When Shallow Wins: Silent Failures and the Depth-Accuracy Paradox in Latent Reasoning

• URL: http://arxiv.org/abs/2603.03475v1
• Date: Tue, 03 Mar 2026 19:43:36 GMT
• Title: When Shallow Wins: Silent Failures and the Depth-Accuracy Paradox in Latent Reasoning
• Abstract summary: We demonstrate that state-of-the-art models (Qwen2.5-Math-7B) achieve 61% accuracy through a mixture of reliable and unreliable predictions.<n>We show that 18.4% of correct predictions employ stable, faithful reasoning while 81.6% emerge through computationally inconsistent pathways.
• Score: 16.505918019260964
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Mathematical reasoning models are widely deployed in education, automated tutoring, and decision support systems despite exhibiting fundamental computational instabilities. We demonstrate that state-of-the-art models (Qwen2.5-Math-7B) achieve 61% accuracy through a mixture of reliable and unreliable reasoning pathways: 18.4% of correct predictions employ stable, faithful reasoning while 81.6% emerge through computationally inconsistent pathways. Additionally, 8.8% of all predictions are silent failures -- confident yet incorrect outputs. Through comprehensive analysis using novel faithfulness metrics, we reveal: (1) reasoning quality shows weak negative correlation with correctness (r=-0.21, p=0.002), reflecting a binary classification threshold artifact rather than a monotonic inverse relationship; (2) scaling from 1.5B to 7B parameters (4.7x increase) provides zero accuracy benefit on our evaluated subset (6% of GSM8K), requiring validation on the complete benchmark; and (3) latent reasoning employs diverse computational strategies, with ~20% sharing CoT-like patterns. These findings highlight that benchmark accuracy can mask computational unreliability, demanding evaluation reforms measuring stability beyond single-sample metrics.

Related papers

• State of Thought Enables Endogenous Reasoning [48.902117018115256]
  Test-time reasoning has emerged as a major approach to improving the capabilities of Large Language Models (LLMs)<n>We propose State of Thought, a new reasoning paradigm that enables endogenous reasoning in LLMs.<n>We show that SoT consistently improves mean-baseline accuracy while reducing generated tokens by 62.6% and end-to-end latency by 44.6%.
  arXiv  Detail & Related papers  (2026-09-13T05:55:03Z)
• Calibration-Aware Uncertainty Cascades for Efficient Heterogeneous Model Collaboration [18.045450334322545]
  Heterogeneous model collaboration seeks to exploit the complementary strengths of different models to balance predictive performance and inference cost.<n>Existing approaches typically rely on trained routers, which tie routing decisions to a fixed task and model pool, or on raw-confidence cascades, whose thresholds lack consistent reliability semantics across heterogeneous models.<n>We propose-Aware Uncertainty Cascades (CAUC), a simple post-hoc framework that independently calibrates each model's confidence and selects deployment policies using validation data.
  arXiv  Detail & Related papers  (2026-09-10T12:14:54Z)
• Cross-Fitted Residual Utility for Primary-Preserving Cognitive Decision Correction in Automatic Modulation Classification [1.547549252134621]
  We study a post-inference problem through cross-fitted residual utility and a primary-preserving cognitive decision policy.<n>A structured KAN-Fourier classifier supplies the default probability, while neural and non-neural candidates provide observable evidence.<n>On RMLA, RMLB, and HISAR, the complete system improves overall accuracy from 63.632% to 66.332%, 65.161% to 66.168%, and 77.769% to 79.867%.
  arXiv  Detail & Related papers  (2026-08-03T11:07:37Z)
• Automated Proving of Shannon-Type Entropy Inequalities via Fine-Tuned Language Models and Guided Tree Search [50.16356451328644]
  Proving Shannon-type entropy inequalities is a fundamental task in information theory.<n>We investigate whether small-scale large language models, fine-tuned on atomic proof steps, can automate this process.<n>We show that GPT-5.5 solves 1.7% samples under zero-shot prompting while Psitip solves 33.3% samples.
  arXiv  Detail & Related papers  (2026-06-04T05:43:12Z)
• Unveiling the Entropy Dynamics of Chain-of-Thought Reasoning [81.57028614960576]
  We show that the Confidence Region possesses two critical properties: 1) High Reliability -- answers in the confidence region become highly accurate and stable, and 2) High Redundancy -- models generate unnecessary tokens long after reaching the correct answer.<n>These properties unlock more efficient and reliable inference strategies.
  arXiv  Detail & Related papers  (2026-06-01T10:11:14Z)
• What Single-Prompt Accuracy Misses: A Multi-Variant Reliability Audit of Language Models [0.0]
  Single-prompt accuracy is the dominant way to benchmark language models, but it can miss reliability failures that matter.<n>We evaluate a 15-model open-weight corpus, with the main reliability analyses focused on 10 instruct models across five classification and reasoning benchmarks.<n>We find three broad results. First, evaluation design can materially change the conclusion.<n>Second, confidence signals are fragile. On MMLU-Pro, every primary model verbally reports confidence substantially above both its accuracy and its token-probability confidence on the same rows, and verbal parse rate can collapse for a single model on a single prompt variant.
  arXiv  Detail & Related papers  (2026-05-03T20:05:08Z)
• Correct Answers from Sound Reasoning: Verifiable Process Supervision for Language Models [94.68358825189738]
  We propose a post-training framework for verifiable domains that jointly optimize prediction accuracy and reasoning quality.<n>We evaluate VPS on chess, a controlled testbed where reasoning steps can be deterministically verified against engine signals.<n>VPS preserves accuracy while significantly improving reasoning quality, reducing win-rate error by up to 30% and restoring consistency to near saturation.
  arXiv  Detail & Related papers  (2026-04-03T15:19:46Z)
• Black-Box Reliability Certification for AI Agents via Self-Consistency Sampling and Conformal Calibration [0.0]
  We answer with a reliability level -- a single number per system-task pair.<n>Self-consistency sampling reduces uncertainty exponentially.<n> conformal calibration guarantees correctness within 1/(n+1) of the target level.
  arXiv  Detail & Related papers  (2026-02-24T21:03:50Z)
• Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models [108.26461635308796]
  We introduce Rationale Consistency, a fine-grained metric that quantifies the alignment between the model's reasoning process and human judgment.<n>Our evaluation of frontier models reveals that rationale consistency effectively discriminates among state-of-the-art models.<n>We introduce a hybrid signal that combines rationale consistency with outcome accuracy for GenRM training.
  arXiv  Detail & Related papers  (2026-02-04T15:24:52Z)
• When Small Models Are Right for Wrong Reasons: Process Verification for Trustworthy Agents [0.0]
  We reveal a critical reliability crisis: 50-69% of correct answers from small language models contain fundamentally flawed reasoning.<n>We introduce the Reasoning Integrity Score (RIS), a process-based metric validated with substantial inter-rater agreement.<n>We show RAG succeeds by grounding calculations in external evidence, reducing errors by 7.6%, while meta-cognition amplifies confusion without sufficient model capacity.
  arXiv  Detail & Related papers  (2026-01-01T23:54:15Z)
• d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models [45.27333046908981]
  emphd-TreeRPO is a reliable reinforcement learning framework for dLLMs.<n>We show that emphd-TreeRPO achieves significant gains on multiple reasoning benchmarks.
  arXiv  Detail & Related papers  (2025-12-10T14:20:07Z)
• Towards a Science of Scaling Agent Systems [79.64446272302287]
  We formalize a definition for agent evaluation and characterize scaling laws as the interplay between agent quantity, coordination structure, modelic, and task properties.<n>We derive a predictive model using coordination metrics, that cross-validated R2=0, enabling prediction on unseen task domains.<n>We identify three effects: (1) a tool-coordination trade-off: under fixed computational budgets, tool-heavy tasks suffer disproportionately from multi-agent overhead, and (2) a capability saturation: coordination yields diminishing or negative returns once single-agent baselines exceed 45%.
  arXiv  Detail & Related papers  (2025-12-09T06:52:21Z)
• Geometric Calibration and Neutral Zones for Uncertainty-Aware Multi-Class Classification [0.0]
  This work bridges information geometry and statistical learning, offering formal guarantees for uncertainty-aware classification in applications requiring rigorous validation.<n> Empirical validation on Adeno-Associated Virus classification demonstrates that the two-stage framework captures 72.5% of errors while deferring 34.5% of samples, reducing automated decision error rates from 16.8% to 6.9%.
  arXiv  Detail & Related papers  (2025-11-26T01:29:49Z)
• Parrot: Persuasion and Agreement Robustness Rating of Output Truth -- A Sycophancy Robustness Benchmark for LLMs [0.0]
  PARROT (Persuasion and Agreement Robustness Rating of Output Truth) is a robustness focused framework designed to measure the degradation in accuracy under social pressure exerted on users.<n>We evaluate 22 models using 1,302 MMLU-style multiple-choice questions across 13 domains and domain-specific authority templates.
  arXiv  Detail & Related papers  (2025-11-21T13:01:28Z)
• Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression [68.69801176669843]
  We propose an online post-training RL method that prunes redundant steps and estimates difficulty.<n> TRAAC (Think Right with Adaptive, Attentive Compression) achieves an average absolute accuracy gain of 8.4%.<n>Although our models are trained on math datasets, they show accuracy and efficiency gains on out-of-distribution non-math datasets.
  arXiv  Detail & Related papers  (2025-10-02T02:00:20Z)
• TrustJudge: Inconsistencies of LLM-as-a-Judge and How to Alleviate Them [58.04324690859212]
  Large Language Models (LLMs) as automated evaluators (LLM-as-a-judge) has revealed critical inconsistencies in current evaluation frameworks.<n>We identify two fundamental types of inconsistencies: Score-Comparison Inconsistency and Pairwise Transitivity Inconsistency.<n>We propose TrustJudge, a probabilistic framework that addresses these limitations through two key innovations.
  arXiv  Detail & Related papers  (2025-09-25T13:04:29Z)
• Benchmarking Reasoning Robustness in Large Language Models [76.79744000300363]
  We find significant performance degradation on novel or incomplete data.<n>These findings highlight the reliance on recall over rigorous logical inference.<n>This paper introduces a novel benchmark, termed as Math-RoB, that exploits hallucinations triggered by missing information to expose reasoning gaps.
  arXiv  Detail & Related papers  (2025-03-06T15:36:06Z)
• Common 7B Language Models Already Possess Strong Math Capabilities [61.61442513067561]
  This paper shows that the LLaMA-2 7B model with common pre-training already exhibits strong mathematical abilities. The potential for extensive scaling is constrained by the scarcity of publicly available math questions.
  arXiv  Detail & Related papers  (2024-03-07T18:00:40Z)
• Don't Just Blame Over-parametrization for Over-confidence: Theoretical Analysis of Calibration in Binary Classification [58.03725169462616]
  We show theoretically that over-parametrization is not the only reason for over-confidence. We prove that logistic regression is inherently over-confident, in the realizable, under-parametrized setting. Perhaps surprisingly, we also show that over-confidence is not always the case.
  arXiv  Detail & Related papers  (2021-02-15T21:38:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.