The Pulse
Apple-Linked RISED Method Trains Agents Across Three Environments
A paper proposes using behavioral rubrics to select training examples and guide a language-model agent across interactive tasks. In tests on shopping, database and household-task environments, the authors report higher average pass rates th
AI.info Team · Written with AI ·
“Both challenges highlight limitations of relying solely on scalar rewards in multi-environment RL: they provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical.”
Jingtan Wang and coauthors, authors of “RISED: RubrIcs for agentic multi-environment Selection and sElf-Distillation” (paper)
A paper submitted to arXiv on October 1 proposes a way to train one language-model agent across different interactive tasks by using behavioral rubrics to choose training examples and guide the model’s own outputs. The method, RISED, is tested on shopping, database and household-task environments, where its authors report higher average pass rates than five baseline approaches on two model backbones.
The author list includes Jingtan Wang, Sirajul Salekin, Young mok Jung, Javier Movellan, Bryan Kian Hsiang Low and Manjot Bilkhu. Apple’s research page marks Jingtan Wang—not Manjot Bilkhu—as having completed work while at Apple; it also identifies National University of Singapore affiliations for Wang and Bryan Kian Hsiang Low. The paper describes a research method, not a product announcement.
Rubrics compare behavior across tasks
Multi-environment training can leave ordinary reward signals short of useful feedback. When every attempt at a prompt fails—or every attempt succeeds—the rewards offer no within-group contrast to tell the model which behaviors to repeat. The paper says such all-failure and all-success groups made up more than half of candidate prompt groups in its experiments.
RISED gives rollouts a shared behavioral description. An LLM judge assigns tags from a predefined vocabulary, including positive behaviors to demonstrate and negative behaviors to avoid. The authors build a vocabulary of 82 rubrics across 11 skills from warm-up and teacher rollouts; for the three-environment experiments, they use 59 rubrics across 10 skills.
The method uses those tags in two ways. First, it selects prompt groups whose combined behavioral profiles fit the broader mix of current rollouts, while penalizing overlap with examples already selected. Second, it conditions a self-distillation teacher on positive rubrics, giving the student token-level guidance; negative rubrics can steer subsequent training rollouts away from recurring failure modes.
RISED’s gains came with extra training time
The main tests train on ALFWorld, WebShop and DBBench, which cover household-object interaction, online shopping and database tasks. The paper evaluates Qwen2.5-3B-Instruct and Qwen3-4B, and compares RISED with five baselines, including standard group-relative policy optimization and methods that select examples based on reward variance.
On Qwen2.5-3B, RISED reaches an overall pass rate of 0.542, against 0.484 for the standard 64-prompt baseline and 0.530 for the strongest competing method reported on that backbone. On Qwen3-4B, it scores 0.585, compared with 0.516 for the standard baseline and 0.579 for the strongest competitor. Those are gains of 5.8 and 6.9 percentage points over the standard baseline, and 1.2 and 0.6 points over the strongest comparison, respectively.
The authors also report that RISED reaches the standard baseline’s final overall pass rate in roughly half as many optimizer steps in the Qwen2.5-3B experiment. The result is not a simple reduction in compute: in the measured setup, a training step takes 252.9 seconds for RISED, versus 155.5 seconds for the standard baseline. The paper attributes the added time mainly to rubric tagging and the teacher forward pass, and notes that its timing compares elapsed time rather than total compute because RISED uses a dedicated judge server.
Evidence remains tied to the tested setup
RISED’s results are limited to the paper’s three benchmarks and two tested backbones; they do not establish that the method improves agents in every environment or with every model. The authors also flag tagging and teacher inference as added overhead, and call for work to improve efficiency and test larger policies and more environments.
The study’s narrower contribution is a training recipe: compare agent rollouts by shared descriptions of what they do, not only by whether they earn the same scalar reward. In the reported trials, that combination of example selection and rubric-guided self-distillation produces the highest mean pass rate among the compared methods on both backbones.