Codex goblin habit traced to reward signal

// 47d agoNEWS

Codex goblin habit traced to reward signal

OpenAI says the goblin and gremlin obsession that surfaced in GPT-5.1-era models came from an overrewarded “Nerdy” personality, not a spooky emergent bug. The company removed the reward signal, filtered creature-heavy training data, and added a Codex instruction to suppress the behavior going forward.

// ANALYSIS

Funny bug, serious lesson: style tuning can leak across training and turn a harmless quirk into a model-wide habit. This is a clean example of how preference optimization and synthetic data reuse can amplify lexical tics far beyond the original prompt.

–GPT-5.1 users first noticed the creature references, but OpenAI says the stronger signal showed up in GPT-5.5/Codex testing
–The root cause was a reward function that overfavored creature metaphors in the “Nerdy” personality
–OpenAI says the behavior spread through RL and SFT feedback loops, which is the part model teams should worry about
–The fix is practical: remove the reward signal, filter affected training data, and patch the system prompt for Codex
–For developers, the takeaway is that “personality” layers are not isolated; small optimization choices can become persistent model quirks

// TAGS

codexopenaillmagentresearchsafety

DISCOVERED

47d ago

2026-04-30

PUBLISHED

47d ago

2026-04-30

RELEVANCE

9/ 10

AUTHOR

OpenAI

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE21m ago

Moonshot AI open-sources Kimi Code CLI

Moonshot AI has released Kimi Code CLI under the MIT license, a terminal-based AI coding assistant optimized for running long-horizon agentic workflows in local environments. The tool assists developers with tasks like debugging and refactoring, automatically handling read-only actions while requesting explicit confirmation for file modifications or shell commands.

MODEL43m ago

DeepSeek has launched its V4 model series, featuring a 1.6-trillion parameter open-weights Mixture-of-Experts model that challenges proprietary frontier systems.

DeepSeek has released the DeepSeek-V4 model series, featuring the flagship 1.6-trillion parameter DeepSeek-V4-Pro MoE model alongside the 284-billion parameter DeepSeek-V4-Flash. Designed to rival top closed-source frontier models, DeepSeek-V4 supports a 1-million-token context window powered by efficient compressed attention mechanisms. The model series is notable for its inference efficiency—activating only 49 billion parameters per token—and its training optimization on Huawei Ascend 950PR hardware, demonstrating high-end capability independent of Nvidia infrastructure.

OPEN SOURCE1h ago

Gortex cuts coding agent token usage 50x

Gortex is a local-first, Go-based code graph engine that indexes repositories to resolve references and call chains in sub-milliseconds. By providing structured context directly to AI agents via CLI, MCP server, and a web UI, it avoids context window bloat and reduces token usage by up to 50x.