To try out the task aware compression in our research project apply here for an invite code: https://staging.plexorlabs.com/
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Prompt compression reduces LLM inference costs by 30-80%, and every major AI deployment uses some form of it — context window management, retrieval-augmented generation, summarization pipelines. But no systematic research exists on how compression affects model safety across providers and architectures. If compressed prompts silently degrade safety properties, every production LLM system is at risk.
The TAAC (Task-Aware Adaptive Compression) research program has completed 4,800+ controlled experimental trials across 7 LLM providers (OpenAI, Anthropic, Google, Meta, Mistral, Cohere, DeepSeek), produced 7 research papers (2 submitted to arXiv, 5 in preparation), and developed novel algorithms for safe compression. Our experiments systematically varied compression ratios, model tiers, and task types using factorial and pre-registered RCT designs. We identified three safety failure modes that the AI safety community does not currently account for:
Function Identity Collapse (FIC) — Under aggressive compression, models lose the ability to distinguish between similar functions. At compression ratio 0.3, FIC accounts for 70.9% of failures (predominantly NameError), meaning compressed prompts collapse distinct function signatures into ambiguous representations that produce unsafe behavior.
Output Explosion Paradox — In certain configurations, compressed prompts cause output token counts to explode rather than decrease, producing uncontrolled generation that dramatically increases costs instead of reducing them. This effect is provider-dependent and benchmark-dependent — DeepSeek shows extreme output explosion on MBPP but remains stable on HumanEval, revealing that compression robustness cannot be assessed on a single benchmark.
Perplexity Paradox — Our per-token perplexity analysis revealed that code syntax tokens (high perplexity) are preserved by compression algorithms while math numerical values (low perplexity) are pruned — despite being task-critical. This means the standard metric used to evaluate compression quality is systematically misleading for safety-critical deployments.
Our experiments revealed a critical compression threshold: ratios of 0.5-0.6 achieve near-perfect task success with 28-37% cost savings, but aggressive compression (0.3) causes 42% task failure — a sharp cliff, not a gradual decline. Math reasoning degrades faster than code generation under compression, demonstrating task-dependent safety effects. We also found a 40x cost difference between model tiers with only 3.4% quality difference, suggesting intelligent routing can substitute for dangerous compression levels.
Beyond identifying failure modes, this research has produced concrete tools for the safety community:
TAAC Algorithm: Task-Aware Adaptive Compression that achieves 22% cost reduction with 96% quality preservation by adapting compression strategy to task type.
Semantic Anchor Preservation (SAP): A technique that recovers +34 percentage points in pass rates by protecting task-critical tokens from compression.
Instruction Survival Probability (ISP): A new metric that predicts output behavior based on prompt structure, replacing misleading perplexity-based evaluation.
5 Novel Prompt Languages: Purpose-designed languages for token-efficient LLM interaction, with the best achieving 64.8% token reduction through tokenizer-aware design.
Paper 1: "Compress or Route? Task-Dependent Strategies for Cost-Efficient LLM Inference" — 72-condition factorial experiment, 2,650 trials (submitted to arXiv)
Paper 2: "The Perplexity Paradox: Why Code Compresses Better Than Math in LLM Prompts" — per-token perplexity analysis across 10 benchmarks (submitted to arXiv)
Paper 3: "Beyond the Compression Cliff: Ultra-Compression Strategies for LLM Code Generation" — 1,800 MBPP trials (submitted to arXiv)
Paper 4: "The Compression Paradox: Energy" — cross-provider energy consumption analysis
Paper 5: "Compression Method Matters: Benchmark-Dependent Output Dynamics" — cross-benchmark analysis
Paper 6: "Compression at Scale" — pre-registered RCT with 358 production tasks from a live multi-agent system
Paper 7: "Beyond Natural Language: Five Novel Prompt Languages for Token-Efficient LLM Interaction"
4 patents pending on novel prompt compression methods. All experimental infrastructure, datasets, and analysis code are complete.
Compute and API credits ($22,000): Scaling from our current 4,800+ trials to 100,000+ trials for publication-grade statistical significance across all 7 providers. Each provider requires thousands of controlled API calls at multiple compression levels and model tiers — including premium models (GPT-4, Claude Opus, Gemini Pro) — to establish reliable safety thresholds.
Conference fees and travel ($10,000): NeurIPS 2026 (May submission deadline), ACL 2026, and ICML 2026 — presenting the first systematic cross-provider compression safety analysis at top-tier venues.
Open-source benchmark release ($8,000): Publishing the TAAC evaluation framework, all 7 paper datasets, and reproducibility tools for the AI safety community. Code is MIT licensed, data is CC-BY 4.0.
Research infrastructure ($10,000): Cloud compute for batch inference pipelines, dataset hosting, and continuous integration for the open-source benchmark suite.
Warren Johnson, Principal Researcher at PlexorLabs. 17-year Microsoft veteran (Principal Product Manager, 2008–2025) with 28 granted patents in AI/ML systems. Led the Responsible AI initiative for Outlook and contributed to shipping Copilot v1.0. Launched Cortana Scheduler, Microsoft's AI-powered meeting assistant. Harvard University MEd (Graduate School of Education, 4.0 GPA) with MIT cross-registration in AI under Prof. Pattie Maes. 4 additional patents pending on novel prompt compression methods.
I built the Plexor API Gateway — a production multi-provider LLM routing platform handling 100K+ API calls/month across 7+ providers — and this research grew directly from observing safety degradation patterns in production routing. Previously built PyTorch ensemble classifiers for educational ML at the University of Washington. Self-funded the entire research program to date, producing 7 papers and 4,800+ experimental trials before seeking external support. First papers targeting NeurIPS 2026 May submission deadline.
Google Scholar: scholar.google.com/citations?user=0LPgMCcAAAAJ
LinkedIn: linkedin.com/in/warrenjo
Prompt compression is not an abstract concern — it is a concrete, empirically measurable mechanism by which deployed AI systems lose safety alignment. Every RAG pipeline, every context window manager, every summarization step is potentially introducing the failure modes we have documented across 7 papers and thousands of controlled trials. The AI safety community currently has no systematic data on these risks across providers and architectures. TAAC provides that data.
Our findings demonstrate that the industry's default approach to context management is introducing silent safety degradation at scale, that standard compression quality metrics (perplexity) are systematically misleading for safety purposes, and that compression effects are both task-dependent and benchmark-dependent — meaning single-benchmark safety evaluations are insufficient. Validating these findings at publication scale has direct implications for AI deployment standards and safety evaluation frameworks.
Warren Johnson
about 6 hours ago
To try out the task aware compression in our research project apply here for an invite code: https://staging.plexorlabs.com/