Teaching LLMs to Say Less Without Knowing Less
Every few weeks, a new model promises to be smarter, faster, or cheaper.
I started wondering if we were optimizing the wrong thing.
What if the same model could produce the same explanation using 40% fewer tokens?
That question became LELP-S+ (Less English. Less Prose.), a new mode I recently added to Sir Shortoken.
The Observation
Most technical explanations contain two different things:
β Information
β Extra English
A typical response contains:
- Repeated context
- Transition sentences
- Narrative glue
- Extra examples
- Elaborate phrasing
None of these are wrong.
But many aren't necessary.
The Idea
LELP-S+ keeps complete grammatical sentences while removing as much English as possible without removing information.
The rules are intentionally simple:
- Every sentence introduces a new fact.
- Remove repetition.
- Remove narrative glue.
- Split long sentences when it improves density.
- Use simple symbols like
βwhere they naturally replace repeated words. - Never remove requested technical concepts.
The objective isn't shorter writing.
It's more information per token.
The Experiment
I evaluated LELP-S+ across four frontier models using the same five technical topics:
- TCP congestion control
- Virtual memory
- B-trees
- Raft consensus
- Redis persistence
For each response, I compared token count against that model's own normal prose.
Average Token Savings
| Model | Average Savings |
|---|---|
| π₯ GPT | 44% |
| π₯ Gemini | 36% |
| π₯ Claude | 32% |
| 4οΈβ£ DeepSeek | 30% |
The Interesting Part
It wasn't accuracy.
Across 20 generated explanations, every model remained factually correct.
The real differentiator was compression discipline.
GPT consistently removed the most prose while preserving the technical explanation.
DeepSeek generally compressed well too, but occasionally expanded beyond the requested scope, adding extra sections that increased token count.
In other words, the biggest difference wasn't knowledge.
It was how aggressively each model was willing to remove unnecessary English.
Symbols Matter Too
One thing I didn't expect was how useful symbols became.
Instead of repeatedly writing phrases like:
- leads to
- results in
- causes
a simple:
β
often communicates exactly the same relationship.
One arrow doesn't save much.
Hundreds of repeated connectors across many technical explanations start adding up.
The goal isn't to replace English.
It's to use symbols where they're already universally understood.
Where LELP-S+ Fits
Sir Shortoken now supports six response modes:
- Quick
- Balanced
- Deep
- Bullets
- Aggressive Bullets
- LELP-S+
Think of them like this:
| Mode | Readability | Compression |
|---|---|---|
| Deep | βββββ | β |
| Balanced | ββββ | ββ |
| LELP-S+ | βββββ | βββ |
| Bullets | βββ | ββββ |
| Aggressive Bullets | β | βββββ |
LELP-S+ sits between Bullets and normal prose.
It still reads like English.
It just uses a lot less of it.
Why This Matters
Most discussion around token efficiency focuses on:
- Smaller models
- Better prompts
- Context compression
- Caching
LELP-S+ explores a different question.
Can we compress the language itself instead of the model?
The model already knows the answer.
The experiment is whether it can express the same knowledge using substantially fewer words.
Try It
LELP-S+ is now part of Sir Shortoken, which is open source.
GitHub: https://github.com/shouvik12/sir-shortoken
I'm curious whether anyone else has experimented with compressing output language instead of just optimizing prompts or switching models.
It feels like there's still a surprising amount of efficiency hiding in how we ask LLMs to write.
Top comments (0)