JOLT+ beats GRPO on all five benchmarks, up 9.9 points on AppWorld
perplexity_ai · x · 2026-10-22
Perplexity reports final-checkpoint evaluations for JOLT+: it beats GRPO on all five benchmarks spanning math, coding, tool use, and terminal tasks.
- GSM8K: 76.1% (+1.5)
- MATH: 72.59% (+4.2)
- LCBv6: 49.08% (+3.0)
- AppWorld state-test pass rate: 72.09% vs GRPO's 62.20% (+9.9)
Results are mean ± 1 SD across 3-5 training runs per configuration, with JOLT and JOLT+ consistently outperforming GRPO, OPSD, and π-Distill baselines.
Related event: JOLT explained: dual teacher-student roles beat GRPO across five benchmarks(2 posts)→
More from coding & agent
- Amp team runs with 21 people, no PRs, one weekly meeting, 100 releases a day — jevon · 2026-10-22
- GPT models reportedly coming to Google's Antigravity coding tool — zacharynado · 2026-10-22
- Slack-threading an AI agent to track customer feedback and bug investigations — craigsdennis · 2026-10-22
- 112 agents wrote 60,000 lines of code in 16 hours — by writing specs, not prompts — colinmcnamara · 2026-10-22
- How Codex Happened: Inside OpenAI's comeback in AI coding, from failed launch to ChatGPT's centerpiece — every · 2026-10-22
- A voice-driven MCP chain: ChatGPT Voice orchestrating multiple CLI coding agents — PurposePossible1649 · 2026-10-22