Stanford Study: LLMs Encode 'Current Year' Inconsistently, and Prompting Can't Fully Fix It
stanfordnlp · x · 2026-08-21
A Stanford NLP paper accepted at COLM 2026 probes how language models represent the current time via two tasks: an associative task (inferring the year from verb tense) and a declarative task (directly querying the year).
Key findings:
- Both tasks estimate the current year within one year of instruction-tuned models' post-training cutoff; for base models, the associative task serves as a strong proxy for the pre-training cutoff, with an average error of only 10 months across 13 models.
- Their internal mechanisms diverge: the associative task resembles factual recall, while the declarative task lacks consistent causal pathways — making the encoded year hard to update.
- Prompting, SFT, and weight editing all fail to shift both years simultaneously: prompting updates the declarative year (94.6% success across 351 target years) but leaves the associative year nearly unchanged (1.7%); year-shifted SFT matches the target in only 1 of 8 models; weight editing works per-task but doesn't generalize across both.
The paper concludes the current year is not consistently encoded in language models.
More from Research
- 4DAnyone turns single monocular video into 4D human models — janusch_patas · 2026-08-21
- Academic study: agents without MCP match reliability and cost 5-28x less on mature CLI tasks — tobowers · 2026-08-21
- SFT Shifts Reasoning Language, RL Fixes Formatting in Low-Resource SFT — KIEFERSA · 2026-08-21
- GOAG: Object-Agnostic Generative Grasp Planner for Robots — Julien Merand · 2026-08-21
- CoToGrasp: Contact-Topology-Conditioned Zero-Shot Dexterous Grasping — Julien Merand · 2026-08-21
- Tool Calls Can Fake Success: Paper Quantifies Agent Failure Modes — silentw111 · 2026-08-21