Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

reinforcement learning with verifiable rewards

Reinforcement learning rewards actions that lead to good outcomes. With verifiable rewards, the outcome is checked automatically, such as a correct answer or a completed task, not judged by a human. The signal is clean but often sparse: the agent only hears at the end, and if it never succeeds, there is little to learn.

Across domains, the papers attack that sparseness by pre-solving environments, warming up with a teacher, shaping rewards from progress, directing exploration, using self-written insights, decomposing checks, sequencing distillation before reinforcement, and training on memory trees. These turn weak or absent final rewards into useful learning signals, beating plain terminal-only feedback.

Papers
160
Released code
13
First seen
June 2026
Latest
Nov 2026

152 papers in the last two months, against 7 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Small language models outperform 763B model at analog circuit design

    A small locally-deployed model fine-tuned with reinforcement learning outperforms a massive cloud API model on multi-turn analog circuit optimization.

    Nov 2026
  2. 2AI teacher adapts its hints as a language model learns, boosting performance across four domains

    AutoScaffold uses a frozen teacher LLM to automatically design adaptive training scaffolds that improve language model RL performance across domains.

    Nov 2026
  3. 3Language models taught by examples can generalize better than reward-trained models

    Supervised fine-tuning can match reinforcement learning generalization when trained on diverse data at sufficient scale.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

group relative policy optimizationA reinforcement learning method that optimizes policy by computing advantages based on group-level reward statistics.58 togethercredit assignmentDetermining which actions or steps in a sequence contributed to a final outcome, often for reinforcement learning.40 togetherverificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.27 together

Who is working on it

Peking University 7University of Chinese Academy of Sciences 7Zhejiang University 7Tencent 7Tsinghua University 6Alibaba Group 6

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.

reinforcement learning with verifiable rewards · Research concepts | Zotpaper