acceptodds
Under review as a conference paper at ICLR 2027

DriveLab: Autonomous Research for Self-Driving Motion-Prediction Models

Abstract

Developing high-performing motion-prediction models for autonomous driving is an iterative and resource-intensive process: researchers must diagnose model failures, modify complex training and inference pipelines, debug implementations, launch expensive experiments, and prioritize next steps. We investigate whether this development loop can be automated using LLM-based autonomous research agents. To this end, we build an agent-operable environment around the Waymo Open Motion Dataset (WOMD) and study the autonomous improvement of two structurally distinct motion-prediction systems: Wayformer and MotionLM. We introduce DriveLab, an auto-research framework that organizes long-horizon exploration through three agentic roles atop a shared human-built infrastructure: Researcher Agents that propose high-quality, diverse hypotheses; sandboxed, skill-equipped Coders that implement changes and diagnose runtime failures; and an Evolution Manager that maintains persistent experimental memory and manages context. Under a fixed 24-hour compute budget on identical hardware, DriveLab achieves the highest Soft Mean Average Precision (Soft mAP) among the evaluated harnesses for all three base LLMs: averaged across base LLMs, it improves Wayformer by 9.5% and MotionLM by 13.0% over their starting checkpoints (up to 18.6% without a fixed time budget), and discovers non-trivial algorithmic improvements, such as new mode-aggregation mechanisms. We also identify the balance between research-idea novelty and code-execution stability as a key design consideration for auto-research and show that our method handles both well. Our results provide strong evidence that autonomous LLM agents can take on substantial parts of the development of complex driving-behavior models.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Auto-Research, Self-Driving, Motion Prediction, LLM Agents

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UabDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces DriveLab, an autonomous research framework designed for long-horizon, computationally expensive machine learning research, specifically instantiated on autonomous vehicle (AV) motion prediction. It organizes exploration via researcher, coder, and evolution-manager agents around WOMD and reports consistent Soft mAP gains for Wayformer and MotionLM under a fixed 24-hour budget, while identifying the balance between idea novelty and code-execution stability as a key design consideration.

Critical strengths

The paper makes a methodological contribution by shifting focus from purely final benchmark numbers to evaluating the research process itself. It introduces metrics such as Unproductive Attempt Rate (UAR), Active Utilization Rate (AUR), and LLM tokens per valid experiment, as well as an independent 2D rubric (Scope and Novelty). Standard baselines exhibited a clear coupling where higher novelty triggered catastrophic code failures (UAR rising from 22.0% to 35.5%). DriveLab effectively breaks this coupling, achieving both the highest Novelty (1.21-1.35) and the lowest UAR (3.8%-10.7%).

Critical weaknesses

Already stated in the Questions.

Questions for authors
  1. From the ablation study in Table 2, it appears that the Exploration term contributes more prominently than the other two components. Could the authors clarify what β = 0 represents in this context and how it relates to βtβ_t in Eq.1? Please add a theoretical analysis explaining why the Exploration term yields such a large improvement in UAR.
  2. If DriveLab were initialized without expert-authored seed skills (specifically without WOMD failure-diagnosis and interactive-kernel-smoke-test), does the automatic skill distillation mechanism alone manage to recover execution stability, or does UAR degrade back to baseline levels?
2 comments

We thank the reviewer for recognizing our process evaluation, especially the distinction between ambitious ideas and reliable execution.

Q1: What does β = 0 mean, and why does UAR change?

In Eq. 1, the Aggregator ranks each parent–hypothesis pair using normalized parent Soft mAP, the Reviewers' expected gain/risk score, and an exploration bonus β_t·u(h, 𝒜). Here, u measures dissimilarity from previously attempted edits using deterministic token-edit similarity (Eq. 3). Full DRIVELAB uses β_t = 1.5 × 0.92^t, encouraging diverse hypotheses early and gradually shifting toward exploitation. β = 0 removes only this bonus; Reviewers, Coders, skills, and execution safeguards remain unchanged.

Our presentation of Table 2 should have made two results clearer. First, removing exploration lowers UAR from 10.5% to 5.7%, rather than increasing it. The largest UAR increase occurs when skill distillation is removed (10.5% → 27.6%). Second, removing exploration produces the smallest observed drop in final Soft mAP (0.0256), compared with removing skill distillation (0.0282) or the Researcher Team (0.0435). Its most pronounced effect is on Novelty (1.48 → 0.75). These differences are descriptive, not estimates of relative causal contributions.

The explanation follows our design rationale in Sec. 3: expected progress from testing hypothesis h depends on both its execution success probability and its potential gain if successful:

E[Δf | h, c] ≈ P(valid | h, c) · E[Δf | h, c, valid].

The exploration bonus changes which hypotheses are selected, not how they are executed. Without exploration, selection relies on parent performance and the Reviewers' gain/risk assessments, which already penalize implementation risk. This tends to favor familiar, incremental changes that are easier to execute but offer limited improvement. With exploration, structurally different hypotheses receive additional weight. These may be harder to implement but offer opportunities for larger gains.

Table 2 is consistent with this trade-off. Without exploration, 94.3% of attempts produce valid evaluations, compared with 89.5% for full DRIVELAB. Yet the full system achieves approximately 1.5× the improvement over the initial checkpoint (+0.0727 vs. +0.0471 Soft mAP). Skill distillation helps manage the additional implementation risk: retaining exploration but removing distillation raises UAR to 27.6%.

The exploration bonus also discourages repeated edits, which could reduce UAR because duplicate attempts count as failures (App. A). Nevertheless, the observed net UAR is higher with exploration, consistent with the greater difficulty of the selected hypotheses.

This is a mechanistic explanation rather than a theoretical guarantee, and Table 2 contains only one open-ended episode per variant. Sec. 4.3 already notes that β = 0 yields the lowest UAR; we will make the comparison and its interpretation clearer in the revision.

Q2: Can automatic distillation recover without expert seed skills?

We have not run a no-seed-skills experiment, so we cannot determine whether automatic skill distillation alone would recover execution stability. This was a deliberate choice in our experimental design. We provide basic infrastructure workflows to all agents because our goal is to study their ability to conduct autonomous model research, rather than requiring them to first discover how to operate a particular computing or evaluation environment. Specifically, the two expert seed skills serve different purposes:

  • interactive-kernel-smoke-test (Listing 8) gives Coders a procedure for hot-reloading JAX modules and checking shapes and gradients before expensive training runs.
  • womd-failure-diagnosis (Listing 10) gives Researchers a procedure for inspecting prediction errors through scenario-level diagnostics and BEV plots. It also contains an expert-defined failure taxonomy, which helped identify the problems in Case 2, but did not prescribe the algorithmic solution.

We view these primarily as environment-specific procedures rather than research solutions. For example, requiring agents to rediscover how to operate an interactive JAX kernel or which diagnostic tools are available would introduce an additional infrastructure-discovery problem. We chose to standardize these procedures so that differences in research performance would not be confounded by differences in how effectively agents learn to operate the environment.

We acknowledge that the distinction between infrastructure and research is not absolute. Better diagnostic tools can influence which hypotheses agents pursue, and improved execution procedures can make more ambitious experiments feasible. Our matched evaluation controls access to these tools and skills across all baselines, but does not measure their absolute contribution. Allowing agents to develop and improve the research infrastructure alongside the models is an interesting direction for future work.

Thank you for your rebuttal. I have read your responses carefully and appreciate the clarifications. After reconsidering the paper in light of the rebuttal, I will maintain my original score.

Reviewer_TVA8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DriveLab, an LLM-based framework for autonomous research on motion prediction models for self-driving. It assigns hypothesis generation, experiment execution, and research state management to specialized agents. The agents use persistent memory, reusable skills, and domain tools to modify model training and trajectory post-processing. The authors evaluate the framework on Wayformer and MotionLM using the Waymo Open Motion Dataset under fixed and open-ended budgets.

Critical strengths

The paper provides a controlled evaluation of the framework’s practical value. The main comparisons use matched time and hardware budgets, shared starting models and domain tools, and five independent runs per setting. A protected evaluation split helps assess whether the gains extend beyond the data used during search. The results show consistent improvements over the single-agent baseline across two prediction models and three base LLMs.

Critical weaknesses
  • The proposed framework mainly combines existing components, including role specialization, persistent memory, parallel execution, and skill reuse. Its adaptation to motion prediction mainly relies on domain tools and expert-written initial skills. It does not explain which research challenge in motion prediction calls for a new design of the autonomous research process. This limits the novelty of the paper.

  • The reported model changes by the auto-research framework mainly involve engineering heuristics, such as adding continuous corrections to discrete action outputs and modifying trajectory post-processing. These examples explore combinations of known techniques and implementation choices rather than resolve a clearly defined scientific question. The resulting gains demonstrate useful engineering optimization. However, they do not provide convincing evidence that the proposed auto-research system can discover previously unexplored methods or new knowledge in the domain.

Questions for authors
  • How much of the observed improvement depends on the domain-specific diagnostic tools and expert-written initial skills?

  • Could the authors provide more examples of model changes discovered by DriveLab that show its ability to analyze model behavior, reason with domain knowledge, and develop new methods or insights?

  • Is the main contribution of the paper a better general-purpose auto-research framework, including its design for role specialization and collaboration, or infrastructure adapted to trajectory prediction? If it is the former, could the authors compare the framework with state-of-the-art auto-research methods on other tasks? If it is the latter, which challenges specific to trajectory prediction research does the infrastructure address, and how?

1 comment

We thank the reviewer for the review.

W1/Q3: Contribution and Motion-Prediction Challenges

Our contribution is investigating autonomous research in driving motion prediction and developing a harness for the challenges this domain exposes. We do not claim cross-domain generality or novelty in individual agent components. As Reviewer 1 recognized, the environment and controlled forecasting study are important contributions.

Three challenges motivated our design:

Cross-stage coupling. Changes to trajectory representations often require coordinated modifications to preprocessing, losses, decoding, and post-processing. Small inconsistencies can invalidate expensive experiments (App. E, Listing 3). We address this through dedicated Coders with sandboxed execution, smoke tests, and automatically distilled implementation skills that preserve cross-stage constraints (e.g., Listing 9).

Difficult verification. Soft mAP aggregates 63 class/horizon/trajectory cells, potentially hiding serious errors. For example, missing agent-type propagation silently applied vehicle parameters to pedestrians and cyclists for 14 experiments (App. E, Listings 4–5). We provide scenario-level diagnostic tools and maintain complete experiment records, while the Evolution Manager supplies Researchers with scientific findings and Coders with relevant debugging history. This helps agents investigate failures without repeatedly rediscovering earlier evidence.

Expensive feedback. Full retraining takes approximately 3–10 hours, making failed structural changes costly. Our Single-Agent pilots frequently retreated to incremental edits (75.0–78.6% Scope 0/1; Table 8). DRIVELAB uses explicit exploration incentives to sustain diverse hypotheses, parallel Coders to evaluate alternatives, and skill distillation to reduce repeated debugging.

W2/Q2: What Did the Agents Actually Develop?

We agree with the reviewer's characterization: the individual techniques are established, and the results are engineering improvements, not new scientific knowledge. Both cases are rated Novelty 2 ("Adapted"), not 3 ("Original"), in Table 4. The question the paper asks is whether agents can carry out the development loop that produces such improvements: find a modeling problem, propose a fix, and refine it against experimental evidence.

For example, in Case 2: expert-assisted diagnosis, agent-designed solution (App. G.2), The expert-provided diagnostic workflow and failure taxonomy helped identify two limitations in MotionLM's post-processing: suppression at 5 s removed trajectories that only diverge by 8 s, and K-means refinement pulled distinct modes toward cluster centers. The skill did not provide a solution. The Researcher proposed velocity-scaled distances across 3/5/8 s, MMR selection for mode diversity, and entropic OT with speed-dependent blending. Two competing designs were implemented and rejected: a strict all-horizon conjunction almost never treated rollouts as duplicates, which starved coverage (0.4170 dev), and unbalanced OT collapsed secondary modes (0.4242 dev). The promoted design reached 0.4538 dev and 0.4473 protected.

We believe this is valuable even though it is not new knowledge. Diagnosis, implementation, and empirical selection are what consume expert time in developing these models, and here that loop ran autonomously within each episode. In 24 hours DRIVELAB improved the seed checkpoints by 9.5% (Wayformer) and 13.0% (MotionLM) on average across base LLMs (Table 1), achieveing single model scores of0.4523 on the official WOMD test server, above ModeSeq (0.4487) and EDA (0.4510).

Q1: Dependence on Tools and Expert Seed Skills

We deliberately provide expert seed skills as part of the research infrastructure. Our goal is to evaluate autonomous research capability, not whether agents can independently discover how to operate a specific computing environment. For example interactive-kernel-smoke-test (Listing 8) explains how to run and validate JAX code efficiently, while womd-failure-diagnosis (Listing 10) guides error analysis using diagnostic tools and BEV visualizations. The latter also includes an expert-defined failure taxonomy that informed Case 2's diagnosis, but neither skill prescribes algorithmic solutions.

These workflows vary across research environments. Requiring agents to rediscover them would introduce a separate infrastructure-discovery task, potentially confounding our evaluation of research capability. Importantly, all evaluated harnesses received identical tools and seed skills (Table 5; App. D). We acknowledge that expert guidance may still indirectly affect execution reliability and hypothesis generation. Our matched comparisons control access to this guidance, but do not measure performance without it. We will clarify this distinction and its limitations in Secs. 3–5.

Reviewer_qsWeOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether LLM research agents can sustain the development loop of autonomous-driving motion-prediction models: diagnosis, implementation, training, evaluation and selection of the next experiment. It argues that long runs suffer from context growth, repeated procedural work and increasingly narrow search. DriveLab provides a WOMD environment with separate development and protected evaluation splits. Proposer and reviewer agents rank hypotheses with an exploration bonus, parallel coders implement changes and distil reusable skills, and a manager curates memory and context. The fixed-budget study uses 24 hours and five episodes per setting on Wayformer and MotionLM. Four harnesses are compared under Gemini 3.6 Flash, with DriveLab versus Single-Agent under two further LLMs. The paper reports average improvements of 9.5% and 13.0% over the two seed models, case-study model changes, and a WOMD test-server score of 0.4523.

Critical strengths

A useful environment for long-horizon automated research on motion forecasting. It goes beyond selecting model configurations by allowing agents to inspect and modify the forecasting pipeline over long episodes. The environment separates development feedback from protected evaluation and documents tools, permissions and concrete code changes. The environment and controlled forecasting study are a clear contribution, even though the agent components have precedents.

The fixed-budget comparison is careful and the lead is consistent. Each setting runs five episodes, and the final development-selected checkpoint is evaluated on a protected split whose scores are not returned to the agents. All harnesses receive the same tools, expert seed skills and promotion rule. DriveLab's worst episode exceeds the tested alternatives' best in five of six model/LLM settings. Gemini 3.6 Flash tests all four harnesses; the other two LLMs test only DriveLab and Single-Agent. The consistent lead within these settings is convincing.

Critical weaknesses

What produces DriveLab's lead is not established. Table 2 runs one open-ended episode per variant, with 61.6 h for the full system and 31.8–55.4 h for its ablations. Component removal therefore changes both the system and the amount of search completed. This does not estimate the component's contribution to the fixed 24-hour result. The memory/context manager is also not separately ablated. DeepScientist (ICLR 2026) already combines scored hypothesis selection, exploration, persistent findings and sandboxed coding; AIRA (NeurIPS 2025) is another relevant published comparator. Please explain the functional differences and compare the important components under a common budget and repeated episodes. A matched published harness would strengthen the broader claim further.

The discovery claim does not separate the agents' contribution from expert guidance. Case 2 applied to seed MotionLM reaches 0.4473 Soft mAP through post-processing alone, above the best alternative MotionLM episode at 0.4332. The authors label it Novelty 2, “Adapted,” and Appendix G.2 traces the diagnosis to an expert-written seed skill whose taxonomy names the defects. All harnesses receive the same skill, and DriveLab successfully turns that shared guidance into a working improvement. However, this case does not establish what the agents identified independently, as implied by the abstract's “new mode-aggregation mechanisms.”

Questions for authors

How do hypothesis selection, memory/context management and skill induction differ from DeepScientist (ICLR 2026), AIRA (NeurIPS 2025) and Agent Workflow Memory (ICML 2025)? What will be released to reproduce these distinctions?

Can the component ablations use the Table 1 protocol of 24 hours and multiple episodes, including a variant without the manager's context separation? Please keep tools and capabilities matched.

How does a published harness such as AIRA perform in the Flash setting under the same tools, expert seed skills and compute budget?

1 comment

We thank the reviewer for the thoughtful feedback.

W1/Q2: Ablation Analysis

We agree that Table 2 does not estimate each component's contribution to the fixed 24-hour result. We ran the ablations open-ended, with one episode per variant because of compute cost (Sec. 5), to see whether each variant keeps finding useful changes or stalls. All variants start from the same checkpoint and stop after 10 consecutive valid non-improving evaluations or 12 hours without improvement, so their runtimes (31.8–55.4 h versus 61.6 h for full DRIVELAB) reflect when each search plateaued, not different preset budgets. Removing skill distillation or exploration still allowed 49.8 h and 55.4 h of research, but both reached lower protected Soft mAP (0.4350/0.4376 versus 0.4632).

The ablations also reveal different failure modes. Removing the Researcher Team or exploration bonus lowers Novelty from 1.48 to 0.58/0.75; β = 0 also reduces UAR to 5.7%, consistent with safer, narrower search. Removing automatic skill distillation while retaining seed skills leaves Novelty high (1.25) but raises UAR from 10.5% to 27.6% and tokens per valid experiment from 0.52M to 0.98M. These single-episode results are diagnostic, not causal estimates. We agree that repeated 24-hour ablations, including an isolated test of the Evolution Manager's context separation, would be a valuable addition; we aim to include them in the camera-ready and will otherwise state their absence as a limitation.

Q1/Q3: Functional Differences from Prior Work

We will expand Sec. 2 to compare DRIVELAB with DeepScientist, AIRA, and AWM.

Hypothesis selection. Both DeepScientist and DRIVELAB combine LLM assessments with exploration scoring. DeepScientist estimates exploration value using an LLM surrogate and acquisition rule; DRIVELAB uses deterministic token-edit dissimilarity against archived attempts, including failures, with an annealed weight and explicit parent performance (Eqs. 1, 3). AIRA studies Greedy, MCTS, and Evolutionary search. Our β = 0 ablation supports exploration within DRIVELAB, not superiority over these approaches.

Context management. DeepScientist retrieves Findings Memory; AIRA maintains operator-specific histories. DRIVELAB separates a full archive from research memory consolidated every four experiments, providing Researchers with scientific findings and Coders with relevant patches and failure reports under bounded context (Sec. 3.3). This addresses the context failures documented in App. E.

Matched published harnesses (Q3). Adapting AIRA or DeepScientist to our job system would require substantial changes to their execution and evaluation interfaces, which are infeasible within the rebuttal period. Instead, our baselines compare fundamental scaffold designs under matched tools, skills, and compute. They are not reproductions of these systems, and we do not claim superiority over untested harnesses. We will clarify this in Sec. 5.

W2: Discovery Versus Expert Guidance

We agree that post-processing accounts for a major part of the MotionLM improvement. Case 2 alone reaches 0.4473 protected Soft mAP, above the best alternative MotionLM episode (0.4332). However, the gains are not limited to post-processing. With post-processing fixed (Table 3), DRIVELAB outperforms Single-Agent in training-only optimization on both MotionLM (+0.0327 vs. +0.0206) and Wayformer (+0.0209 vs. +0.0175). The post-processing-only gains are larger (+0.0572 vs. +0.0180 and +0.0403 vs. +0.0151, respectively). We acknowledge that these comparisons do not isolate the influence of expert guidance.

We also want to clarify what the agents contributed. In Case 2, the expert-written skill helped the agent identify two problems: suppression at 5 seconds missed trajectories that diverged later, and K-means averaging distorted distinct modes. The skill described these problems, not how to fix them. The agent developed a new aggregation procedure combining multi-horizon distances, MMR, and entropic OT. It tested and rejected two alternatives (0.4170/0.4242 dev) before reaching 0.4538 dev and 0.4473 protected (App. G.2). In Case 1, the agent identified the problem itself: token loss kept improving while minADE plateaued, suggesting an action-quantization bottleneck. It then proposed a continuous residual head, improving protected Soft mAP from 0.3905 to 0.4284 (App. G.1). We will make this distinction clearer in Sec. 4.4 and change "new mode-aggregation mechanisms" to "domain-adapted mode-aggregation mechanisms."

Finally, we are unable to release the code as it depends on proprietary infrastructure. The paper documents the scoring rule (Eqs. 1, 3), the episode and Coder procedures (Algorithms 1–2), harness parameters (Table 6), permissions and tools (Table 5; Listing 2), the prompt (Listing 1), and representative skills (Listings 8–10). If the reviewer finds an implementation detail needed for replication to be missing, we will gladly add it.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their constructive feedback. Across the reviews, we see three central questions: what DRIVELAB contributes beyond existing auto-research frameworks, how much its discoveries depend on expert guidance, and which components explain its performance gains. We address these shared concerns here, with further details in the individual responses.

1. Contribution and Scope

We agree that role specialization, persistent memory, and reusable skills are not individually new. Our contribution is an integrated auto-research harness for complex ML pipelines and, to our knowledge, the first controlled long-horizon study of autonomous research on driving motion prediction. We do not claim cross-domain generality.

Several challenges in motion prediction directly motivated our design:

  • Cross-stage coupling. Changes to trajectory representations often require coordinated changes to losses, decoding, and post-processing. We separate hypothesis design from implementation through Researcher and Coder teams. Coders use sandboxed execution, smoke tests, and automatically distilled skills to preserve implementation constraints across stages (e.g., quantization requirements in Listing 9).

  • Difficult verification. Soft mAP aggregates 63 class/horizon/trajectory cells, potentially concealing failures. Appendix E documents a missing agent-type argument that silently applied vehicle parameters to pedestrians and cyclists for 14 experiments. We therefore provide scenario-level diagnostic tools, maintain complete experimental records, and separate scientific findings from debugging histories so agents can revisit earlier evidence without overwhelming their context.

  • Expensive feedback. Training takes approximately 3–10 hours, making failed experiments costly. Single-Agent pilots often retreated to incremental edits after failures (App. E). DRIVELAB addresses this with an explicit exploration bonus that discourages repeating previous attempts, parallel Coders to evaluate competing hypotheses, and skill distillation to reduce repeated debugging.

2. Dependence on Expert Seed Skills and Component Attribution

While We have not evaluated DRIVELAB without expert seed skills, it is important to distinguish that those skills are designed as Procedural infra-level skills, such as interactive-kernel-smoke-test, help agents operate the execution environment. Diagnostic skills, such as womd-failure-diagnosis, additionally provide environment specific workflows to find recognizable failure modes. Neither prescribes the algorithmic solutions developed in our case studies, and those would vary case by case depending on the specific computing infra around the evaluations, and removing them would have require a high level of autonomy, which requires agent to find those workflow on different environment, and coufound our result.

We also acknowledge two limitations of component attribution: Table 2 reports single open-ended ablations rather than repeated 24-hour experiments, and the Evolution Manager's context separation has not been independently ablated. Likewise, we have not directly evaluated published harnesses such as DeepScientist or AIRA under our protocol. Our matched baselines establish improvements over the evaluated scaffold designs, not superiority over every existing auto-research system.

In response to the reviews, we have updated the submitted paper pdf on:

  1. Abstract: change “new mode-aggregation mechanisms” to “domain-adapted mode-aggregation mechanisms.”
  2. Secs. 1–3: connect the motion-prediction challenges to the harness design, and state that we do not claim generality across domains.
  3. Sec. 2: add a functional comparison with DeepScientist, AIRA, and Agent Workflow Memory, and state that our baselines are alternative scaffold designs, not reproductions of published harnesses.
  4. Sec. 3.2 and Appendix F: describe the seed skills as environment-specific procedures, separating procedural, diagnostic, and automatically distilled skills.
  5. Sec. 4.3: make explicit that β = 0 gives the lowest UAR, explain the trade-off, and note that Table 2 reports one open-ended episode per variant.
  6. Sec. 4.4: credit the expert diagnostic taxonomy in Case 2 and distinguish diagnosis from mechanism design.
  7. Sec. 5: state as limitations the absence of ablations without seed skills or domain tools, of repeated 24-hour component ablations, of an isolated context-separation ablation, and of a comparison with published harnesses at current state.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.