{"built":"2026-10-15","rows":[{"band":"Hard DecisionGym tasks","label":"Chosen combination: tool and argument cascades + reflection trigger","who":"jev","n":200.0,"ref_success":66.0,"with_success":69.5,"delta_pt":3.5,"ref_time_s":41.6,"with_time_s":71.5,"src":{"n":{"key":"e7b/jevagent_best/n","file":"research/paper/generated/e7b_hard.tex"},"ref_success":{"key":"e7b/llm_only/success","file":"research/paper/generated/e7b_hard.tex"},"with_success":{"key":"e7b/jevagent_best/success","file":"research/paper/generated/e7b_hard.tex"},"delta_pt":{"key":"e7b/jevagent_best/delta","file":"research/paper/generated/e7b_hard.tex"},"ref_time_s":{"key":"e7b/llm_only/wall_mean","file":"research/paper/generated/e7b_hard.tex"},"with_time_s":{"key":"e7b/jevagent_best/wall_mean","file":"research/paper/generated/e7b_hard.tex"},"saving_pct":{"key":"e7b/jevagent_best/saving","file":"research/paper/generated/e7b_hard.tex"}},"saving_pct":4.6},{"band":"Hard DecisionGym tasks","label":"Tool cascade refitted on the new rule","who":"jev","n":100.0,"ref_success":87.0,"with_success":89.0,"delta_pt":2.0,"ref_time_s":63.5,"with_time_s":104.1,"src":{"n":{"key":"e7e/jev_refit/n","file":"research/paper/generated/e7e_hard.tex"},"ref_success":{"key":"e7e/llm_only/success","file":"research/paper/generated/e7e_hard.tex"},"with_success":{"key":"e7e/jev_refit/success","file":"research/paper/generated/e7e_hard.tex"},"delta_pt":{"key":"e7e/jev_refit/delta","file":"research/paper/generated/e7e_hard.tex"},"ref_time_s":{"key":"e7e/llm_only/wall_mean","file":"research/paper/generated/e7e_hard.tex"},"with_time_s":{"key":"e7e/jev_refit/wall_mean","file":"research/paper/generated/e7e_hard.tex"},"saving_pct":{"key":"e7e/jev_refit/saving","file":"research/paper/generated/e7e_hard.tex"}},"saving_pct":30.5},{"band":"Hard DecisionGym tasks","label":"GPT-6-Luna as agent LLM: tool cascade refitted","who":"jev","n":100.0,"ref_success":45.0,"with_success":83.0,"delta_pt":38.0,"ref_time_s":53.9,"with_time_s":67.4,"src":{"n":{"key":"e13/luna/jev_refit/n","file":"research/paper/generated/e13_llms.tex"},"ref_success":{"key":"e13/luna/llm_only/success","file":"research/paper/generated/e13_llms.tex"},"with_success":{"key":"e13/luna/jev_refit/success","file":"research/paper/generated/e13_llms.tex"},"delta_pt":{"key":"e13/luna/jev_refit/delta","file":"research/paper/generated/e13_llms.tex"},"ref_time_s":{"key":"e13/luna/llm_only/wall_mean","file":"research/paper/generated/e13_llms.tex"},"with_time_s":{"key":"e13/luna/jev_refit/wall_mean","file":"research/paper/generated/e13_llms.tex"},"saving_pct":{"key":"e13/luna/jev_refit/saving","file":"research/paper/generated/e13_llms.tex"}},"saving_pct":-737.4,"cost_multiple":8.4,"cost_multiple_src":[{"key":"e13/luna/jev_refit/cost_k","file":"research/paper/generated/e13_llms.tex"},{"key":"e13/luna/jev_refit/cost_ref_k","file":"research/paper/generated/e13_llms.tex"}]},{"band":"Coding agents (strong or fast LLM per step)","label":"mini-SWE-agent, SWE-bench Verified: DeepSeek-V4-Pro or -Flash","who":"jev","n":300.0,"ref_success":77.3,"with_success":76.3,"delta_pt":-1.0,"ref_time_s":457.0,"with_time_s":436.0,"src":{"n":{"key":"e4r/n_eval","file":"research/paper/generated/e4r_swe.tex"},"ref_success":{"key":"e4r/always_deep/success","file":"research/paper/generated/e4r_swe.tex"},"with_success":{"key":"e4r/jevagent/success","file":"research/paper/generated/e4r_swe.tex"},"delta_pt":{"key":"e4r/jevagent/vs_deep/delta","file":"research/paper/generated/e4r_swe.tex"},"ref_time_s":{"key":"e4r/always_deep/wall_mean_s","file":"research/paper/generated/e4r_swe.tex"},"with_time_s":{"key":"e4r/jevagent/wall_mean_s","file":"research/paper/generated/e4r_swe.tex"},"saving_pct":{"key":"e4r/jevagent/vs_deep/saving_pct","file":"research/paper/generated/e4r_swe.tex"}},"saving_pct":10.5},{"band":"Coding agents (strong or fast LLM per step)","label":"mini-SWE-agent: Gemini 3.8 Flash, high or low thinking","who":"jev","n":100.0,"ref_success":70.0,"with_success":75.0,"delta_pt":5.0,"ref_time_s":886.0,"with_time_s":728.0,"src":{"n":{"key":"e4b/jevagent/n","file":"research/paper/generated/e4b_swe.tex"},"ref_success":{"key":"e4b/always_deep/success","file":"research/paper/generated/e4b_swe.tex"},"with_success":{"key":"e4b/jevagent/success","file":"research/paper/generated/e4b_swe.tex"},"delta_pt":{"key":"e4b/jevagent/vs_deep/delta","file":"research/paper/generated/e4b_swe.tex"},"ref_time_s":{"key":"e4b/always_deep/wall_mean_s","file":"research/paper/generated/e4b_swe.tex"},"with_time_s":{"key":"e4b/jevagent/wall_mean_s","file":"research/paper/generated/e4b_swe.tex"},"saving_pct":{"key":"e4b/jevagent/vs_deep/saving_pct","file":"research/paper/generated/e4b_swe.tex"}},"saving_pct":7.0},{"band":"Coding agents (strong or fast LLM per step)","label":"SWE-agent, SWE-bench Verified: Gemini 3.5 Flash or Flash-Lite","who":"jev","n":100.0,"ref_success":63.0,"with_success":57.0,"delta_pt":-6.0,"ref_time_s":110.0,"with_time_s":124.0,"src":{"n":{"key":"e25/swe/jev_router/n","file":"research/paper/generated/e25_harnesses.tex"},"ref_success":{"key":"e25/swe/baseline/success","file":"research/paper/generated/e25_harnesses.tex"},"with_success":{"key":"e25/swe/jev_router/success","file":"research/paper/generated/e25_harnesses.tex"},"delta_pt":{"key":"e25/swe/jev_router/delta","file":"research/paper/generated/e25_harnesses.tex"},"ref_time_s":{"key":"e25/swe/baseline/wall_mean_s","file":"research/paper/generated/e25_harnesses.tex"},"with_time_s":{"key":"e25/swe/jev_router/wall_mean_s","file":"research/paper/generated/e25_harnesses.tex"},"saving_pct":{"key":"e25/swe/jev_router/saving","file":"research/paper/generated/e25_harnesses.tex"}},"saving_pct":7.7},{"band":"Coding agents (strong or fast LLM per step)","label":"pi, Terminal-Bench 2.0: Gemini 3.5 Flash or Flash-Lite","who":"jev","n":89.0,"ref_success":42.7,"with_success":50.6,"delta_pt":7.9,"ref_time_s":257.0,"with_time_s":260.0,"src":{"n":{"key":"e25/tb/jev_router/n","file":"research/paper/generated/e25_harnesses.tex"},"ref_success":{"key":"e25/tb/baseline/success","file":"research/paper/generated/e25_harnesses.tex"},"with_success":{"key":"e25/tb/jev_router/success","file":"research/paper/generated/e25_harnesses.tex"},"delta_pt":{"key":"e25/tb/jev_router/delta","file":"research/paper/generated/e25_harnesses.tex"},"ref_time_s":{"key":"e25/tb/baseline/wall_mean_s","file":"research/paper/generated/e25_harnesses.tex"},"with_time_s":{"key":"e25/tb/jev_router/wall_mean_s","file":"research/paper/generated/e25_harnesses.tex"},"saving_pct":{"key":"e25/tb/jev_router/saving","file":"research/paper/generated/e25_harnesses.tex"}},"saving_pct":9.6},{"band":"ALFWorld text games","label":"Action cascade with Jev","who":"jev","n":140.0,"ref_success":91.4,"with_success":91.4,"delta_pt":0.0,"ref_time_s":24.1,"with_time_s":22.3,"src":{"n":{"key":"e6a/cascade/n","file":"research/paper/generated/e6a.tex"},"ref_success":{"key":"e6a/llm_only/success","file":"research/paper/generated/e6a.tex"},"with_success":{"key":"e6a/cascade/success","file":"research/paper/generated/e6a.tex"},"delta_pt":{"key":"e6a/cascade/delta","file":"research/paper/generated/e6a.tex"},"ref_time_s":{"key":"e6a/llm_only/wall_mean","file":"research/paper/generated/e6a.tex"},"with_time_s":{"key":"e6a/cascade/wall_mean","file":"research/paper/generated/e6a.tex"},"saving_pct":{"key":"e6a/cascade/saving","file":"research/paper/generated/e6a.tex"}},"saving_pct":14.0},{"band":"ALFWorld text games","label":"Action cascade with a 4B model trained on its states","who":"ours","n":140.0,"ref_success":90.7,"with_success":95.0,"delta_pt":4.3,"ref_time_s":37.0,"with_time_s":18.2,"src":{"n":{"key":"e11/cascade/n","file":"research/paper/generated/e11_alfworld.tex"},"ref_success":{"key":"e11/llm_only/success","file":"research/paper/generated/e11_alfworld.tex"},"with_success":{"key":"e11/cascade/success","file":"research/paper/generated/e11_alfworld.tex"},"delta_pt":{"key":"e11/cascade/delta","file":"research/paper/generated/e11_alfworld.tex"},"ref_time_s":{"key":"e11/llm_only/wall_mean","file":"research/paper/generated/e11_alfworld.tex"},"with_time_s":{"key":"e11/cascade/wall_mean","file":"research/paper/generated/e11_alfworld.tex"},"saving_pct":{"key":"e11/cascade/saving","file":"research/paper/generated/e11_alfworld.tex"}},"saving_pct":89.0},{"band":"ALFWorld text games","label":"Unseen rooms: the same model and threshold","who":"ours","n":134.0,"ref_success":88.8,"with_success":89.6,"delta_pt":0.7,"ref_time_s":35.1,"with_time_s":18.2,"src":{"n":{"key":"e23/cascade/n","file":"research/paper/generated/e23_alfworld_unseen.tex"},"ref_success":{"key":"e23/llm_only/success","file":"research/paper/generated/e23_alfworld_unseen.tex"},"with_success":{"key":"e23/cascade/success","file":"research/paper/generated/e23_alfworld_unseen.tex"},"delta_pt":{"key":"e23/cascade/delta","file":"research/paper/generated/e23_alfworld_unseen.tex"},"ref_time_s":{"key":"e23/llm_only/wall_mean","file":"research/paper/generated/e23_alfworld_unseen.tex"},"with_time_s":{"key":"e23/cascade/wall_mean","file":"research/paper/generated/e23_alfworld_unseen.tex"},"saving_pct":{"key":"e23/cascade/saving","file":"research/paper/generated/e23_alfworld_unseen.tex"}},"saving_pct":92.0}],"link":"https://github.com/JevAgentHarness/JevAgentHarness/blob/main/research/paper/generated/","note":"Table 1 of the paper: the LLM alone (ref.) against the same LLM with the delegation."}