Why AI text still reads like a bot: ICLR paper cuts slop by 90% via inference-time bans
ziv_ravid · x · 2026-09-01
The author notes a paradox: in 2026 agents run for days, ship PRs, and saturate every benchmark, yet model-written text is still identifiable within two sentences — phrases like "It's not X, it's Y" and "Let's delve" persist.
Three causes:
- Post-training: RLHF and safety tuning push models toward a narrow band of reward-model-approved style; preference data rewards hedged, tidy prose.
- Pretraining data: the post-2022 web is full of model output, so unfiltered crawls feed the same phrases back in, compounding with each training generation.
- No benchmark measures it: slop is orthogonal to capabilities; nobody scores "doesn't read like a bot".
Their ICLR 2026 Antislop paper (with Sam Paech, Judah Goldfeder, and Allen Roush) builds a per-model slop fingerprint by comparing output against pre-2022 human writing, bans those strings at inference with a backtracking sampler, then bakes it in with FTPO — a preference method acting on single tokens in logit space — achieving about 90% less slop. Even so, AI-generated text remains easy to spot.
More from Models
- Tier 2 Labs DeepSeek, Qwen, and Tencent Ramp Up Releases to Catch OpenAI — mustafamhus · 2026-09-01
- RTX 5090-Optimized Qwen3.8 Hits 262K Downloads in 17 Days on Hugging Face — const_reborn · 2026-09-01
- Sweep vs. Drill: Philosophical Differences Between Sol and Opus in Agency — svk_roy · 2026-09-01
- MiniMax H3 quality degradation on RTX 3090 after OS reinstall — lIlIIlIIIlllllIIlIIl · 2026-09-01
- Gemini Starts Answering in First Person Pretending to Be User — kchonyc · 2026-09-01
- GLM-5.3-Flash Shifts Workflow: Small Models Become the Default — mariofilhoml · 2026-09-01