IBM's ALTK-Evolve beats ACE on AppWorld using ~40% of the tokens, and it's the cleanest evidence yet that retrieval beats hoarding

Numbers first, because they're the whole argument.

On AppWorld's 168 tasks with DeepSeek-V3.2: ALTK-Evolve hit 89.3% goal completion at 263K tokens per task. ACE hit 80.4% at 634K (IBM Research on HF). Nine points better, 41% of the tokens. On the weaker gpt-oss-120b: 56.0% at 116K versus 54.8% at 777K. Roughly one-seventh the inference cost for a small accuracy win.

The architectural difference is the point. ACE builds one playbook from failures and injects the whole thing at every inference step. ALTK-Evolve consolidates lessons into individually retrievable guidelines with support counts, then sends only a curated subset sized to the model's capacity. That's it. Same learning-from-failure premise, different delivery.

Notice what happened on the weak model: the gap in accuracy nearly vanished but the cost gap exploded. Always-on context degrades faster the less capacity the model has to ignore irrelevant parts of it. If you're routing cheap models to high-volume roles (see story 1), this compounds.

Three other findings this cycle point the same direction, which is why I'm giving this a Top 5 slot instead of burying it in Research.

"Not Worth Another Token" tested marginal-value pruning at three points in a deep-research agent pipeline: pre-retrieval, post-retrieval, pre-synthesis (arXiv 2608.08389). Lightweight heuristics alone cut token usage up to 73% with little quality loss, and early-stage pruning produced by far the largest end-to-end savings. The cheapest win is deciding what not to retrieve. Honest caveat the authors flag: no single strategy won on quality, efficiency, and faithfulness at once.

A second paper attacks turn-by-turn context interference in search agents, arguing that accumulated irrelevant documents actively distract rather than merely waste tokens, and adds a distill-based context refiner into the RL training pipeline (arXiv 2608.10743). And READ replaces dense top-K retrieval with deterministic agentic navigation over document structure, answering 58.8% of financial document questions against 15.7% for dense retrieval, p=2×10⁻⁵, with tuned baselines only reaching 35.3% (arXiv 2608.06305).

Four independent results, one shape: your context window is a budget, not a bucket. Stop appending.

What I'd do Monday: instrument what fraction of injected context your agent actually references in its output. If you're running a monolithic playbook or CLAUDE.md that's grown past a couple thousand tokens, split it into retrievable chunks with a cheap relevance gate. IBM's support-count idea is the underrated bit, tracking how often each guideline actually earned its place gives you a pruning signal for free.