body { background: #06080c; color: #e5e9f0; margin: 0; } .rw-nojs-bar { max-width: 880px; margin: 0 auto; padding: 18px 16px 14px; } .rw-nojs-bar .rw-nojs-brand { font: 700 20px/1 Inter, system-ui, sans-serif; color: #e5e9f0; text-decoration: none; } .rw-nojs-nav { max-width: 880px; margin: 0 auto; padding: 0 16px 14px; border-bottom: 1px solid #1c2230; font: 500 14px/1.4 Inter, system-ui, sans-serif; } .rw-nojs-nav a { color: #6f9bff; margin: 0 14px 6px 0; text-decoration: none; display: inline-block; } .rw-nojs-nav a:hover { text-decoration: underline; } .rw-nojs-note { max-width: 880px; margin: 12px auto 0; padding: 0 16px; font: 400 13px/1.5 Inter, system-ui, sans-serif; color: #8a93a6; } #root [data-rw-crawler] { max-width: 880px; margin: 0 auto; padding: 8px 16px 48px; font: 400 16px/1.65 Inter, system-ui, sans-serif; color: #e5e9f0; } #root [data-rw-crawler] a { color: #6f9bff; } #root [data-rw-crawler] h1 { font-size: 28px; line-height: 1.2; } #root [data-rw-crawler] h2 { font-size: 20px; margin-top: 28px; } #root [data-rw-crawler] img { max-width: 100%; height: auto; } #root [data-rw-crawler] ul { padding-left: 0; list-style: none; } #root [data-rw-crawler] li { margin: 0 0 18px; } #root [data-rw-crawler] .rw-pagination { margin: 28px 0 0; display: flex; flex-wrap: wrap; gap: 12px; align-items: baseline; } #root [data-rw-crawler] .rw-pagination strong { color: #e5e9f0; } .rw-nojs-footer { max-width: 880px; margin: 40px auto 0; padding: 22px 16px 44px; border-top: 1px solid #1c2230; font: 400 13px/1.6 Inter, system-ui, sans-serif; color: #8a93a6; } .rw-nojs-footer .rw-nojs-fcols { display: flex; flex-wrap: wrap; gap: 28px 40px; margin-bottom: 20px; } .rw-nojs-footer h2 { font-size: 11px; letter-spacing: 0.05em; text-transform: uppercase; color: #b7c0d3; margin: 0 0 8px; } .rw-nojs-footer a { color: #6f9bff; text-decoration: none; display: block; margin: 0 0 5px; } .rw-nojs-footer a:hover { text-decoration: underline; } .rw-nojs-footer .rw-nojs-legal { font: 400 12px/1.6 Inter, system-ui, sans-serif; color: #6b7384; margin: 0; } .rw-nojs-footer .rw-nojs-legal a { display: inline; } RuntimeWire Investigations Stories AI Startups Venture Products Funding Exits Models Head-to-Head About You're browsing RuntimeWire with JavaScript disabled. Articles and navigation work fully. Interactive features — search, comments, and newsletter signup — require JavaScript.

Hugging Face fixes TRL CISPO default that shrank gradients fivefold in its reproduction

Hugging Face's TRL v1.18.0 sets the CISPO default to 5.0 after an inherited 0.2 ceiling clipped on-policy weights near 1.

By RuntimeWire Staff · Published Nov 14th, 2026, 6:15am CT

A smiling yellow plush holds a glowing blue network-patterned sphere beside a broken metal chain; “Rw” appears on a shelf.

Hugging Face's TRL v1.18.0 sets the CISPO default to 5.0 after an inherited 0.2 ceiling clipped on-policy weights near 1.

Sections

Investigations Stories AI Startups Venture Products Funding Exits

Publication

About FAQ Contact Editorial Policy Corrections Policy Ethics

Tools

AI Model Pricing Head-to-Head

Legal

Privacy Terms

© 2026 RuntimeWire, Inc. All rights reserved.
An independent startup and technology publication based in Austin, Texas and San Francisco, California. Send tips to tips@runtimewire.com.