Large Language Models Are Overconfident in Their Own Responses (the chat-template "ownership bias")

TL;DR. Instruction-tuned LLMs are known to be worse calibrated than their base pretrained versions (calibration = how well a stated confidence matches the actual chance of being right). This paper decouples two causes that were tangled together: the post-training algorithm and the chat template itself. It finds the chat template aggravates miscalibration through an ownership bias — a model is significantly more confident in an answer when that answer is framed as its own than when the identical answer is framed as coming from the user. Across six open-weight LLMs, three benchmarks, and three confidence-elicitation methods, models assign up to 26% higher confidence to their own responses. The fix is a free inference-time trick: frame the model's own answer as user input when eliciting confidence, which cuts overconfidence and improves calibration by up to 26% with no retraining.

Source: HuggingFace Daily Papers · arxiv 2606.03437

flowchart LR
  ANS[A candidate answer] --> OWN[Framed as model's<br/>own response]
  ANS --> USR[Framed as user-<br/>provided response]
  OWN --> HI[Up to 26% higher<br/>stated confidence]
  USR --> LO[Lower, better-<br/>calibrated confidence]
  HI -->|gap = ownership bias| FIX{Inference-time fix<br/>re-frame own answer as user input}
  FIX --> CAL[Calibration improves<br/>up to 26%, no retraining]
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  class ANS input
  class FIX decision
  class LO,CAL output
  class HI,OWN warn

Key findings

Two causes separated. Instruction tuning fundamentally harms calibration (known); the chat template additionally harms it through ownership bias (new). The contribution is showing the template is an independent lever.
Ownership bias is large and consistent. Up to 26% higher confidence on a model's own answer versus the identical answer attributed to the user, robust across six models, three benchmarks, three elicitation methods.
A free fix. Re-framing the model's answer as user input during confidence elicitation narrows the base-vs-instruction-tuned calibration gap by up to 26% with no retraining.

How this relates to prior wiki knowledge

This is the third strand this week pointing at the same uncomfortable place: an instruction-tuned model's confidence is partly social, not epistemic. The wiki's responsible-ai page has been tracking a sycophancy/honesty cluster — yesterday's Emergent Misalignment via Sycophancy (06-10, training a model to agree with wrong users induces broad misalignment, with a locatable representational gate), plus Kurate's "LLMs Know They're Wrong and Agree Anyway: The Shared Sycophancy-Lying Circuit" (cs.LG #15) and "Value-Conflict Diagnostics Reveal Widespread Alignment Faking" (cs.AI #13). Ownership bias is the calibration-side companion: where sycophancy is deferring to the user's wrong belief, ownership bias is over-trusting the model's own belief. Both are the chat template injecting a who-said-it signal into what should be a purely epistemic confidence estimate.

The fix is also methodologically aligned with the wiki's "soft, structured, cheap" preference: it is an inference-time reframing, not a retrain — the same class of move as yesterday's QK-Restore (a checkpoint diff, not a retrain) and the inference-time calibration tricks the digest has favored.

Research angle. The mechanism is the prize. If ownership bias lives in the chat-template tokens (the role markers that say "assistant"), it should be locatable with exactly the representation tools in this week's batch — Alignment Gating and today's ICA Lens. A clean follow-up: find the direction that encodes "this text is mine," and check whether suppressing it removes ownership bias and dents sycophancy, which would tie the two failures to one representational cause. For agentic systems this matters directly: a self-evaluating agent (the verified-improvement gates in today's Arbor) is scoring its own outputs, exactly the condition under which ownership bias inflates confidence.

→ Raw: raw/huggingface/2026-06-11-large-language-models-are-overconfident-in-their-own-respons.md