The Two Sides of Reasoning Traces: Risks of Overexplaining

I. Research background

Problem and importance

As LLMs enter human decision workflows, a central question emerges: Does showing model reasoning actually help users make better decisions?

Since Wei et al.'s 2022 “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” CoT has become standard for improving reasoning. A natural conjecture follows: if full reasoning helps models reason, might showing these traces help users understand and verify answers?

However, this assumption faces a fundamental challenge: a large gap may separate displaying reasons from understanding them. Human understanding differs from internal model computation. The core paper systematically tests that gap.

Field status and disagreements

Important debates include:

• Cognitive cost of traces: Do full steps overload users, or provide useful transparency?
• Calibration and confidence: Do traces improve assessment of answer correctness, or merely inflate confidence?
• Trace–performance decoupling: Why can more reasoning steps worsen user performance, as found in 2505.13792?
• Test-time-compute limits: Is more reasoning always the best capability-scaling route, as questioned by 2604.10739?

Three main research directions

• Causal trace–performance relationships: Randomized controlled trials compare shown versus hidden reasoning for human decisions.
• Verifiers and reward models: Learned verifiers replace intuitive human judgments; test-time scaling/process supervision optimize reasoning quality.
• Internal reasoning mechanisms: Activations, gradients, and tokens reveal whether traces reflect actual model computation.

II. Core paper deep dive

Paper information

Title: Explaining Too Much? Understanding How LLM Reasoning Traces Influence Performance and Metacognition
Authors: Fernandes, Buschek, Tankelevitch, Kosch, Welsch
Publication date: May 25, 2026 (arXiv:2605.25856)
Study type: Preregistered human-subject empirical study

Abstract and findings

Narrative and methodology

Clear comparisons and preregistration underpin this study. Its argument develops as follows:

Step 1: Origins and premise checking

The paper identifies an intuitive but untested assumption: If CoT helps LLMs reason, showing it should help humans reason. This overlooks fundamentally different human/machine cognition. The novelty is treating it as an empirical question to test rather than an assumed fact.

Step 2: Key experimental variables

The design covers three core dimensions:

• Independent variable: Presentation format—answer only, full trace, condensed trace
• Dependent variable:
  • Objective performance: Actual answer accuracy
  • Subjective assessment: Confidence in answer correctness
  • Interface perception: Appeal, trust, and professionalism ratings
  • Calibration quality: Agreement between confidence and actual performance

Step 3: Preregistration and robustness

Hypotheses, questions, and analyses were registered before collection, an important transparency commitment that prevents p-hacking through selective significant-result reporting.

Key results and data

Performance metrics

Confidence and calibration (Confidence & Calibration)

Core finding: Participants substantially overestimate themselves in every condition:

• They claim 70–75% accuracy on average but achieve only 55–62%
• Full-trace participants perform worst yet feel most confident—the most dangerous mismatch
• No format improves self-assessment of knowledge

Interface-perception mechanism (Hedonic Appeal vs. Trust)

Mediation analysis reveals a surprising mechanism:

• Hedonic appeal, rather than genuine trust, primarily drives false confidence
• Full traces look attractive/professional, directly increasing confidence ratings
• This confidence gain is entirely disconnected from actual performance improvement

Contributions and innovations

Main contributions:

1. Challenges intuition: Establishes that CoT benefiting LLMs does not imply benefiting humans, rejecting an uncritical analogy.
2. Quantifies false-confidence harm: Shows traces can create strongest confidence when performance is weakest, relevant to deployment safety.
3. Identifies a mechanism: Attributes false confidence to aesthetics rather than cognition, guiding targeted interventions.
4. Preregistration exemplar: Promotes rigorous preregistration at the HCI/AI-safety intersection.

III. Related work and technical lineage

Five key stages

Stage 1: Discovery of reasoning prompts (2022–2023)

Representative work: Wei et al. (2022)《Chain-of-Thought Prompting Elicits Reasoning in Large Language Models》

The pioneering finding was that stepwise reasoning rather than direct answers improves math/common-sense tasks. Its insight: Explicit intermediate steps help not only human understanding but model reasoning itself.

Implicit assumption then: Reasoning traces ≈ transparency, trustworthiness, usefulness.

Stage 2: Verifiers and process supervision (2023–2024)

Related work: Process Reward Models rise, with verification explored as a scaling axis, e.g. 2607.05391 “LLM-as-a-Verifier.”

Because outcome-only supervision is inefficient, researchers train Process Reward Models (PRMs) to assess each reasoning step, introducing the distinction: Not every step is correct, and plausible-looking steps are not always useful.

The shift: better step evaluation, not merely more steps.

Stage 3: Recognizing test-time-compute harm (2024–2026)

Representative work: 2604.10739《When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling》

An alarming finding emerges: More reasoning steps or computation can worsen performance on some tasks/models. Overthinking derails reasoning or creates erroneous loops.

This motivates the core paper: if excessive reasoning harms models, what happens when humans see it?

Stage 4: Structural trace–performance decoupling (2026)

Representative work: 2505.13792《Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation》

This work finds in knowledge distillation that highly interpretable-looking traces can worsen student performance. Why?

• Traces capture surface-level discrete steps, while actual model computation is continuous and high-dimensional interval's
• Human-understood logical flow differs from optimized gradient flow
• Apparently correct explanations can guide students toward suboptimal learning paths

Stage 5: User-side calibration (current, 2026)

Core paper(this report): 2605.25856, “Explaining Too Much?”

The focus extends from distillation to the human side. Not only model-to-model transfer can fail; humans also mistake apparent clarity for understanding. Traces become a confidence pump rather than an understanding aid.

2508.06225, “Overconfidence in LLM-as-a-Judge,” likewise finds widespread overconfidence when LLMs evaluate.

Unresolved questions along the lineage

Three key questions ranked by benefit / difficulty:

IV. Structured review and value analysis

A rigorous point-by-point review follows:

1. Omitted metrics and hidden failures

Full traces may require more reading yet worsen performance, a serious mismatch between effort and benefit. Equal-performing summaries taking less time would be superior. This omission may conceal an important engineering tradeoff.

High- and low-ability participants may react very differently: some learn from traces, others are misled. Missing subgroup analysis means conclusions do not cover everyone, despite generalizing language.

2. Parameters versus setup-free claims

Many implementations qualify as full traces. If summaries have unusually good formatting and full versions are crowded, differences may reflect layout rather than content quantity. Interface specifications or appendix screenshots are needed.

3. Are sources of effects isolated?

Poor full-trace performance may reflect these particular low-quality generated traces, not the trace concept itself. Would high-quality human explanations differ? Without this control, findings apply to LLM-generated traces, with uncertain applicability to ideal traces.

4. Curves rather than single-point metrics

Do users adapt over more questions, or does harm persist? If initial harm diminishes, long-term conclusions may differ. Initial performance may not represent users familiar with the interface.

5. Failures and edge cases

What traces arise for ambiguous or unsolvable real questions? Might they mislead especially strongly? The narrow task space limits practical generalization.

6. Do conclusions exceed evidence?

“Explaining Too Much?” suggests broadly harmful traces, but evidence covers only LSAT + English speakers + one particular LLM. Other tasks, languages, and models could differ substantially. Generality in title/text exceeds the evidence.

7. Reproducibility and audit transparency

Preregistration supports honest planning; full reproduction also needs:

• Exact questions and generated traces, potentially anonymized
• Raw participant questionnaire responses
• Accessible interface code/prototypes

Without these, replicators must guess many details.

Value analysis: Layers and durability

What the paper is not

❌ Not a study of LLM reasoning ability; model reasoning quality is not tested.
❌ Not a systematic trace-design guide; full/condensed designs are examined, not the full design space.
❌ Not a universal conclusion about explanations; only sequential reasoning displays, not attention visualization or concept-activation vectors.
❌ Not about every task; LSAT questions may be particularly vulnerable to step overload.

V. Papers and resource links

Core paper

Key related work by technical lineage

🏛️ Foundations: Discovery of reasoning prompts

🔴 Core connections: Test-time-compute harm and trace decoupling

🟢 Verifiers and process reward models

🟡 Calibration and overconfidence

Further reading

• Test-time-compute survey: GitHub repository Awesome-Inference-Time-Scaling
• Practical guide: Test-Time Compute in 2026: The Complete Practitioner's Guide
• LLM interpretability survey: A Gentle Primer on LLM Explainability - KDnuggets

VI. Verification notes and confirmed items

Overall assessment: Core problem, design, and directional findings are highly credible. Exact numbers/mechanistic details warrant moderate confidence because of PDF-access limits. Readers are strongly encouraged to consult the original paper for complete firsthand information.