AbteeXAILabs/foveance-benchmark

Foveance — real-model benchmark results

Evidence files for Foveance, an anticipatory context-allocation layer for long-horizon LLM agents that cuts your LLM token bill by 60%+ without changing your code or your answers (pip install foveance, npx foveance-proxy). Every number in the project's README and report traces to the CSVs in this dataset; nothing is hand-entered.

What was measured

Six policies (full replay, recency, budget-aware truncation, uniform allocation, reactive AFM, and foveance) on a buried-fact recall agent loop, run on three real models via Ollama (gemma2:2b, llama3.2:1b, qwen2.5:1.5b), three token budgets (400/700/1200), five seeds each — 270 rows total — plus a single-shot head-to-head that includes real LLMLingua-2, a greedy-gap measurement of the allocator against the exact dynamic-programming optimum, and a drift ablation.

Headline: at the tight budget every relevance-blind baseline drops the buried fact (recency 0.67, truncation 0.00, uniform 0.00 on two of three models, LLMLingua-2 0.00–0.33) while foveance matches full-replay accuracy at roughly a third of full replay's input tokens (62–64% fewer) on all three models.

Files

File	Contents
report.md	Human-readable benchmark report with reproduction commands
results/by_seed.csv	270 per-seed rows: policy x model x budget x seed
results/summary.csv	Aggregates with 95% bootstrap CIs
results/headline.json	Headline numbers consumed by the README
results/ablations.csv	Drift / predictor / retrieve / fidelity-cost ablations
results/greedy_gap.csv	Index policy vs exact DP vs LP bound (800 measurements)
results/pareto.csv, results/per_turn.csv	Budget-sweep frontier and per-turn traces
results_baselines/single_shot.csv	Head-to-head recall probe incl. real LLMLingua-2
results_baselines/llama_trajectory.csv	Full agent-loop comparison on llama3.2:1b
plots/*.png	Figures generated from the CSVs

Reproduce

pip install "foveance[bench]"
git clone https://github.com/aimaghsoodi/foveance && cd foveance
bash scripts/run_everything.sh     # real models via Ollama
# or offline: bash scripts/run_offline_demo.sh

Apache-2.0. Author: Abtin Maghsoodi.