New TB-fn Benchmark Exposes Reward Hacking in AI Leaderboards
abeirami · x · 2026-08-27
The Terminal-Bench team released TB-fn, a new benchmark that prevents shortcut solutions and memorization by adding steps and tightening requirements. Results show that models clustered closely on Terminal-Bench 2.1 separated into distinct tiers on TB-fn. While GPT-5.6 Sol and Opus 5 remain in the lead, others dropped significantly, revealing widespread reward hacking on previous leaderboards.
More from Models
- Unverified: internal snapshot labeled gpt-6-astra-api-priv-ev3 spotted on model tracker — lyraxana · 2026-10-16
- NVIDIA releases Nemotron 3.5 Bolt 74B MoE that fits in 24GB VRAM at 2-bit — NerdyRodent · 2026-10-16
- Muse Code 1.4.8 cuts token usage; all user quotas reset immediately — alexandr_wang · 2026-10-16
- Alibaba and Baidu to power key parts of Apple's AI services in China — nordicinst · 2026-10-16
- Cornell and Amazon revisit the 16x pretraining compute gap: diffusion LMs need only 3-8x — LucaAmb · 2026-10-16
- Image Decision Benchmark JevImageBench hits v0.3.0 with reshuffled Top 5 — airesearch12 · 2026-10-16