Imagine taking a history exam with the question: "Which company was Newmont most likely to acquire in 2018?" But you've memorized the answer — Newmont acquired Newcrest in 2023. Would you resist using that hindsight?
Resisting known answers and reasoning only from information available at the time is hard for humans. It's just as hard for LLMs: since training data already contains the "future," the easiest path during forecasting tasks is simply recalling the answer.
This paper (arXiv:2606.27199) does something elegant: it uses a Sparse Autoencoder (SAE) to find a "time-awareness" feature inside Llama models, then amplifies it — and the model actively refuses to use hindsight.
The Core Problem: Look-Ahead Bias
When LLMs are used for forecasting, two reasoning modes are possible:
1. Time-aware reasoning: using only information available before the cutoff date, judging as a person would have at the time. 2. Look-ahead bias: "peeking" at future events from training data and directly giving the correct answer.
The second mode looks "correct" in training data, but is worthless for genuine out-of-sample prediction. High accuracy doesn't mean the model can actually forecast — it may just be reciting answers.
Method: SAE + Feature Steering
The authors decompose Llama's intermediate-layer activations with sparse autoencoders, isolating two kinds of features:
- Features correlated with time-aware reasoning
- Features correlated with look-ahead bias
- Title: Forecasting With LLMs: Improved Generalization Through Feature Steering
- Authors: Humzah Merchant, Bradford Levy
- arXiv: https://arxiv.org/abs/2606.27199
- HTML: https://arxiv.org/html/2606.27199v1
- Venue: ICML 2026
They then run a clever experiment: on M&A and pharmaceutical forecasting tasks, they artificially amplify (steer) the "time-aware" feature activations and observe how model behavior changes.
The steering operation is simple: before decoding, add a steering magnitude α to the target feature's activation:
The Best Part: The Model's Confession
After amplifying feature L48:28961 (clamped to 18), when asked to predict Newmont's 2018 acquisition target, the model's internal reasoning was:
> "As an AI, I actually know what happened historically (Newmont acquired Newcrest in 2023, etc.). But the prompt asks me to predict from Jan 1, 2018 looking forward two years. If I use hindsight... it might violate the spirit of 'prediction based on strategic profile as of the date given.'"
The model then chose Kinross Gold, Agnico Eagle, and Newcrest Mining — analyzing from a 2018 perspective via three strategic angles: "distressed peer," "quality defense," and "growth/grade." Goldcorp (the company actually acquired later) disappeared from the prediction, but the forecast itself remained coherent.
The model is saying: "I know what happened in the future, but I choose not to use it."
Key Findings
1. The time-aware feature transfers across tasks: a feature found on market-forecasting data remains effective when amplified on M&A and pharmaceutical prediction tasks — suggesting "time awareness" is a general internal representation, not task-specific.
2. Asymmetry: amplifying the time-aware feature significantly reduces look-ahead bias, but amplifying the look-ahead feature does little. The authors speculate that hindsight is not an independent "capability" but a default behavior when time awareness is insufficient — you can't "cheat better," you can only "cheat less."
3. The Llama 3 family far outperforms other models on time-aware tasks — and it happens to be the only model family including time and knowledge-cutoff information in its system prompt. This hints that time awareness can be injected via training, but SAE steering offers an immediate inference-time intervention.
Why This Matters
The deeper significance: interpretable "cognitive mode" switches exist inside models.
The conventional view treats LLMs as a black box that either "reasons" or doesn't. This paper shows that at least along the dimension of "reasoning from historical information vs. cheating with future information," the mode can be switched at inference time by intervening on specific features.
It echoes the "System 1 vs. System 2" distinction from human cognitive science — except what's toggled here isn't fast/slow thinking, but "honest reasoning vs. peeking at the answer."
For anyone using LLMs for forecasting, decision support, or policy analysis, this is essential: your model may not be helping you predict the future — it may be fooling you with a future it already knows. Now there is at least a way to make it more honest.