Researchers Identify Neural Pathways Behind Anchoring Bias in Language Models
A new study identifies the internal circuits within large language models (LLMs) responsible for anchoring bias, where irrelevant numbers in a prompt distort a model's numerical judgments. Using attribution-based circuit localization on 7B–8B parameter Qwen and Llama models, researchers found that edge-level analysis methods more accurately trace the bias signal than node-level methods. The findings offer a mechanistic explanation of a known LLM vulnerability and reveal that post-training fine-tuning meaningfully alters which internal pathways drive the effect.
Researchers at arXiv have published a study examining how anchoring bias — the tendency for irrelevant numerical information in a prompt to skew model outputs — is encoded within the internal architecture of language models. The team designed a controlled multiple-choice experimental setup with shared answer options and developed a logit-difference metric to track when a model's output shifts toward an anchor value rather than the correct answer. Applying attribution-based circuit localization to 7B–8B parameter versions of Qwen and Llama models, both in base and instruction-tuned forms, they found that edge-level circuit methods recovered the anchoring signal more faithfully than node-level approaches. Circuits identified for low- and high-anchor conditions transferred strongly within the same model, suggesting a shared underlying pathway structure regardless of anchor direction. However, transfer between base and instruction-tuned variants of the same model was sparse and less reliable, indicating that post-training processes such as RLHF or supervised fine-tuning reorganize which pathways carry decision-relevant signals. The study contributes to the growing field of mechanistic interpretability by providing a concrete account of how a specific cognitive bias manifests in transformer architectures.
What's missing
The study does not report whether the identified circuits can be surgically suppressed or edited to reduce anchoring behavior, leaving open the question of practical mitigation. It is also unclear whether findings generalize beyond the two model families tested (Qwen and Llama) or to models larger than 8B parameters. The authors note that post-training changes pathway structure but do not quantify how much anchoring bias itself changes behaviorally after instruction tuning.
What different sources said
- arXiv cs.AICenter
Localizing Anchoring Pathways in Language Models
Related
Gut Bacteria Enzyme Found to Break Down Heat-Processed Food Compounds, Producing Novel Biogenic Amines
Researchers have discovered that an enzyme in common gut bacteria can degrade N-epsilon-carboxymethyllysine (CML), a compound formed during thermal food processing, producing previously unknown biogenic amines. The enzyme, ornithine decarboxylase SpeC from enterobacteria, acts on CML and related modified lysine derivatives through a low-level 'underground' catalytic activity. This finding suggests a previously unrecognized communication axis between thermally processed dietary compounds and gut microbial physiology, with potential implications for host health.
Full-Length Gene Sequencing Reveals Two Distinct Bacterial Communities in Black-Legged Ticks Expanding Into Canada
Researchers used Oxford Nanopore full-length 16S rRNA gene sequencing to characterize the microbiome of Ixodes scapularis black-legged ticks collected in Nova Scotia, Canada, distinguishing between tick-adapted bacteria and environmentally acquired bacteria. The study comes as I. scapularis — the primary vector of Lyme disease — is rapidly expanding northward into Canada due to climate change. The findings suggest that environmentally derived bacteria in tick microbiomes are not mere contamination, which has implications for how tick microbiome data is collected and interpreted across surveillance studies.
Study Identifies Metabolic Link Between Cell Envelope Stress and Biofilm Formation in Bacteria
Researchers have discovered that the metabolite acetyl-CoA directly inhibits enzymes that degrade the bacterial signaling molecule c-di-GMP, connecting cell envelope biosynthesis stress to biofilm formation in Pseudomonas aeruginosa. The study found that sub-inhibitory concentrations of antibiotics targeting early peptidoglycan biosynthesis — but not other antibiotic classes — elevate c-di-GMP levels by reducing phosphodiesterase activity, with acetyl-CoA competing for the enzyme active site. Because the relevant enzyme domain is broadly conserved across bacterial species, this checkpoint mechanism may be widespread and could have implications for understanding antibiotic-induced biofilm responses.