Do Language Models Consistently Encode the Current Year?
Suze van Adrichem, Aditi Bhaskar, Diyi Yang, Christopher Potts, Jing Huang
COLM) 2026
cs.CL, cs.LG
2026-08-16
Two probes split the current year: prompting hits the stated year on 94.6% of 351 targets, the verb-tense year on 1.7%. SFT and weight edits cannot move both at once.
Language models need a working notion of "now" for temporal reasoning. Who the president is, whether a policy still holds, and whether a question is even about the present all hang on that anchor. Treating the data cutoff as the present produces over-refusal. Writing a year into the prompt is also a known handle for temporal backdoors.
Most existing work either scores temporal QA or tries to refresh models with fine-tuning and prompting. It treats "the current year" as one concept. This COLM 2026 paper from Stanford splits that assumption. The same model carries at least two years, learned at different stages, and they do not move together.
Two probes.
Associative current year (ACY): prompt In {year} there and read the next-token tense. was/were count as past; is/are/will count as present/future. ACY is the first year where P(past) drops below P(future). Those five verb forms hold more than 90% of next-token mass on 1950-2050 for every model in the 13-model set. A month-level variant is ACYM.
Declarative current year (DCY): ask the model what year it is, using 10 continuation or instruction prompts, then average greedy-decoded years.
For mechanisms they run distributed alignment search (DAS) to find a one-dimensional "year-type" subspace in the residual stream. For updates they try three families: a system prompt that names a target year, SFT on year-shifted Tulu3, and low-rank edits (ROME for DCY, MEMIT for ACY).
OLMo2 is the main subject because its two-stage pre-training mix is public. Behavioral probes cover 13 base models.
ACY is a tight proxy for the pre-training cutoff. Mean error across 13 base models is 0.85 years (about 10 months); ACYM is 0.75 years. The WIKISPAN perplexity estimator from Cheng et al. is off by 2.71 years on the same 13. OLMo2-7B is the outlier: ACY is only one year off, but ACYM lands on January 2029, 61 months past a December 2023 cutoff.
On OLMo2, ACY tracks the mix. Stage 1 sits at 2021-2022; stage 2 moves to 2023-2024. N-gram and co-occurrence count models shift in the same direction.
DCY is a post-training object. OLMo2-1B at the end of pre-training reports 2011±15.0 across the ten prompts. After SFT it is 2023±0.0, and confidence jumps from about 6% to 60%. Dropping every Tulu3 example that contains a number ≥1000 (76% of data remains) still collapses variance to zero, but DCY stops at 2021 rather than the control's 2023. Instruction following makes the answer consistent; the year distribution picks which year.
The internal routes do not match. ACY looks like factual recall: on OLMo2-7B a 1-D subspace on the last year token, layers 4-22, is enough to swap tense predictions. DCY has no stable pathway. Knockouts often hit the first token of the prompt, not the word "year", and the map changes when the question is rephrased.
Updates make the split operational.
| Method | DCY | ACY |
| System prompt (351 target years) | 94.6% exact, 0.45-year mean error | 1.7% exact, predictions stuck in 2019-2030 |
| Year-shifted SFT (8 targets) | almost exact shift | 1 of 8 matches the target |
| ROME (edit DCY) | hits the target | ACY barely moves |
| MEMIT (edit ACY) | stays at 2022/2023 | all 8 targets within ±6 years |
OLMo2-1B-SFT-control, never trained with a dated system prompt, still hits DCY 50.7% of the time, off by 6.82 years on average. Gemma-3 and Qwen-3.5 repeat the split: Qwen3.5-4B/9B/27B hit DCY 100%; the best ACY number is 53.3% on Qwen3.5-27B, and every other model is below 40%.
Putting today's date in the system prompt only moves the year the model is willing to say. The year sitting in grammar was welded in during pre-training. Anyone doing knowledge refresh, temporal backdoors, or cutoff alignment who only asks "what year is it" will over-count success.
A cheap cutoff probe falls out as a side effect: In {year} there is 10 months off on average across 13 bases, tighter than a WIKISPAN perplexity sweep.
This is mechanistic evidence, not a new alignment method. None of the three current update tools moves both years at once.
There is no Limitations section. The edges are still visible.
The associative probe is English verb tense. Other languages and templates other than In {year} there are not tested at the same strength. Causal localization and weight editing are almost entirely on OLMo2-1B/7B; Table 5's cross-family numbers cover prompting only.
The "year" subspace also fires on four-digit numbers that are not years, with 90%+ interchange accuracy in early layers. It is closer to a four-digit-number-to-tense channel than an abstract clock. ACYM blowing out by 61 months on OLMo2-7B means month granularity is not robust.
Target years run to 2250, far from the training mix, so failures mix "cannot edit" with "out of distribution". Weight edits are task-specific; there is no joint objective that constrains both years.