Temporal Awareness

"I want my time to grant me what time can't grant itself." — Al-Mutannabi

Research on detecting and steering temporal awareness in LLMs.

Overview

This project investigates how LLMs encode temporal reasoning and whether we can:

1. Detect temporal preference from internal representations
2. Steer temporal orientation via activation engineering
3. Measure divergence between stated and internal time horizons

Key findings:

• GPT-2 encodes temporal scope with 92.5% linear separability
• Steering validation: r=0.935 correlation between steering and probe predictions
• Late layers (6-11) encode semantic temporal features robust to keyword removal

Program

Research Program

Framework

We ground temporal awareness in intertemporal preference:

U(o_i; θ) = u(r_i) · D(t_i; θ)     # Value function
t_internal = inf{t : D(t) ≤ α}     # Internal horizon

Key questions:

• Does t_internal ≈ t_h (stated horizon)?
• Can we detect divergence between stated and internal preference?

See docs/research_plan.md for full framework.

Setup

pip install -e .
cp .env.example .env  # Add API keys

Structure

temporal-awareness/
├── data/
│   ├── raw/                 # Intertemporal preference datasets
│   ├── validated/           # Human-validated
│   └── processed/           # Train/val/test splits
├── scripts/
│   ├── probes/              # Probe training & validation
│   └── analysis/            # Figures, metrics
├── results/checkpoints/     # Trained probes & steering vectors
├── docs/
│   ├── research_plan.md     # Full framework & roadmap
│   └── RELATED_WORK.md      # Literature review
└── paper/                   # Manuscript

Quick Start

from latents import SteeringFramework
from latents.model_adapter import get_model_config

# Use latents library for extraction and steering

# Train probes
python scripts/probes/train_temporal_probes_caa.py

Related Work

See docs/RELATED_WORK.md:

• Zhu et al. 2025: Steering Risk Preferences via Behavioral-Neural Alignment
• Mazyaki et al. 2025: Temporal Preferences in LLMs for Long-Horizon Assistance
• Time-R1: Comprehensive temporal reasoning (arXiv:2505.13508)

Public Datasets

Dataset	Source	Link
Time-Bench	Time-R1	HuggingFace
Test of Time	Google	HuggingFace

License

MIT