Executing an idea often changes how promising it looks, so the next step in autonomous research should rest on results, not on proposals. After each round of experiments, the system must decide whether to refine the best solution or start a new direction, which solution to build from, and what goal to set. We call this the between-round decision. Existing systems wrap it in machinery such as rules, scores, search trees, and schedules; where an LLM makes it, what the LLM needs to read is left open. We ask whether an LLM can make this decision from accumulated evidence, without a search structure around it, and how much of that evidence it needs to read. We present BRED (Between-Round Evidence-based Decision), an LLM-based controller that maintains the best verified solution so far (the Champion) and a log of past rounds, and makes one decision per round. On three widely used research-agent benchmark tasks, BRED improves on the task baselines by 4.80% in MAE on ETTh1 (to 0.4172) and by 3.44 and 3.43 points of accuracy on SST-2 (to 95.07%) and CIFAR-100 (to 84.63%), ahead of the published results. When the controller does not read the accumulated evidence, the same loop gains less: with no history it reaches 0.4299 on ETTh1 and 83.71% on CIFAR-100, whereas the full controller's MAE is 2.95% lower and its accuracy 0.92 points higher. On a new task, speech-based depression detection under a leakage-free protocol, it reaches a macro-F1 of 0.6164 on seeds held out from selection, 5.5 points above its baseline and 4.0 above the strongest published system. Code is available at https://github.com/learninginvision/BRED.
Figures & tables
Figure 1: Overview of the loop, shown on round 7 of the AutoClsSST run. (1) The research controller reads the accumulated evidence, namely the Champion and the log, together with the task and a survey of its published lines, and commits one decision: one of four actions (start a new line from the baseline, extend the Champion, refine it, or stop), an instruction, a rationale, and hard constraints. (2) For a new line or an extension, candidate generation plans queries, retrieves papers, and drafts ideas; refinement skips retrieval. (3) Each of the N ideas is implemented and run in its own copy of the starting workspace, and failures are repaired. (4) Record promotes the round’s best candidate only if it beats the Champion, and appends the round and its lesson to the log. No weights are updated.
InternAgent
Δ vs.
Task (dataset)
Metric
Baseline
Dolphin
published
re-run †
Ours
best prior
AutoTSF (ETTh1)
MAE ↓
0.4382
0.4627
0.4331
0.4245
0.4172
−1.72%
AutoClsSST (SST-2)
Acc. ↑
91.63
92.50
93.50
92.78
95.07
+1.57 pp
AutoCls2D (CIFAR-100)
Acc. ↑
81.20
82.00
83.30
83.44
84.63
+1.19 pp
AutoDepDAIC (DAIC-WOZ)
Macro-F1 ↑
56.10
—
—
—
61.64
+4.04 pp
Table 1: Main results. Dolphin and published InternAgent numbers are taken verbatim from Tables 1–2 of InternAgent Team et al. (2025) , reported there and here as maximum performance across runs; re-run † is InternAgent’s released code run with our models (Appendix D ). The first three tasks share that benchmark’s datasets, baselines, and metrics. AutoDepDAIC lies outside it and is compared against HAREN-CTC ( Li et al., 2026 ) ( 57.6 ), the strongest published system under the same leakage-free protocol; our entry there is measured on five seeds that took no part in selection. AutoTSF, AutoClsSST, and AutoDepDAIC baselines are our runs’ verified floors. The last column compares our result with the strongest prior entry in the row (relative reduction for MAE, gap in points otherwise).
Figure 2: The AutoCls2D run as a lineage of code. The top row gives each round’s action and the controller’s decision. Each candidate hangs from the code it starts from: new lines (R1, R5–R7) from the baseline, in the upper band, and extensions and refinements from the Champion, below the green Champion lineage, whose nodes mark each new Champion and its gain. Pills give each candidate’s top-1 accuracy (%); dashed pills mark candidates without a valid result. R1–R3 raise the Champion to 83.91% . After R4 fails to extend it, the controller leaves the stalled lineage and opens three new lines from the baseline (R5–R7); none beats the Champion, which is kept unchanged. R8–R10 fall in the last 30% of rounds, reserved for refining or stopping, and two refinements lift the Champion to 84.63% , 3.43 points above the published WideResNet baseline ( 81.20% ). The run ends when its ten-round budget is used up. Round labels are read back from the logged ideas and each candidate’s code.
Decide reads
AutoTSF
AutoCls2D
MAE ↓
Acc. ↑
No history
0.4299
83.71
1-round history
0.4244
84.01
Full history (ours)
0.4172
84.63
Table 2: Best verified result of each context setting. Baselines: AutoTSF 0.4382 ; AutoCls2D 81.20% ( InternAgent Team et al., 2025 ) .
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Left out
Intended purpose
Replacement
Pattern–action table
Map history to next action
LLM decision over the log
Research forest, revive states
Return to old directions
The log plus a new line from the baseline
Composite mechanism stack
Track accumulated solution
Champion signature plus code
Multi-axis round plan
Express fine-grained intent
Goal-level instruction plus hard constraints
Appendix
Table 3: Structures left out of the controller and what replaces them.
Phase
Implementation role
Decide
New controller LLM node over the Champion and the log
Generate
Existing search planner, retriever, and idea generator
Run
Existing code runner and scheduler
Record
New log writer and Champion updater
Base copying
Existing parent-workspace copy mechanism, if available
Appendix
Table 4: Mapping the loop’s phases to common AI-scientist execution components.
R
Decide (mode/base)
#cand.
#fail
Best MAE ↓
1
explore /baseline
3
0
0.4608
2
explore /baseline
3
0
0.4947
3
explore /baseline
3
1
0.5443
4
explore /baseline
3
0
0.4226
5
explore / iter_4_idea_3
3
0
0.4193
6
explore / iter_5_idea_1
3
1
0.4184
Appendix
Table 5: Main run on AutoTSF ( 20260731_080205 ); Decide reads the Champion and the full log. The controller sets the first Champion at R4 and improves it in each of R5–R8, reaching MAE 0.4172 (R8 improves on R7 only in the fifth decimal, 0.41724→0.41721 ), and stops at R9 with budget left, after eight of ten budgeted rounds. Baseline MAE: 0.4382. Three candidates per round. Bold marks rounds that improved the Champion.
R
Decide (mode/base)
#cand.
#fail
Best Acc ↑
1
explore /baseline
3
1
94.61
2
exploit / iter_1_idea_1
3
0
94.15
3
explore /baseline
3
0
94.95
4
exploit / iter_3_idea_3
3
0
94.04
5
explore /baseline
3
0
93.81
6
explore / iter_3_idea_3
3
0
94.27
Appendix
Table 6: Main run on AutoClsSST ( 20260803_001220 ), ten-round budget, three candidates per round; Decide reads the Champion and the full log. Accuracy (%) on SST-2. The controller stops at R10, after nine executed rounds. Baseline accuracy: 91.63% . Bold marks rounds that improved the Champion.
R
Decide (mode/base)
#cand.
#fail
Best Acc ↑
1
explore /baseline
4
0
82.20
2
explore / iter_1_idea_5
4
1
83.57
3
explore / iter_2_idea_2
4
1
83.91
4
explore / iter_3_idea_4
4
0
83.39
5
explore /baseline
4
0
81.62
6
explore /baseline
4
1
80.51
Appendix
Table 7: Main run on AutoCls2D ( 20260923_131120 ), ten rounds, four candidates per round; Decide reads the Champion and the full log. Top-1 accuracy on CIFAR-100. The in-run baseline is a three-epoch smoke configuration ( 49.23% ) that serves only as the floor before the first Champion. Bold marks rounds that improved the Champion.
R
Decide (mode/base)
#cand.
#fail
Best Macro-F1 ↑
1
explore /baseline
5
0
56.58
2
explore / iter_1_idea_6
5
0
59.94
3
explore / iter_2_idea_2
5
1
55.65
4
explore /baseline
5
5
—
5
explore / iter_2_idea_2
5
5
—
6
explore /baseline
5
0
56.52
Appendix
Table 8: Main run on AutoDepDAIC ( 20260814_164352 ), ten-round budget, five candidates per round (four in R9); Decide reads the Champion and the full log. Nine rounds executed, after which the controller stopped on convergence. R8 improves on R7 only in the third decimal ( 60.188→60.191 ). Values are macro-F1 ×100 under pooled 188-subject subject-stratified 5-fold CV over five seeds; baseline floor: 56.10 . Bold marks rounds that improved the Champion.
Autonomous research agents use language models and tools to search literature, design and run experiments, analyze results, and support scientific discovery.