Researchers Identify Neural Pathways Behind Anchoring Bias in Language Models
A new study using circuit localization techniques has identified where anchoring effects—where irrelevant numbers shift model judgments—originate within language models like Qwen and Llama. The researchers found that edge-level methods better capture these anchor-sensitive signals than node-level approaches, and that anchoring pathways show consistent structure within individual models. The findings provide mechanistic insight into how language models process numerical context, which has implications for understanding and potentially mitigating reasoning biases in AI systems.
Researchers studying language model behavior have pinpointed the internal neural circuits responsible for anchoring effects, where irrelevant numbers in prompts influence model outputs on numerical reasoning tasks. Using a controlled multiple-choice experimental setup with shared answer options across 7B–8B parameter Qwen and Llama models (both base and instruction-tuned variants), the team developed a logit-difference metric to track which answer options the models favored relative to anchor-aligned choices. Through attribution-based circuit localization, they discovered that edge-level methods—which track connections between neurons—recover anchoring signals more reliably than node-level methods that examine individual neurons. The study found that anchoring circuits transfer strongly within a single model across different anchor directions, suggesting a consistent underlying pathway structure. However, transfer between base and instruction-tuned versions of the same model was less reliable, indicating that instruction-tuning modifies which neural pathways drive anchoring behavior.
What's missing
The study does not discuss potential practical applications for mitigating anchoring bias in deployed language models, nor does it address whether findings generalize to larger models (13B+ parameters) or other model families beyond Qwen and Llama.
What different sources said
- arXiv cs.AICenter
Localizing Anchoring Pathways in Language Models
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.