robust to the specific phrasing of the early-stopping prompt.