Can confidence patterns reveal overthinking versus underthinking?
This explores whether real-time confidence signals can diagnose when a reasoning model is trapped in redundant deliberation versus committing prematurely, and whether steering based on these signals can balance both failure modes.
Overthinking and underthinking are dual failures, and existing methods that suppress one often induce the other. Suppressing reflective keywords or truncating reasoning length reduces overthinking but causes underthinking — the model doesn't explore enough. Forcing longer chains reduces underthinking but generates redundancy. ReBalance resolves this by treating confidence as a continuous diagnostic signal rather than using binary interventions.
The diagnostic: Confidence values correlate with reasoning behavior in interpretable ways:
- High confidence variance — frequent indecisive switching between reasoning paths, causing redundant steps and delayed convergence. This IS overthinking: the model knows something is wrong but can't commit.
- Consistent overconfidence — premature commitment to an incorrect reasoning path. This IS underthinking: the model commits too early without adequate exploration.
The mechanism: From a small-scale dataset, identify reasoning steps indicating each mode. Aggregate their hidden states into reasoning mode prototypes. Compute a steering vector encoding the transition from overthinking to underthinking. A dynamic control function modulates the vector's strength and direction based on real-time confidence: pruning redundancy during overthinking, promoting exploration during underthinking.
Why it's training-free: The steering vector captures the model's inherent reasoning dynamics — it's extracted from the model's own hidden states, not trained. Because it operates on intrinsic representations, it generalizes across unseen data and tasks (math, QA, coding). This makes it plug-and-play across models from 0.5B to 32B.
Since Can we steer reasoning toward brevity without retraining?, ReBalance extends the activation-steering approach from length compression to reasoning quality management. ASC steers between verbose and concise modes; ReBalance steers between overthinking and underthinking — a qualitative distinction, not just quantitative.
Since Does more thinking time always improve reasoning accuracy?, ReBalance provides the dynamic mechanism the threshold finding calls for: instead of a fixed cutoff, confidence-based steering continuously adjusts the reasoning trajectory.
Inquiring lines that read this note 77
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How should models express uncertainty rather than forced confident answers?- Can we measure sophistry by tracking conviction density in model outputs?
- Why do weak belief tracking and conservative actions trap agents in low-information states?
- Why do linguistic hedging markers correlate with internal confidence signals in reasoning traces?
- Can agents escape weak belief tracking and conservative action selection traps?
- Does distillation strip away uncertainty signals that reasoning actually needs?
- Can penalizing reasoning transitions fix underthinking without fine-tuning models?
- Can activation-space steering vectors replicate thinking model performance without retraining?
- Why does per-step deliberation lose global perspective compared to dynamic discovery?
- When should action deliberation trigger during reasoning steps?
- Do high-influence thoughts align with SAND deliberation triggers?
- Can high-entropy tokens and step-level confidence identify the same critical reasoning forks?
- Does performative reasoning mask underlying uncertainty even on easy problems?
- What distinguishes metacognitive regulation from standard chain-of-thought reasoning?
- How does step-level confidence filtering compare to global confidence averaging?
- How do we assign confidence and polarity scores to belief edges?
- How reliable is the top-2 confidence gap as a stopping signal across tasks?
- What does it mean when a user's signal has low confidence?
- What role does confidence play in balancing overthinking versus underthinking?
- What makes accurate confidence different from confident-but-wrong predictions?
- How does model confidence relate to exemplar brittleness in chain-of-thought?
- Can intrinsic confidence signals improve both calibration and reasoning performance?
- Can confidence levels reliably detect when a model is overthinking?
- Can runtime confidence signals detect when reasoning has crossed the overthinking threshold?
- Can step-level confidence filtering work better than global confidence scoring?
- How do miscalibrated confidence signals affect the success of SmartPause routing?
- Can calibrated confidence reduce misleading consensus in group deliberation?
- Can log-probability confidence be separated from decision-aligned signals?
- What signals should systems use to predict the right moment for intervention?
- Can real-time detection identify when users have incomplete or underdeveloped intent?
- How do we measure the cognitive flow cost of different intervention strategies?
- How does timing AI assistance based on cognitive signals affect user autonomy?
- Why is metacognition neglected as a foundational AI research area?
- What triggers overthinking versus underthinking in reasoning models?
- Does distillation from reasoning models spread overthinking to smaller models?
- Can extended deliberation in agents become counterproductive like human overthinking?
- Why do models overthink easy problems and underthink difficult ones?
- Why does overthinking degrade performance at extreme recursion depths?
- Do search agents face their own overthinking threshold like reasoning models do?
- Can models overthink and underthink at the same time?
- Why do different model training approaches produce different overthinking thresholds?
- Can conditioning generation on difficulty probes reduce overthinking on simple tasks?
- What happens when models overthink during test-time search?
- How does self-revision in reasoning chains amplify confidence in wrong answers?
- Why does single-model self-revision amplify confidence in incorrect answers?
- Why does single-agent self-revision amplify confidence in wrong answers over time?
- Why do reasoning models amplify confidence in incorrect answers during self-revision?
- Why do reasoning models exhibit self-doubt about their own early assessments?
- How does self-distillation degrade reasoning by suppressing uncertainty signals?
- Can AI distinguish when validation helps versus when confrontation is needed?
- Can metacognitive categories be learned instead of fixed by human designers?
- Does high model confidence increase the risk of human overreliance?
- What role does real-time accuracy feedback play in reducing user overreliance?
- Why do models overthink underspecified problems instead of rejecting them?
- Can layer-wise prediction stabilization identify when genuine reasoning has stopped?
- How does Self-Discover compare to the cognitive tools approach?
- How do surface signals like confidence override actual quality in user judgment?
- Why is confidence a dangerous proxy for accuracy in human-AI interaction?
Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can we steer reasoning toward brevity without retraining?
This explores whether model reasoning style occupies learnable geometric directions in activation space, and whether we can shift toward concise thinking by steering through that space without expensive retraining.
ASC compresses length; ReBalance steers reasoning quality
-
Does more thinking time always improve reasoning accuracy?
Explores whether extending a model's thinking tokens linearly improves performance, or if there's a point beyond which additional reasoning becomes counterproductive.
ReBalance provides the dynamic mechanism: confidence-based steering vs fixed threshold
-
Do reasoning models switch between ideas too frequently?
Research explores whether o1-like models abandon promising reasoning paths prematurely by switching to different approaches without sufficient depth, and whether penalizing such transitions could improve accuracy.
complementary: underthinking penalty addresses premature switching; ReBalance addresses premature commitment via overconfidence detection
-
When should an agent actually stop and deliberate?
How can models detect when deliberation over action choices is genuinely needed versus wasteful? This matters because unbounded action spaces make universal deliberation intractable, yet skipping it entirely risks missing critical errors.
both use uncertainty/confidence as the trigger for compute allocation
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Efficient Reasoning with Balanced Thinking
- Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models
- Understanding and Mitigating Premature Confidence for Better LLM Reasoning
- Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought
- Missing Premise exacerbates Overthinking: Are Reasoning Models losing Critical Thinking Skill?
- Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties
- A Comment On "The Illusion of Thinking": Reframing the Reasoning Cliff as an Agentic Gap
- Test-time Prompt Intervention
Original note title
ReBalance uses confidence as continuous indicator to dynamically steer between overthinking and underthinking — training-free balanced reasoning via hidden state steering vectors