Home › Code › compute_approx_kl

compute_approx_kl

Syntologyentry name in harvested coderead from the graph 2026-09-24

compute_approx_kl appears in the code Syntology harvested for 20 papers, as 12 distinct code bodies found in 22 places (a place is one code body under one paper). At least one of them ran in 8 of the papers; 2 of the code bodies carry a behaviour fingerprint.

What this page is not. Routines are grouped here by the exact string of their function or class name. Nothing asserts that two samples named compute_approx_kl do the same thing, share code, or are comparable; the name is a string, not an identity. Behaviour outputs (what a fingerprinted sample returned on the shared battery) are not in this export and are not shown here; the graph at syntology.ai holds them. "Ran" means executed on a synthesized fixture, not that the code is correct or reproduces a paper.

Samples Syntology

Syntology ran 5 of the 12 distinct code bodies named compute_approx_kl; 7 are unverified. One tile per status, in the site's fixed vocabulary, each code body counted once:

0ran · honoured contract
0ran · violated contract
1ran · our draft was wrong
0ran · fixture could not drive it
4ran
7unverified
2fingerprinted

Licence is a property of each copy, so it is counted per place: 2 of the 22 places are pointer only (Syntology does not serve that copy's text). This site shows no code text for any sample; every row below links to the file in its repository where the record names one.

“Ran” means the sample executed on a synthesized input; it does not mean the output is correct. “Honoured” and “violated” refer to a contract Syntology drafted from the code itself; “our draft was wrong” and “fixture could not drive it” are failures of Syntology's instrument, not of the code, and those samples did run. The ran count above is every status except unverified, the same rule as each paper page.

Papers

20 papers shown of 20, newest first; 22 places in the table. A paper with no recorded date is placed by the month its arXiv id encodes, shown in the Date column as YYYY-MM (from id). One row per place: a paper whose repository defines the name more than once appears more than once, and the same code body held for several papers appears once under each, with the same status. Titles and dates are the archive's archive 2025-07-28 for papers in the archive, and the graph's for 6 papers added by Syntology. Status and fingerprint are Syntology's record of each code body; licence is recorded for each place. The File cell ends with the code body's code_sha256, Syntology's identity for that exact code: an agent fetches the stored sample with Syntology's MCP tool get_code(code_sha256="…") (how to connect); click an identity to copy that call.

PaperDateFileStatus SyntologyLicence
Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning added by Syntology 2026-07 (from id) NVIDIA-NeMo/labs-molt/molt/models/utils.py 6c9f2419f316153a ran Apache-2.0 (permissive)
A Regret Minimization Framework on Preference Learning in Large Language Models added by Syntology 2026-06 (from id) OpenRLHF/OpenRLHF/openrlhf/models/utils.py c146f3cb0e5259a0 ran Apache-2.0 (permissive)
A Regret Minimization Framework on Preference Learning in Large Language Models added by Syntology 2026-06 (from id) YSLIU627/Regularized-Preference-Optimization/OpenRLHF/openrlhf/models/utils.py 60d73ddf6afc1bc3 ran fingerprinted Apache-2.0 (permissive)
GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models added by Syntology 2026-05 (from id) GaryBall/GDSD/gdsd/trainers/espo_trainer.py 773d9371a4bdce62 ran Apache-2.0 (permissive)
WIST: Web-Grounded Iterative Self-Play Tree for Domain-Targeted Reasoning Improvement added by Syntology 2026-03 (from id) lfy-123/WIST/openrlhf/models/utils.py d6e23db8c6a74238 unverified Apache-2.0 (permissive)
Align to the Pivot: Dual Alignment with Self-Feedback for Multilingual Math Reasoning added by Syntology 2026-01 (from id) Rover912/PASMR/openrlhf/models/utils.py de2fefc774bb0362 unverified Apache-2.0 (permissive)
Milestones over Outcome: Unlocking Geometric Reasoning with Sub-Goal Verifiable Reward added by Syntology 2026-01 (from id) FrontierX-Lab/SGVR/MM-EUREKA/openrlhf/models/utils.py 95084b4b49902e3d unverified Apache-2.0 (permissive)
TreeRL: LLM Reinforcement Learning with On-Policy Tree Search 13 Jun 2025 THUDM/TreeRL/openrlhf/models/utils.py afb9321e3d2d6a5f unverified Apache-2.0 (permissive)
Play to Generalize: Learning to Reason Through Game Play 9 Jun 2025 yunfeixie233/vigal/openrlhf/models/utils.py 95084b4b49902e3d unverified Apache-2.0 (permissive)
Teaching Language Models to Evolve with Users: Dynamic Profile Modeling for Personalized Alignment 21 May 2025 XingYuSSS/RLPA/openrlhf/models/utils.py 95084b4b49902e3d unverified Apache-2.0 (permissive)
VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning 10 Apr 2025 tiger-ai-lab/vl-rethinker/openrlhf/models/utils.py de2fefc774bb0362 unverified Apache-2.0 (permissive)
Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining 10 Apr 2025 rosieyzh/openrlhf-pretrain/openrlhf/models/utils.py 95084b4b49902e3d unverified Apache-2.0 (permissive)
Statistical Impossibility and Possibility of Aligning LLMs with Human Preferences: From Condorcet Paradox to Nash Equilibrium 14 Mar 2025 szk123456789/Nash_RS/openrlhf/models/utils.py 60d73ddf6afc1bc3 ran fingerprinted MIT (permissive)
Statistical Impossibility and Possibility of Aligning LLMs with Human Preferences: From Condorcet Paradox to Nash Equilibrium 14 Mar 2025 szk123456789/Nash_RS/openrlhf/models/NashRS_utils.py 477b4c962ccae21e unverified MIT (permissive)
Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning 20 Feb 2025 OpenLLMAI/OpenLLaMA2/openrlhf/models/utils.py c146f3cb0e5259a0 ran Apache-2.0 (permissive)
Demystifying Long Chain-of-Thought Reasoning in LLMs 5 Feb 2025 eddycmu/demystify-long-cot/openrlhf/openrlhf/models/utils.py b6035617946dd129 unverified MIT (permissive)
Offline Reinforcement Learning for LLM Multi-Step Reasoning 20 Dec 2024 jwhj/oreo/openrlhf/models/utils.py 60d73ddf6afc1bc3 ran fingerprinted Apache-2.0 (permissive)
Free Process Rewards without Process Labels 2 Dec 2024 lifan-yuan/implicitprm/train/openrlhf/models/utils.py b507127549cdea93 unverified Apache-2.0 (permissive)
Matryoshka: Learning to Drive Black-Box LLMs with LLMs 28 Oct 2024 lichangh20/Matryoshka/models/utils.py 60d73ddf6afc1bc3 ran fingerprinted no licence file found · pointer only
DPO Meets PPO: Reinforced Token Optimization for RLHF 29 Apr 2024 zkshan2002/RTO/openrlhf/models/utils.py b507127549cdea93 unverified Apache-2.0 (permissive)
InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling 14 Feb 2024 miaoyuchun/InfoRM/openrlhf/models/utils.py 95084b4b49902e3d unverified no licence file found · pointer only
Preference-grounded Token-level Guidance for Language Model Fine-tuning 1 Jun 2023 yinyueqin/denserewardrlhf-ppo/denserlhf/trainer/ppo_utils/experience_maker.py 8572f73b51db94c4 ran · our draft was wrong fingerprinted Apache-2.0 (permissive)

This site shows no code text; each File cell links to the file on GitHub at the repository's current default branch, which may have changed since the harvest. "Pointer only" means Syntology does not serve that copy's text, for one of four reasons: no licence file was found; the licence was not identified; the licence is recorded as permissive but that copy's record is not marked cleared; or the licence is outside the permissive list Syntology serves text under (MIT, Apache-2.0, BSD and similar). Some licences outside that list permit redistribution, such as WTFPL, and GPL-3.0 under its conditions; they are simply not on the list. Hover a licence cell for the reason. Per-sample records for a paper are on its paper page under "Code Syntology ran".

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections