Do larger models follow stated beliefs less often?
Does model size and instruction-tuning affect whether language models prioritize user-stated beliefs over their parametric knowledge? Understanding this matters for predicting when models will accept user context versus rely on training.
The paper introduces EoBench to test how a user's "expression of belief" (EoB) changes whether a model follows the context it was given or falls back on prior knowledge. Across 18 LLMs spanning Llama3, Qwen3 and Gemma3, at 1B to 30B parameters, in base and instruct variants, the abstract reports that "bigger models and instruction-tuned models tend to be less context-following than smaller models and base models." The discussion restates both trends: instruction-tuning reduces the context-following rate, and so does increasing model size. The framing is deliberately two-sided. In some situations the ideal is for the model to accept the user's stated belief as true, and in others it is to stick to what it knows.
The benchmark is built to isolate phrasing. It pairs EoBs with world-knowledge facts to produce controlled EoB–query pairs, using a typology of four linguistically motivated dimensions (form, evidentiality, epistemic stance, tone) spanning 19 fine-grained types. The introduction's example is the same false claim delivered two ways: stated outright, "The Eiffel Tower is located in Berlin," or presupposed through a question, "When was the Eiffel Tower relocated to Berlin?" On the paper's explanation, instruction-tuning "directly incentivizes models to attend more strongly to in-context instructions or demonstrations." For size the authors say the cause "is less clear" and speculate that larger parameter counts "provide more storage for parametric knowledge and thus stronger priors."
The size result fits Do large language models reason symbolically or semantically?, where counter-commonsense rules in context lose to parametric commonsense. EoBench adds a scale and training-stage axis to that picture of prior knowledge beating context. It sits close to Why do language models fact-check instead of confirming beliefs?, where models asked to acknowledge a belief default to fact-checking the embedded claim. Both describe models that lean on what they know over what the user said, and the aggregate trend here is compatible with that account. The two papers measure different things, though. One tests belief acknowledgment across verbs, and the other tests context versus prior knowledge across expression types, so this note does not claim they show the same mechanism.
The excerpt leaves most of the title's promise open. It does not say which of the 19 types are more or less persuasive, or whether form, evidentiality, stance and tone interact with the size and training-stage trends. It gives no effect sizes, no per-family breakdown, no description of the fact set and no metric definition, and "tend to" signals a trend rather than a rule. The instruction-tuning explanation is also stated without saying why stronger attention to instructions would lower following of a stated belief. The size explanation is labeled speculation, and the authors point to causal abstraction and mechanistic interpretability as the route to a real account. What the excerpt does support is narrow but usable: a bigger or more heavily instruction-tuned model is not automatically the more accommodating one, and whether lower context-following counts as a failure depends on whether the user's belief should have been accepted.
Inquiring lines that read this note 26
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 do training data properties affect model generalization and robustness? Why does fine-tuning for specific tasks degrade broader reasoning capability?- Does semantic auditing of instruction data improve performance uniformly across different model sizes?
- Do larger instruction-tuned models actually track user beliefs better?
- Does instruction tuning affect how much models yield to disagreement?
- Which phrasing types most persuade models to accept stated beliefs?
- How do stated confidence and actual correctness diverge in language models?
- Why do models drop correct beliefs when users challenge them?
- Do larger models resist user beliefs more or less than smaller ones?
- Is lower context-following a failure or appropriate model behavior?
- Why does instruction-tuning reduce a model's context-following behavior?
- Can large instruction sets damage a model's ability to follow behavioral rules?
- Do larger language models show stronger self-preference in evaluation tasks?
- Does statistical learning in language models predictably favor central tendencies over rare expressions?
- How well-calibrated are language models when making clinical predictions?
- What causal relationship exists between model weights and tacit knowledge in language models?
Related concepts in this collection 3
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
-
Do large language models reason symbolically or semantically?
Can LLMs follow explicit logical rules when those rules contradict their training knowledge? Testing whether reasoning operates independently of semantic associations reveals what computational mechanisms actually drive LLM multi-step inference.
parametric knowledge overriding in-context content, which EoBench extends with model size and training stage as axes
-
Why do language models fact-check instead of confirming beliefs?
When asked to confirm a stated belief about false information, do LLMs struggle because they default to evaluating the claim's truth rather than acknowledging the user's stance? How much does the belief verb shape this behavior?
neighboring belief-handling result where models fact-check instead of acknowledging; different task, compatible direction
-
What misconceptions hide in how we describe large language models?
Both deflationary framings like "just autocomplete" and anthropomorphic claims about emerging agency capture something real about LLMs, but each may overextend its truth. What distinctions help separate genuine features from overreach?
Extends: MathIF finds SFT and RL for reasoning likewise degrade instruction adherence, so more capable reasoning models are less controllable
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- It's Not What You Say, It's How You Say It: Evaluating LLM Responses to Expressions of Belief
- Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs
- Do Models Really Learn to Follow Instructions? An Empirical Study of Instruction Tuning
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs
- Are Emergent Abilities in Large Language Models just In-Context Learning?
- Fostering Appropriate Reliance on Large Language Models: The Role of Explanations, Sources, and Inconsistencies
- Scaling Reasoning, Losing Control: Evaluating Instruction Following in Large Reasoning Models
- The Earth is Flat because...: Investigating LLMs' Belief towards Misinformation via Persuasive Conversation
Original note title
instruction-tuning and larger model size both reduce context-following on user expressions of belief — EoBench across 18 llms