{"built":"2026-10-15","source":{"dataset":"https://huggingface.co/datasets/JevAgentHarness/AgentDecisionBench","configs":{"rows":"leaderboard (leaderboard/headline_numbers.jsonl.gz)","models":"models (models.jsonl.gz)"},"note":"Rows copied verbatim; each row names the paper macro (tex_key, tex_file) that holds its value."},"models":[{"model":"jev","name":"Jev (typesafe/jev-1.13)","kind":"teacher","notes":"commercial decision model (TypeSafe), via OpenRouter; the teacher of the training data","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"rl","name":"AgentDecision-4B v1","kind":"ours","notes":"Qwen3-4B + SFT + GRPO on agent states; the checkpoint frozen before Stage 2 (code: JevAgentRL-4B, frozen, rl)","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e12","name":"AgentDecision-4B","kind":"ours","notes":"the released 4B: AgentDecision-4B v1 continued on public tool-use data (E12, 'mixed')","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e28","name":"AgentDecision-0.8B","kind":"ours","notes":"StartLux-Decision-0.8B + the recipe (E28); base licence CC BY-NC 4.0","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e28_vendor","name":"AgentDecision-0.8B (vendor serving engine)","kind":"ours","notes":"the same checkpoint as e28 served by the vendor's engine (cost basis of the 0.8B in the paper)","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"base","name":"Qwen3-4B (untrained)","kind":"ablation","notes":"the 4B's base model before training","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"sft_v1","name":"Qwen3-4B + SFT (v1)","kind":"ablation","notes":"SFT stage of the recipe, before GRPO","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"v2","name":"JevAgentRL-v2 (exploratory)","kind":"ablation","notes":"SFT v2 on StartLux-4B trained with Miles (exploratory track A7.19)","isolated_L1":false,"common_states_L2":true,"shift_L2":false},{"model":"e12c","name":"StartLux-4B + recipe","kind":"ablation","notes":"StartLux-Decision-4B + the recipe with public tool-use data (E12c); base licence CC BY-NC 4.0","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e16a","name":"AgentDecision-4B (mixed+)","kind":"ablation","notes":"E16 arm A: longer continuation on public data (appendix only)","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"e16b","name":"StartLux-9B + recipe","kind":"ablation","notes":"E16 arm B (appendix only); base licence CC BY-NC 4.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"e31","name":"Qwen3-4B + GRPO without SFT","kind":"ablation","notes":"E31","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e7d_tier","name":"AgentDecision-4B v1 + tier-rule training","kind":"ablation","notes":"E7d (DecisionGym-Hard rule); offline only","isolated_L1":false,"common_states_L2":true,"shift_L2":false},{"model":"e14a","name":"Qwen3.5-0.8B scoring head (agent states)","kind":"ablation","notes":"E14 arm A","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e14ag","name":"Qwen3.5-0.8B scoring head (agent states + generic mix)","kind":"ablation","notes":"E14 arm AG","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"e14g","name":"Qwen3.5-0.8B scoring head (generic mix only)","kind":"ablation","notes":"E14 arm G","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"e17gen","name":"Qwen3.5-0.8B one-token read-out","kind":"ablation","notes":"E17 generative read-out on the same rows as E14 A","isolated_L1":false,"common_states_L2":true,"shift_L2":false},{"model":"e18f","name":"Qwen3.5-0.8B with a trained defer option","kind":"ablation","notes":"E18 arm F (appendix only)","isolated_L1":false,"common_states_L2":true,"shift_L2":false},{"model":"e27s","name":"Qwen3.5-0.8B + surface augmentation","kind":"ablation","notes":"E27 arm S","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e27sp","name":"Qwen3.5-0.8B + surface augmentation + public data","kind":"ablation","notes":"E27 arm SP","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"e5_q17","name":"Qwen3-1.7B student","kind":"ablation","notes":"E5 small decision models","isolated_L1":false,"common_states_L2":true,"shift_L2":false},{"model":"e5_q06","name":"Qwen3-0.6B student","kind":"ablation","notes":"E5 small decision models","isolated_L1":false,"common_states_L2":true,"shift_L2":false},{"model":"e5_mc1","name":"MiniCPM5-1B student","kind":"ablation","notes":"E5 small decision models","isolated_L1":false,"common_states_L2":true,"shift_L2":false},{"model":"intern","name":"Intern-Decision-4B","kind":"off_the_shelf","notes":"internlm/Intern-Decision-4B","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"sl0.8b","name":"StartLux-Decision-0.8B","kind":"off_the_shelf","notes":"CC BY-NC 4.0 weights","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"sl2b","name":"StartLux-Decision-2B","kind":"off_the_shelf","notes":"CC BY-NC 4.0 weights","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"sl4b","name":"StartLux-Decision-4B","kind":"off_the_shelf","notes":"CC BY-NC 4.0 weights","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"sl9b","name":"StartLux-Decision-9B","kind":"off_the_shelf","notes":"CC BY-NC 4.0 weights","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"sl27b","name":"StartLux-Decision-27B (Q8_0)","kind":"off_the_shelf","notes":"CC BY-NC 4.0 weights, Q8_0 GGUF","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"oaidec","name":"OpenAI Decisions API (gpt-6-luna)","kind":"off_the_shelf","notes":"proprietary API, POST /v1/decisions","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"openjev","name":"OpenJev (DiffusionGemma-26B-A4B)","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"openjevhf","name":"OpenJev-27B (HF, FP8)","kind":"off_the_shelf","notes":"CC BY-NC 4.0 weights","isolated_L1":true,"common_states_L2":true,"shift_L2":true},{"model":"jevk5","name":"JevK5 v0.2 (4B)","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"clm","name":"CLM v0.1 (8B)","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"laya","name":"Laya (421M)","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"verdict","name":"Verdict 1.4 (151M)","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"decider","name":"decider-4b v2.1","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"imajev","name":"Imajev-4B","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"bosun","name":"Bosun 3.1 1.7B","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"clef","name":"Clef 27B (8-bit)","kind":"off_the_shelf","notes":"Apache-2.0; isolated level only","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"clefflash","name":"Clef-Flash 9B","kind":"off_the_shelf","notes":"Apache-2.0","isolated_L1":true,"common_states_L2":true,"shift_L2":false},{"model":"flash","name":"Gemini 3.5 Flash (multiple choice)","kind":"llm","notes":"the agent LLM answering the typed question","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"lite","name":"Gemini 3.5 Flash-Lite (multiple choice, distribution)","kind":"llm","notes":"","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"litelabel","name":"Gemini 3.5 Flash-Lite (label + confidence)","kind":"llm","notes":"","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"gemma","name":"Gemma 4 26B (multiple choice)","kind":"llm","notes":"unpriced","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"flashfc","name":"Gemini 3.5 Flash, own function call","kind":"control","notes":"native function calling, not a multiple-choice decider","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"embed","name":"Gemini embedding ranker","kind":"control","notes":"","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"bm25","name":"BM25 keyword ranker","kind":"control","notes":"","isolated_L1":true,"common_states_L2":false,"shift_L2":false},{"model":"tfidf","name":"TF-IDF ranker","kind":"control","notes":"","isolated_L1":true,"common_states_L2":false,"shift_L2":false}],"rows":[{"model":"base","model_name":"Qwen3-4B (untrained)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"71.4","value_num":71.4,"paper_display":"Table 3","tex_key":"e9/base/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"base","model_name":"Qwen3-4B (untrained)","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"52.9","value_num":52.9,"paper_display":"Table 3","tex_key":"e21/base/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"base","model_name":"Qwen3-4B (untrained)","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"27.6","value_num":27.6,"paper_display":"Table 3","tex_key":"e21/base/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"base","model_name":"Qwen3-4B (untrained)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"13.5","value_num":13.5,"paper_display":"Table 3","tex_key":"e21/base/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"sft_v1","model_name":"Qwen3-4B + SFT (v1)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"78.4","value_num":78.4,"paper_display":"Table 3","tex_key":"e9/sftv1/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"sft_v1","model_name":"Qwen3-4B + SFT (v1)","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"98.4","value_num":98.4,"paper_display":"Table 3","tex_key":"e21/sft_v1/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"sft_v1","model_name":"Qwen3-4B + SFT (v1)","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"95.0","value_num":95.0,"paper_display":"Table 3","tex_key":"e21/sft_v1/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"sft_v1","model_name":"Qwen3-4B + SFT (v1)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"96.0","value_num":96.0,"paper_display":"Table 3","tex_key":"e21/sft_v1/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"rl","model_name":"AgentDecision-4B v1","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"78.8","value_num":78.8,"paper_display":"Table 3","tex_key":"e9/rl/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"rl","model_name":"AgentDecision-4B v1","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"98.6","value_num":98.6,"paper_display":"Table 3","tex_key":"e21/rl/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"rl","model_name":"AgentDecision-4B v1","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"90.8","value_num":90.8,"paper_display":"Table 3","tex_key":"e21/rl/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"rl","model_name":"AgentDecision-4B v1","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"96.0","value_num":96.0,"paper_display":"Table 3","tex_key":"e21/rl/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"e12","model_name":"AgentDecision-4B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"85.1","value_num":85.1,"paper_display":"Table 3","tex_key":"e9/e12/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"e12","model_name":"AgentDecision-4B","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"94.9","value_num":94.9,"paper_display":"Table 3","tex_key":"e21/e12/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"e12","model_name":"AgentDecision-4B","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"67.4","value_num":67.4,"paper_display":"Table 3","tex_key":"e21/e12/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"e12","model_name":"AgentDecision-4B","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"91.5","value_num":91.5,"paper_display":"Table 3","tex_key":"e21/e12/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"86.4","value_num":86.4,"paper_display":"Table 3","tex_key":"e9/jev/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"91.8","value_num":91.8,"paper_display":"Table 3","tex_key":"e21/jev/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"58.3","value_num":58.3,"paper_display":"Table 3","tex_key":"e21/jev/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"82.0","value_num":82.0,"paper_display":"Table 3","tex_key":"e21/jev/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"73.4","value_num":73.4,"paper_display":"Table 3","tex_key":"e9/intern/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"75.9","value_num":75.9,"paper_display":"Table 3","tex_key":"e21/intern/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"33.9","value_num":33.9,"paper_display":"Table 3","tex_key":"e21/intern/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"34.0","value_num":34.0,"paper_display":"Table 3","tex_key":"e21/intern/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"83.1","value_num":83.1,"paper_display":"Table 3","tex_key":"e9/sl4b/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"86.5","value_num":86.5,"paper_display":"Table 3","tex_key":"e21/sl4b/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"36.7","value_num":36.7,"paper_display":"Table 3","tex_key":"e21/sl4b/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"49.0","value_num":49.0,"paper_display":"Table 3","tex_key":"e21/sl4b/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"e12c","model_name":"StartLux-4B + recipe","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"85.5","value_num":85.5,"paper_display":"Table 3","tex_key":"e9/e12c/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"e12c","model_name":"StartLux-4B + recipe","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"98.5","value_num":98.5,"paper_display":"Table 3","tex_key":"e21/e12c/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"e12c","model_name":"StartLux-4B + recipe","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"100.0","value_num":100.0,"paper_display":"Table 3","tex_key":"e21/e12c/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"e12c","model_name":"StartLux-4B + recipe","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"98.0","value_num":98.0,"paper_display":"Table 3","tex_key":"e21/e12c/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl0.8b","model_name":"StartLux-Decision-0.8B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"70.0","value_num":70.0,"paper_display":"Table 3","tex_key":"e9/sl0.8b/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"sl0.8b","model_name":"StartLux-Decision-0.8B","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"51.7","value_num":51.7,"paper_display":"Table 3","tex_key":"e28/base/l2/all","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"sl0.8b","model_name":"StartLux-Decision-0.8B","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"27.2","value_num":27.2,"paper_display":"Table 3","tex_key":"e28/base/l2/mustask","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"sl0.8b","model_name":"StartLux-Decision-0.8B","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"13.0","value_num":13.0,"paper_display":"Table 3","tex_key":"e28/base/l3/success","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"e28","model_name":"AgentDecision-0.8B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"80.5","value_num":80.5,"paper_display":"Table 3","tex_key":"e9/e28/acc","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"e28","model_name":"AgentDecision-0.8B","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"95.6","value_num":95.6,"paper_display":"Table 3","tex_key":"e28/l2/all","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"e28","model_name":"AgentDecision-0.8B","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"93.5","value_num":93.5,"paper_display":"Table 3","tex_key":"e28/l2/mustask","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"e28","model_name":"AgentDecision-0.8B","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"82.0","value_num":82.0,"paper_display":"Table 3","tex_key":"e28/l3/success","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"e12","model_name":"AgentDecision-4B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"89.3","value_num":89.3,"paper_display":"Figure 4b","tex_key":"e12/e12/e1/acc/toolchoice","tex_file":"paper/generated/e12_train.tex"},{"model":"e28","model_name":"AgentDecision-0.8B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"87.1","value_num":87.1,"paper_display":"Figure 4b","tex_key":"e28/l1/tool","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"89.9","value_num":89.9,"paper_display":"Figure 4b","tex_key":"explore/startlux_live/jev/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"oaidec","model_name":"OpenAI Decisions API (gpt-6-luna)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"90.6","value_num":90.6,"paper_display":"Figure 4b","tex_key":"e15/oaidec/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjevhf","model_name":"OpenJev-27B (HF, FP8)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"90.2","value_num":90.2,"paper_display":"Figure 4b","tex_key":"e15/openjevhf/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"90.8","value_num":90.8,"paper_display":"Figure 4b","tex_key":"explore/startlux_live/sl27b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"92.2","value_num":92.2,"paper_display":"Figure 4b","tex_key":"explore/startlux_live/sl9b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"91.3","value_num":91.3,"paper_display":"Figure 4b","tex_key":"explore/startlux_live/sl4b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"88.2","value_num":88.2,"paper_display":"Figure 4b","tex_key":"explore/startlux_live/intern/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"oaidec","model_name":"OpenAI Decisions API (gpt-6-luna)","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"86.1","value_num":86.1,"paper_display":"Figure 4b","tex_key":"e21/oaidec/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"oaidec","model_name":"OpenAI Decisions API (gpt-6-luna)","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"41.7","value_num":41.7,"paper_display":"Appendix / text","tex_key":"e21/oaidec/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"oaidec","model_name":"OpenAI Decisions API (gpt-6-luna)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"66.0","value_num":66.0,"paper_display":"Figure 4b","tex_key":"e21/oaidec/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"openjevhf","model_name":"OpenJev-27B (HF, FP8)","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"92.4","value_num":92.4,"paper_display":"Figure 4b","tex_key":"e21/openjevhf/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"openjevhf","model_name":"OpenJev-27B (HF, FP8)","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"58.3","value_num":58.3,"paper_display":"Appendix / text","tex_key":"e21/openjevhf/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"openjevhf","model_name":"OpenJev-27B (HF, FP8)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"69.5","value_num":69.5,"paper_display":"Figure 4b","tex_key":"e21/openjevhf/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"94.7","value_num":94.7,"paper_display":"Figure 4b","tex_key":"e21/sl27b/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"65.0","value_num":65.0,"paper_display":"Appendix / text","tex_key":"e21/sl27b/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"79.9","value_num":79.9,"paper_display":"Figure 4b","tex_key":"e21/sl27b/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"90.9","value_num":90.9,"paper_display":"Figure 4b","tex_key":"e21/sl9b/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L2","task_set":"DecisionGym-Shift open-loop states, must-ask subset","metric":"step accuracy, must-ask states (%)","value":"55.0","value_num":55.0,"paper_display":"Appendix / text","tex_key":"e21/sl9b/l2/mustask","tex_file":"paper/generated/e21_shift.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"69.5","value_num":69.5,"paper_display":"Figure 4b","tex_key":"e21/sl9b/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"e14a","model_name":"Qwen3.5-0.8B scoring head (agent states)","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"73.2","value_num":73.2,"paper_display":"Appendix / text","tex_key":"e21/e14a/l2/all","tex_file":"paper/generated/e21_shift.tex"},{"model":"e14a","model_name":"Qwen3.5-0.8B scoring head (agent states)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"17.5","value_num":17.5,"paper_display":"Appendix / text","tex_key":"e21/e14a/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"e31","model_name":"Qwen3-4B + GRPO without SFT","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"71.4","value_num":71.4,"paper_display":"Appendix / text","tex_key":"e31/rlonly/l1/acc","tex_file":"paper/generated/e31_rl_only.tex"},{"model":"e31","model_name":"Qwen3-4B + GRPO without SFT","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"50.2","value_num":50.2,"paper_display":"Appendix / text","tex_key":"e31/rlonly/l2/all","tex_file":"paper/generated/e31_rl_only.tex"},{"model":"e31","model_name":"Qwen3-4B + GRPO without SFT","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"11.5","value_num":11.5,"paper_display":"Appendix / text","tex_key":"e31/rlonly/l3/success","tex_file":"paper/generated/e31_rl_only.tex"},{"model":"e27s","model_name":"Qwen3.5-0.8B + surface augmentation","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"67.7","value_num":67.7,"paper_display":"Appendix / text","tex_key":"e27/s/l1/acc","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"e27s","model_name":"Qwen3.5-0.8B + surface augmentation","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"93.7","value_num":93.7,"paper_display":"Appendix / text","tex_key":"e27/s/l2/all","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"e27s","model_name":"Qwen3.5-0.8B + surface augmentation","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"72.0","value_num":72.0,"paper_display":"Appendix / text","tex_key":"e27/s/l3/success","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"e27sp","model_name":"Qwen3.5-0.8B + surface augmentation + public data","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"77.2","value_num":77.2,"paper_display":"Appendix / text","tex_key":"e27/sp/l1/acc","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"e27sp","model_name":"Qwen3.5-0.8B + surface augmentation + public data","level":"L2","task_set":"DecisionGym-Shift open-loop states (100 tasks)","metric":"step accuracy, all states (%)","value":"96.3","value_num":96.3,"paper_display":"Appendix / text","tex_key":"e27/sp/l2/all","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"e27sp","model_name":"Qwen3.5-0.8B + surface augmentation + public data","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%)","value":"80.0","value_num":80.0,"paper_display":"Appendix / text","tex_key":"e27/sp/l3/success","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"none","model_name":"Gemini 3.5 Flash alone (no decision model)","level":"L3","task_set":"DecisionGym-Shift closed-loop (200 tasks)","metric":"task success (%), agent LLM alone (Gemini 3.5 Flash)","value":"100.0","value_num":100.0,"paper_display":"Appendix / text","tex_key":"e21/llm_only/l3/success","tex_file":"paper/generated/e21_shift.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"89.9","value_num":89.9,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/jev/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"91.5","value_num":91.5,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/jev/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"jev","model_name":"Jev (typesafe/jev-1.13)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"83.0","value_num":83.0,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/jev/live_success","tex_file":"paper/generated/explore.tex"},{"model":"rl","model_name":"AgentDecision-4B v1","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"88.4","value_num":88.4,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/frozen/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"rl","model_name":"AgentDecision-4B v1","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.2","value_num":99.2,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/frozen/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"rl","model_name":"AgentDecision-4B v1","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"100.0","value_num":100.0,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/frozen/live_success","tex_file":"paper/generated/explore.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"88.2","value_num":88.2,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/intern/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"75.2","value_num":75.2,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/intern/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"intern","model_name":"Intern-Decision-4B","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"26.5","value_num":26.5,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/intern/live_success","tex_file":"paper/generated/explore.tex"},{"model":"sl0.8b","model_name":"StartLux-Decision-0.8B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"88.2","value_num":88.2,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl0.8b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl0.8b","model_name":"StartLux-Decision-0.8B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"48.3","value_num":48.3,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl0.8b/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl0.8b","model_name":"StartLux-Decision-0.8B","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"17.5","value_num":17.5,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl0.8b/live_success","tex_file":"paper/generated/explore.tex"},{"model":"sl2b","model_name":"StartLux-Decision-2B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"89.9","value_num":89.9,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl2b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl2b","model_name":"StartLux-Decision-2B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"68.4","value_num":68.4,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl2b/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl2b","model_name":"StartLux-Decision-2B","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"6.5","value_num":6.5,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl2b/live_success","tex_file":"paper/generated/explore.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"91.3","value_num":91.3,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl4b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"89.1","value_num":89.1,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl4b/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl4b","model_name":"StartLux-Decision-4B","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"60.0","value_num":60.0,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl4b/live_success","tex_file":"paper/generated/explore.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"92.2","value_num":92.2,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl9b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"91.8","value_num":91.8,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl9b/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"74.0","value_num":74.0,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl9b/live_success","tex_file":"paper/generated/explore.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"90.8","value_num":90.8,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl27b/e1_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"93.9","value_num":93.9,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl27b/corpus_acc","tex_file":"paper/generated/explore.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"63.0","value_num":63.0,"paper_display":"Appendix / text","tex_key":"explore/startlux_live/sl27b/live_success","tex_file":"paper/generated/explore.tex"},{"model":"e12","model_name":"AgentDecision-4B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.1","value_num":99.1,"paper_display":"Appendix / text","tex_key":"e12/e12/c/acc","tex_file":"paper/generated/e12_train.tex"},{"model":"e12","model_name":"AgentDecision-4B","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"100.0","value_num":100.0,"paper_display":"Appendix / text","tex_key":"e12/e12/live/success","tex_file":"paper/generated/e12_train.tex"},{"model":"e12c","model_name":"StartLux-4B + recipe","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.6","value_num":99.6,"paper_display":"Appendix / text","tex_key":"e12c/e12c/c/acc","tex_file":"paper/generated/e12_train.tex"},{"model":"bosun","model_name":"Bosun 3.1 1.7B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"36.3","value_num":36.3,"paper_display":"Appendix / text","tex_key":"e15/bosun/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clefflash","model_name":"Clef-Flash 9B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"47.7","value_num":47.7,"paper_display":"Appendix / text","tex_key":"e15/clefflash/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clm","model_name":"CLM v0.1 (8B)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"11.9","value_num":11.9,"paper_display":"Appendix / text","tex_key":"e15/clm/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"decider","model_name":"decider-4b v2.1","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"63.2","value_num":63.2,"paper_display":"Appendix / text","tex_key":"e15/decider/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"imajev","model_name":"Imajev-4B","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"76.9","value_num":76.9,"paper_display":"Appendix / text","tex_key":"e15/imajev/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"jevk5","model_name":"JevK5 v0.2 (4B)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"65.0","value_num":65.0,"paper_display":"Appendix / text","tex_key":"e15/jevk5/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"laya","model_name":"Laya (421M)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"11.0","value_num":11.0,"paper_display":"Appendix / text","tex_key":"e15/laya/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"oaidec","model_name":"OpenAI Decisions API (gpt-6-luna)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"90.9","value_num":90.9,"paper_display":"Appendix / text","tex_key":"e15/oaidec/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjev","model_name":"OpenJev (DiffusionGemma-26B-A4B)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"91.1","value_num":91.1,"paper_display":"Appendix / text","tex_key":"e15/openjev/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjevhf","model_name":"OpenJev-27B (HF, FP8)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"92.4","value_num":92.4,"paper_display":"Appendix / text","tex_key":"e15/openjevhf/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"verdict","model_name":"Verdict 1.4 (151M)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"19.0","value_num":19.0,"paper_display":"Appendix / text","tex_key":"e15/verdict/corpus_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"bosun","model_name":"Bosun 3.1 1.7B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"85.8","value_num":85.8,"paper_display":"Appendix / text","tex_key":"e15/bosun/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clef","model_name":"Clef 27B (8-bit)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"86.3","value_num":86.3,"paper_display":"Appendix / text","tex_key":"e15/clef/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clefflash","model_name":"Clef-Flash 9B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"87.5","value_num":87.5,"paper_display":"Appendix / text","tex_key":"e15/clefflash/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clm","model_name":"CLM v0.1 (8B)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"71.0","value_num":71.0,"paper_display":"Appendix / text","tex_key":"e15/clm/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"decider","model_name":"decider-4b v2.1","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"90.8","value_num":90.8,"paper_display":"Appendix / text","tex_key":"e15/decider/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"imajev","model_name":"Imajev-4B","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"87.9","value_num":87.9,"paper_display":"Appendix / text","tex_key":"e15/imajev/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"jevk5","model_name":"JevK5 v0.2 (4B)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"86.4","value_num":86.4,"paper_display":"Appendix / text","tex_key":"e15/jevk5/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"laya","model_name":"Laya (421M)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"79.8","value_num":79.8,"paper_display":"Appendix / text","tex_key":"e15/laya/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"verdict","model_name":"Verdict 1.4 (151M)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"78.3","value_num":78.3,"paper_display":"Appendix / text","tex_key":"e15/verdict/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjev","model_name":"OpenJev (DiffusionGemma-26B-A4B)","level":"L1","task_set":"isolated: BFCL v4 tool choice only","metric":"accuracy (%)","value":"91.2","value_num":91.2,"paper_display":"Appendix / text","tex_key":"e15/openjev/e1_acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"oaidec","model_name":"OpenAI Decisions API (gpt-6-luna)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"78.5","value_num":78.5,"paper_display":"Appendix / text","tex_key":"e15/oaidec/live_success","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjev","model_name":"OpenJev (DiffusionGemma-26B-A4B)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"62.5","value_num":62.5,"paper_display":"Appendix / text","tex_key":"e15/openjev/live_success","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjevhf","model_name":"OpenJev-27B (HF, FP8)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"63.0","value_num":63.0,"paper_display":"Appendix / text","tex_key":"e15/openjevhf/live_success","tex_file":"paper/generated/e15_baselines.tex"},{"model":"base","model_name":"Qwen3-4B (untrained)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"50.1","value_num":50.1,"paper_display":"Appendix / text","tex_key":"stage2_rl/base/corpus_acc","tex_file":"paper/generated/stage2.tex"},{"model":"sft_v1","model_name":"Qwen3-4B + SFT (v1)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.2","value_num":99.2,"paper_display":"Appendix / text","tex_key":"stage2_rl/sft_v1/corpus_acc","tex_file":"paper/generated/stage2.tex"},{"model":"v2","model_name":"JevAgentRL-v2 (exploratory)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.4","value_num":99.4,"paper_display":"Appendix / text","tex_key":"stage2_v2/corpus_acc","tex_file":"paper/generated/stage2.tex"},{"model":"e14a","model_name":"Qwen3.5-0.8B scoring head (agent states)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.2","value_num":99.2,"paper_display":"Appendix / text","tex_key":"e14/a/c/acc","tex_file":"paper/generated/e14_small.tex"},{"model":"e14a","model_name":"Qwen3.5-0.8B scoring head (agent states)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"100.0","value_num":100.0,"paper_display":"Appendix / text","tex_key":"e14/a/live/success","tex_file":"paper/generated/e14_small.tex"},{"model":"e14ag","model_name":"Qwen3.5-0.8B scoring head (agent states + generic mix)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.1","value_num":99.1,"paper_display":"Appendix / text","tex_key":"e14/ag/c/acc","tex_file":"paper/generated/e14_small.tex"},{"model":"e14g","model_name":"Qwen3.5-0.8B scoring head (generic mix only)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"23.9","value_num":23.9,"paper_display":"Appendix / text","tex_key":"e14/g/c/acc","tex_file":"paper/generated/e14_small.tex"},{"model":"e14g","model_name":"Qwen3.5-0.8B scoring head (generic mix only)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"0.0","value_num":0.0,"paper_display":"Appendix / text","tex_key":"e14/g/live/success","tex_file":"paper/generated/e14_small.tex"},{"model":"e16a","model_name":"AgentDecision-4B (mixed+)","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.1","value_num":99.1,"paper_display":"Appendix / text","tex_key":"e16/a/c/acc","tex_file":"paper/generated/e16_best_open.tex"},{"model":"e16a","model_name":"AgentDecision-4B (mixed+)","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"100.0","value_num":100.0,"paper_display":"Appendix / text","tex_key":"e16/a/live/success","tex_file":"paper/generated/e16_best_open.tex"},{"model":"e16b","model_name":"StartLux-9B + recipe","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.3","value_num":99.3,"paper_display":"Appendix / text","tex_key":"e16/b/c/acc","tex_file":"paper/generated/e16_best_open.tex"},{"model":"e16b","model_name":"StartLux-9B + recipe","level":"L3","task_set":"DecisionGym closed-loop (first 200 Stage-2 tasks)","metric":"task success (%)","value":"100.0","value_num":100.0,"paper_display":"Appendix / text","tex_key":"e16/b/live/success","tex_file":"paper/generated/e16_best_open.tex"},{"model":"e17gen","model_name":"Qwen3.5-0.8B one-token read-out","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.0","value_num":99.0,"paper_display":"Appendix / text","tex_key":"e17/gen/c/acc","tex_file":"paper/generated/e17_e18.tex"},{"model":"e18f","model_name":"Qwen3.5-0.8B with a trained defer option","level":"L2","task_set":"common states (200 Stage-2 tasks)","metric":"step accuracy (%)","value":"99.1","value_num":99.1,"paper_display":"Appendix / text","tex_key":"e17/e18f/c/acc","tex_file":"paper/generated/e17_e18.tex"},{"model":"e28_vendor","model_name":"AgentDecision-0.8B (vendor serving engine)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"80.8","value_num":80.8,"paper_display":"Appendix / text","tex_key":"e28/l1/acc_vendor","tex_file":"paper/generated/e28_sl08b.tex"},{"model":"e16a","model_name":"AgentDecision-4B (mixed+)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"84.3","value_num":84.3,"paper_display":"Appendix / text","tex_key":"e9/e16a/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"e16b","model_name":"StartLux-9B + recipe","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"84.8","value_num":84.8,"paper_display":"Appendix / text","tex_key":"e9/e16b/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"e14a","model_name":"Qwen3.5-0.8B scoring head (agent states)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"67.8","value_num":67.8,"paper_display":"Appendix / text","tex_key":"e9/e14a/acc","tex_file":"paper/generated/e14_small.tex"},{"model":"e14ag","model_name":"Qwen3.5-0.8B scoring head (agent states + generic mix)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"68.8","value_num":68.8,"paper_display":"Appendix / text","tex_key":"e9/e14ag/acc","tex_file":"paper/generated/e14_small.tex"},{"model":"e14g","model_name":"Qwen3.5-0.8B scoring head (generic mix only)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"52.3","value_num":52.3,"paper_display":"Appendix / text","tex_key":"e9/e14g/acc","tex_file":"paper/generated/e14_small.tex"},{"model":"e27s","model_name":"Qwen3.5-0.8B + surface augmentation","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"67.7","value_num":67.7,"paper_display":"Appendix / text","tex_key":"e9/e27s/acc","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"e27sp","model_name":"Qwen3.5-0.8B + surface augmentation + public data","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"77.2","value_num":77.2,"paper_display":"Appendix / text","tex_key":"e9/e27sp/acc","tex_file":"paper/generated/e27_small_shift.tex"},{"model":"sl2b","model_name":"StartLux-Decision-2B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"75.7","value_num":75.7,"paper_display":"Appendix / text","tex_key":"e9/sl2b/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"sl9b","model_name":"StartLux-Decision-9B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"83.4","value_num":83.4,"paper_display":"Appendix / text","tex_key":"e9/sl9b/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"sl27b","model_name":"StartLux-Decision-27B (Q8_0)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"87.7","value_num":87.7,"paper_display":"Appendix / text","tex_key":"e9/sl27b/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"oaidec","model_name":"OpenAI Decisions API (gpt-6-luna)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"79.2","value_num":79.2,"paper_display":"Appendix / text","tex_key":"e9/oaidec/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjev","model_name":"OpenJev (DiffusionGemma-26B-A4B)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"78.8","value_num":78.8,"paper_display":"Appendix / text","tex_key":"e9/openjev/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"openjevhf","model_name":"OpenJev-27B (HF, FP8)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"82.7","value_num":82.7,"paper_display":"Appendix / text","tex_key":"e9/openjevhf/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"jevk5","model_name":"JevK5 v0.2 (4B)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"70.3","value_num":70.3,"paper_display":"Appendix / text","tex_key":"e9/jevk5/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clm","model_name":"CLM v0.1 (8B)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"46.2","value_num":46.2,"paper_display":"Appendix / text","tex_key":"e9/clm/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"laya","model_name":"Laya (421M)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"49.4","value_num":49.4,"paper_display":"Appendix / text","tex_key":"e9/laya/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"verdict","model_name":"Verdict 1.4 (151M)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"47.2","value_num":47.2,"paper_display":"Appendix / text","tex_key":"e9/verdict/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"decider","model_name":"decider-4b v2.1","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"79.3","value_num":79.3,"paper_display":"Appendix / text","tex_key":"e9/decider/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"imajev","model_name":"Imajev-4B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"78.9","value_num":78.9,"paper_display":"Appendix / text","tex_key":"e9/imajev/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"bosun","model_name":"Bosun 3.1 1.7B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"63.3","value_num":63.3,"paper_display":"Appendix / text","tex_key":"e9/bosun/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clef","model_name":"Clef 27B (8-bit)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"84.0","value_num":84.0,"paper_display":"Appendix / text","tex_key":"e9/clef/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"clefflash","model_name":"Clef-Flash 9B","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"81.4","value_num":81.4,"paper_display":"Appendix / text","tex_key":"e9/clefflash/acc","tex_file":"paper/generated/e15_baselines.tex"},{"model":"flash","model_name":"Gemini 3.5 Flash (multiple choice)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"88.1","value_num":88.1,"paper_display":"Appendix / text","tex_key":"e9/flash/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"lite","model_name":"Gemini 3.5 Flash-Lite (multiple choice, distribution)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"83.3","value_num":83.3,"paper_display":"Appendix / text","tex_key":"e9/lite/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"litelabel","model_name":"Gemini 3.5 Flash-Lite (label + confidence)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"84.1","value_num":84.1,"paper_display":"Appendix / text","tex_key":"e9/litelabel/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"gemma","model_name":"Gemma 4 26B (multiple choice)","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"80.2","value_num":80.2,"paper_display":"Appendix / text","tex_key":"e9/gemma/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"flashfc","model_name":"Gemini 3.5 Flash, own function call","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"78.7","value_num":78.7,"paper_display":"Appendix / text","tex_key":"e9/flashfc/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"embed","model_name":"Gemini embedding ranker","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"51.7","value_num":51.7,"paper_display":"Appendix / text","tex_key":"e9/embed/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"bm25","model_name":"BM25 keyword ranker","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"39.4","value_num":39.4,"paper_display":"Appendix / text","tex_key":"e9/bm25/acc","tex_file":"paper/generated/e9_frontier.tex"},{"model":"tfidf","model_name":"TF-IDF ranker","level":"L1","task_set":"isolated: 3,991 BFCL v4 + When2Call questions (pooled)","metric":"accuracy (%)","value":"39.9","value_num":39.9,"paper_display":"Appendix / text","tex_key":"e9/tfidf/acc","tex_file":"paper/generated/e9_frontier.tex"}]}