Researchers map how language models choose between equally valid numbers
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have developed a mechanistic framework for understanding how language models form numerical preferences when all outputs are technically valid. Using Qwen3-4B, they traced preference formation across layers via perturbation probing and gradient attribution, identifying specific neurons and attention heads that drive choice. By surgically editing activations mid-generation, they demonstrated that individual attention heads can shift final outputs while maintaining similar immediate contributions, revealing a distributed decision-making process. This work advances interpretability by showing how models resolve ambiguity at the computational level, with implications for understanding and controlling LLM behavior in open-ended tasks.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper's core contribution is not just identifying which neurons drive numerical choices, but demonstrating that attention heads can shift outputs while maintaining similar immediate layer contributions, suggesting preference decisions are distributed across the network rather than localized. This challenges simpler models of how LLMs resolve ambiguity.
This work sits directly alongside the KV-cache patching study from late October, which showed that fine-tuned behaviors originate in earlier token processing rather than solely at output generation. Both papers use surgical intervention (cache swapping vs. activation editing) to isolate where decisions actually form. The current study goes further by mapping the distributed circuit structure itself. It also complements the commitment direction paper from two days ago, which found that models lock themselves into interpretations mid-generation; this new work reveals the computational mechanics of how that locking happens when multiple valid paths exist.
If the authors can replicate the same attention head circuits across different model sizes or architectures (not just Qwen3-4B), that confirms the preference mechanism is a general feature of how LLMs handle ambiguity. If they cannot, the finding may be model-specific and less actionable for interpretability broadly.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
KV cache swapping reveals where fine-tuned behaviors originate in language models
Researchers have developed a technique to isolate where fine-tuned behaviors originate within language models by swapping KV cache components between base and adapted versions. The work reveals that some learned behaviors, including language preferences and misalignment patterns, are encoded in earlier token processing rather than solely in output generation. This finding has direct implications for…
MentionsQwen3-4B · arXiv
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “When Every Choice Is Correct: A Causal Study of Numerical Preference in Language Models”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.