Python‑Driven Robot Learns Better When Shown the Right Criteria
arXiv: 2611.08318v1
The study tests whether a robot‑coding agent performs better when given the success criteria for its tasks. In the experiments, a large language model writes and runs Python code that controls a robot on five development layouts. After each batch the agent’s learning module revises its library of skills based on the outcomes. The only difference between the two setups is whether the learner can see the success criteria and its scores. The “Criteria Exposure” (CE) setting shows them, while the “w/o CE” setting does not.
Figure 2. Figure 2 illustrates the self‑improvement loop: a frozen actor (GPT‑6 Astra) writes and runs Python cells on five development layouts, and a learner (Claude Code) revises the library ℒk based solely on episode evidence. The only switching point between the two experimental settings is whether the learner sees the success criteria and its scores.Huang, Wenbo, Bi, Jianxin, Yang, Shunyu, Si, Ruiyang, Ni, Rui, Wang, Qiang et al. · ·
On a set of 16 tasks whose success depends on the criteria, the CE model achieves a 51.6 % success rate under the task’s original criteria, compared with 27.8 % when the criteria are hidden. This contrasts with the 9 “faithful” tasks, where success rates are similar (52.5 % vs 52.2 %) regardless of CE.
Table listing four task groups (All, Criteria‑limited, Faithful) with success rates for empty library, without Criteria Exposure, and with Criteria Exposure. The 16‑task column shows 51.6 % success with CE and 27.8 % without CE, highlighting the effect of Criteria Exposure on these tasks.
Tasks
Empty library ℒ∅
w/o CE
w/ CE
All (25)
24.6
47.8 (+23.2)
62.0 (+37.4)
Criteria-limited (16)
8.4
27.8 (+19.4)
51.6 (+43.2)
Faithful (9)
53.3
83.3 (+30.0)
80.6 (+27.3)
Table 1. Table 1 shows test success under the original criteria for the 16 criteria‑limited tasks: 51.6 % when the library was trained with Criteria Exposure (CE) versus 27.8 % when CE was absent.Huang, Wenbo, Bi, Jianxin, Yang, Shunyu, Si, Ruiyang, Ni, Rui, Wang, Qiang et al. · arXiv:2611.08318v1 · Table 1: · CC BY 4.0
The gain from CE is limited to tasks whose criteria differ from the plain instruction. When the criteria are only slightly altered, the performance advantage disappears. Moreover, libraries trained with CE do not transfer well to new tasks or other benchmarks; the benefit vanishes when applied to unseen tasks.
This work demonstrates that showing an agent the exact verification rules can sharpen its coding toward higher success on those specific rules, but it does not universally improve general problem solving or transfer to new contexts.
Python‑Driven Robot Learns Better When Shown the Right Criteria | arXiv News
