PACE: Two-Timescale Self-Evolution for Small Language Model Agents
PACE is a framework that lets frozen small language models improve agent behavior by refining prompts first, then cautiously updating control logic.
Intelligence analysis by GPT-5.4 Mini
The paper argues that small language model agents can self-improve without weight updates or frontier-model teachers. PACE separates low-risk prompt changes from higher-risk control-logic changes, using validation to decide when to advance.
A small AI helper can sometimes get better at its job by trying new instructions and changing how it makes decisions.
PACE is like fixing a bike in two steps. First, it adjusts the handlebars and seat, which is safer. Only later does it replace bigger parts, and only if a test ride says the change is good.
The paper says this helps small AI helpers improve without needing a giant teacher model or changing the model’s brain. That could make them cheaper and easier to use in real systems.
Analysis
What the paper proposes
PACE stands for Prompt And Control Logic Evolution. The core idea is to split agent improvement into two stages with different risk levels. First, the system keeps the control logic fixed and lets the agent refine prompts. Only after prompt gains slow down does it consider changing the control logic itself, and even then the change must pass held-out validation.
Why this is different
The paper targets frozen small language models, not frontier models and not weight updates. That matters because many existing self-improvement methods assume a stronger model can reliably diagnose failures, suggest fixes, and judge whether those fixes are good. PACE is presented as a way to get self-evolution under tighter compute and human-labor limits.
Reported results
Across three frozen backbones from 4B to 14B parameters and four controlled benchmarks, the paper says PACE performs best on all 12 backbone-benchmark pairings. It reports up to a 9.2% relative gain over vanilla SLM agents and up to a 5.4% gain over a stronger single-mode evolution baseline. A tau-bench case study also suggests better multi-turn tool use than vanilla agents and prompt-only evolution.
Takeaway
The paper’s main claim is not that one final prompt or control policy is magic. It is that validated, autonomous search over agent strategies can uncover better inference behavior even when the model itself stays frozen.
Key points
- PACE uses a two-stage system: prompt changes first, control-logic changes later.
- The method is designed for frozen small language models, not weight updates.
- The paper reports best results across 12 backbone-benchmark combinations.
- A tau-bench case study suggests better multi-turn tool use.
- The authors frame this as self-evolution without frontier-model teachers.
