Temporal Horizon and how to steer it

A dedicated study of the temporal horizon and how to steer it with minimum noise.

Project Details

This project was a done as a part of SPAR cohort. I have already performed all the required representational and geometrical analysis of temporal horizon across 7 different LLMs. The focus of this project is to understand how temporal horizon is encoded in LLMs and how can we better steer it. A synthetic dataset was generated which classified temporal horizon into short and long (call it H1) and a graded dataset (call it H2) which quantized the temporal horizon spectrum into points for better analysis. Activation Patching, Divergence tests, Probe training, modality tests and representational alignment tests agree and establish that temporal horizon is a property that is build upon through the model depth where the most focus is given to it in the middle layers of the model. A geometrical analysis of the model shows that the temporal horizon get stretched out progressively while also being unstretched and untwisted through the model depth. A quantitative analysis of manifold and linear steering is done which reveals linear steering has lower-to-similar naturaleness compared to manifold steering. A qualitative analysis has to be done to understand where is steering the most effective and help generalize temporal horizon manipulation.

Theory of Impact

A better understanding of geometry of temporal horizon will increase out knowlege of how LLMs encode world knowledge in their weights. having the ability to steering it effectively will help improve their performance.

People

vansh gupta

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$5,000

Ideal

$5,000

How the money will be spent

Three sets of experiments with one ablation.

models used are (Qwen3-1.7B, Qwen3-4B, Qwen3-8B, Qwen3-14B, Pythia-12B, Llama3.1-8B, Gemma2-2B)

Behaviour steering ~ 6h
component ablation ~ 400h (due to scale of the models)
LLM as a judge evaluation of steered outputs ~ 5h (with a 30B or 40B parameter model)
score component ablation (but runs on CPU).

The datasets each have 5000 datapoints. Running a large scale experiments will require GPUs with 80 to 100GB of RAM and good throughput.

It can easily cost around $600-$1000.

Rest of the money will be used to address any rebuttal experiments if needed. I am planning to submit this work to NeurIPS workshop and ICLR so rest of the funds will also cover travel and accomodation costs.

Discussion

No comments yet. Be the first to share your thoughts.