Does the way you phrase, frame, and sound out a request matter less as AI models get bigger?
Do form, evidentiality, and tone interact with the size effect?
This explores whether the three levers of a message (its form, its evidentiality and its tone) work differently depending on model size, and how big those effects are next to other factors. The corpus has no study that crosses all three, so this reads across adjacent findings.
This explores whether form, evidentiality and tone change with model size, and how big they are next to other factors. No note in the collection tests that interaction directly. Read side by side, though, the notes suggest a pattern: humans keep responding to these surface signals, while bigger models lean on them less.
**Form.** Structure often does the work that content seems to do. Chain-of-thought examples with broken logic Does logical validity actually drive chain-of-thought gains? performed nearly as well as valid ones, so the model was picking up the shape of reasoning, not the inference. In persuasion, wrapping a claim as a presupposition instead of stating it outright Why are presuppositions more persuasive than direct assertions? persuades more, especially for new information, because it skips scrutiny. Size does interact with form. The best chain-of-thought length follows an inverted U, and it gets shorter as models get more capable Why does chain of thought accuracy eventually decline with length?. A trillion-parameter model found self-verification habits on its own that a 104B model needed hand-built rewards to learn Does scale alone teach models to reason without hand-crafted rewards?. Scale can also be replaced: sub-2B models trained on the right supervision format learned to ground answers in passages Can small models learn to ground answers in context?.
**Evidentiality.** Evidence works as a signal on the human side. Users preferred responses with more citations even when the citations were irrelevant, and irrelevant ones helped almost as much as relevant ones Do users trust citations more when there are simply more of them?. On the model side, larger and instruction-tuned models followed a user's stated beliefs less often when those beliefs clashed with world knowledge Do larger models follow stated beliefs less often?. That is a hint, not proof, that in-context claims carry less weight as models grow. The small-model grounding result above shows this can be trained in either direction.
**Tone.** Conviction is the clearest case. LLMs win persuasion partly by sounding more certain, and that helps whether or not the claim is true Does linguistic conviction explain why LLMs persuade more effectively?. The note traces this assertive register to RLHF, not to scale, so tone looks like a training artifact more than a size effect. A separate finding points the other way for models as listeners. Higher model confidence, which tends to come with larger models, predicts steadier answers when a prompt is reworded Does model confidence predict robustness to prompt changes?. Bigger models therefore seem less swayed by phrasing.
**The reader can outweigh all three.** In debate data, voters' political and religious ideology predicted outcomes better than linguistic features did. Language effects that look real without controlling for the audience can be confounded by who was listening Does what readers believe matter more than what debaters say?. If size means how big the effect is, form, evidentiality and tone may be modest next to what the audience already believes.
The open gap is that nobody in this set varies model size and message style together. What the notes support is narrower. Bigger models seem harder to move with surface polish, while people still respond to it, and RLHF-trained models supply that polish through confident tone and citations.
Sources 10 notes
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
Experimental evidence shows presuppositions with additive, iterative, and factive triggers persuade audiences more than assertions, especially for discourse-new content. The mechanism: presuppositions bypass evaluative scrutiny by presenting claims as already-accepted background.
Task accuracy peaks at intermediate CoT length, with optimal length increasing alongside task difficulty but decreasing with model capability. RL training naturally gravitates toward shorter chains as models improve, revealing that simplicity emerges from reward signals rather than explicit training.
Ring-Zero found a scale threshold where pure zero RL becomes sufficient: a 104B model required hand-designed rewards for structured reasoning and self-verification, but a 1T model discovered these strategies autonomously. This suggests reasoning-scaffolding research has value tied to model size.
Sub-2B models trained on synthetic multi-hop QA can ground answers in passages, cite literal quotes, and abstain from confabulation. The OCC-RAG work shows faithfulness emerges from training curriculum design, not parameter count.
Show all 10 sources
Analysis of 24,000 Search Arena interactions shows irrelevant citations boost user preference (β=0.273) nearly as much as relevant citations (β=0.285), indicating citation count functions as a decoupled trust heuristic.
Across 18 LLMs tested with EoBench, bigger models and instruction-tuned variants showed lower rates of context-following when users expressed beliefs that contradicted world knowledge. The effect suggests instruction-tuning strengthens reliance on parametric knowledge.
Linguistic analysis shows LLMs express higher conviction than human persuaders, and this confidence-loading directly correlates with persuasive outcomes regardless of whether claims are true or false. RLHF training installs an assertive register that functions as a content-independent persuasion amplifier.
ProSA found that when models are highly confident, they resist prompt rephrasing; low confidence causes major output swings. Larger models, few-shot examples, and objective tasks all correlate with higher confidence and greater robustness.
Analysis of debate corpora shows that political and religious ideology labels of voters outpredict linguistic features when modeling debate outcomes. Language effects observed without reader controls are confounded by audience composition correlated with debate topics.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Exploring the Role of Prior Beliefs for Argument Persuasion
- When More is Less: Understanding Chain-of-Thought Length in LLMs
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens
- Measuring Faithfulness in Chain-of-Thought Reasoning
- Presuppositions are more persuasive than assertions if addressees accommodate them: Experimental evidence for philosophical reasoning
- VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs
- ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning