BAGEN: LLM Agents Waste 44% of Tokens on Tasks They'll Fail
Frontier LLM agents waste 28–64% of tokens on doomed tasks, BAGEN reveals. Learn how budget interval estimation and early stopping work in practice.
Tag archive
Frontier LLM agents waste 28–64% of tokens on doomed tasks, BAGEN reveals. Learn how budget interval estimation and early stopping work in practice.
May 2026 paper shows LLM agents lose 30+ percentage points on structured backend tasks as constraints accumulate. PoC reproduces the exponential decay curve.
How CMU's AutoExperiment benchmark uses progressive code masking to measure AI agents' ability to replicate ML research from paper descriptions alone.
New research shows RL post-training only modifies 1–3% of token positions, always within the base model's existing top-5 candidates. Here's what it means.
How to diagnose chain, star, and mesh LLM agent topologies before inference using spectral analysis. PoC of arXiv 2605.11453 in pure NumPy.
LMR-BENCH (EMNLP 2025) benchmarks LLM agents on reproducing code from 23 NLP papers. This PoC explains the masking methodology, evaluation axes, and what the results mean for AI-assisted research.
RRO replaces costly process reward model exploration with a rising-reward filter. Effloow Lab reproduced the trajectory selection logic in Python.
How multi-agent debate improves LLM factuality by 8+ points on math benchmarks. Paper breakdown and 90-line Python PoC implementation.
ARES dynamically selects reasoning effort per agent step — high for complex decisions, low for navigation — reducing tokens 52.7% on TAU-Bench Retail.
MemMachine stores raw conversation episodes instead of LLM-extracted summaries, reaching 93% on LongMemEvalS with 80% fewer tokens than Mem0.
The 'Don't Waste Bits' paper (arxiv 2604.04722) shows adaptive per-token KV precision beats static quantization by 17.75% latency and 7.6 accuracy points. Here's how it works.