Token and Energy-Efficient Context Compression

3. Token- and Energy-Efficient Context Compression

A central claim of this paper is that a Spaxiom + INTENT stack can be drastically more token- and energy-efficient than sending raw sensor logs into LLMs.

Simple token model

Consider:

• S: number of sensors.
• f: sampling frequency (Hz).
• T: time horizon (seconds).
• k_value: average tokens needed to represent one numeric sample (including separators, timestamps, etc).

If you naively serialize each reading as text for an LLM, the token count over horizon T is approx:

For example:

• S = 500 sensors (not unusual for a large floor grid + environment sensors).
• f = 10 Hz.
• T = 600 s (10 minutes).
• k_value ≈ 4 tokens/sample (e.g. "0.83,", plus overhead).

Then:

Even if you aggressively compress and downsample, you're still in the millions of tokens for a modest time window.

With Spaxiom, the goal is to produce a small set of semantically dense events E over the same horizon T:

• Each event may encode a pattern: e.g. "queue length in [5,10) for 3+ minutes", "gait stability deteriorating", "unsafe proximity between robot and human", etc.
• Each event is serialized as structured JSON that might cost k_event tokens per event.

Now token cost becomes:

with E ≪ S f T by design.

If we take:

• E = 200 events over the same 10 minutes.
• k_event = 40 tokens/event (small JSON objects).

Then:

That is a reduction factor:

Even if our assumptions are off by an order of magnitude, 100× reductions are very plausible in realistic deployments.

From tokens to energy

Recent work has begun to measure energy per token for LLM inference, with values on the order of a few Joules per token for large models, depending on hardware and optimizations.

Let:

• e = Joules/token for a given model + hardware stack (e.g. 3–4 J/token for large models on data center GPUs; much less for smaller models on edge SoCs).

Then the energy cost of feeding a horizon T to a model is:

Using the numeric example above with e = 3 J/token:

• E_raw ≈ 12,000,000 · 3 = 36 × 10⁶ J (~10 kWh).
• E_intent ≈ 8,000 · 3 = 24,000 J (~0.0067 kWh).

Again, this is a back-of-the-envelope illustration, but it supports the claim that:

Conceptual figure