TAS: Temporal Attractor Steering
for Open-Weight Language Models

Right knowledge, wrong answer

Language models are trained on corpora spanning many years, so a model often encodes both an earlier fact and its later replacement — an older CEO and a newer CEO, an earlier head of state and a later one. Under standard prompting the model surfaces the outdated answer, but a simple temporal cue (“As of 2024, who is…”) reveals the newer fact is already recoverable from parametric memory. We call this Parametric Temporal Conflict (PTC): the right knowledge is in the network, but the default forward pass favors the older basin.

How TAS works

TAS is a three-stage inference-time pipeline: detect a likely conflict with a lightweight probe, locate a conflict-critical layer once per model via causal activation patching, and steer the hidden state toward the newer-fact direction with a precomputed Δ. No retraining, no retrieval, no weight edits.

Figure 1. Overview of Parametric Temporal Conflict (PTC) and Temporal Attractor Steering (TAS). (A) Standard prompting retrieves an outdated answer from parametric memory. (B) Conflict localization compares standard and temporally cued prompts and identifies a conflict-critical layer ℓ* through activation patching. (C) Temporal steering shifts the hidden state toward the newer-fact representation without retrieval or retraining.

Verified evaluation across four open-weight LMs

We evaluate on an 8,746-record verified benchmark mined from Wikidata, spanning five superseding-update relations (head of state, head of government, CEO, head coach, chairperson). All four models exhibit PTC on this benchmark, all four have a recoverable conflict-critical layer, and end-to-end TAS recovers a substantial fraction of PTC cases while preserving non-conflict behavior.

Run TAS in minutes

Load an open-weight base LM, point the pipeline at a PTC record, and apply the precomputed steering vector at the model's conflict-critical layer.

# Clone the repository
git clone https://github.com/eliashossain001/temporal-attractor-steering.git
cd temporal-attractor-steering
pip install -e .

# Pull the verified PTC benchmark
huggingface-cli download EliasHossain/ptc-benchmark \
  --repo-type dataset \
  --local-dir data/ptc-benchmark/

from tas import load_model, TASPipeline
from datasets import load_dataset

# Load any of the four evaluated open-weight base LMs
model, tokenizer = load_model("Qwen/Qwen2.5-7B")

# Pipeline auto-loads per-model l*, alpha*, tau*, and Delta
tas = TASPipeline.from_pretrained(
    model, tokenizer,
    variant="V2",   # per-relation Delta
    tau=0.15,
    alpha=2.0,
)

records = load_dataset("EliasHossain/ptc-benchmark", split="test")

for r in records:
    out = tas.evaluate(r)
    # out.steered in {True, False}; out.answer is the post-steering prediction
    print(r["query"], "=>", out.answer,
          "(steered)" if out.steered else "(passthrough)")

The PTC benchmark

8,746 verified records mined from Wikidata, spanning five superseding-update relations across politics, corporate, and sports domains. Each record carries a verified (a_old, a_new, t_update) triple with non-overlapping validity windows and both standard and temporally cued prompts.

# Verified PTC benchmark (8,746 records, 5 Wikidata relations)
huggingface-cli download \
  EliasHossain/ptc-benchmark \
  --repo-type dataset \
  --local-dir data/ptc-benchmark/

Cite TAS

If you find TAS or the PTC benchmark useful in your research, please consider citing the paper.