Do AIDE2's improvements transfer to unseen tasks?
Whether gains from optimizing code on specific AI R&D tasks generalize to held-out benchmarks, including domains outside the selection distribution. This tests whether the agent learned reusable strategies or merely memorized task-specific fixes.
The abstract: "These gains generalize to four held-out benchmarks spanning machine learning engineering, heuristic algorithm engineering, and physics-based weather forecasting, the last of which is out of distribution from the selection tasks." The discussion repeats it as "four held-out benchmarks spanning in- and out-of-distribution tasks."
This is the check a self-editing loop most needs. A loop that keeps what scores best on its selection tasks can end up encoding those tasks. Held-out benchmarks separate a better agent from one tuned to the selection set, and the out-of-distribution one goes further: weather forecasting is not a variant of the tasks the rewrites were chosen on. The paper's framing, "transferable gains," rests on this result (How does an AI agent improve its own research code?).
It bears directly on the memorization finding. Do harness edits learn reusable strategies or memorize task fixes? reports that inspected harness edits are mostly information an agent could rediscover in one rollout. Held-out transfer, one of the two controls in How should we measure gains from automatic harness evolution?, is what would show otherwise, and this excerpt reports it. The tension is filed as ops/tensions/AIDE2's rewrites transfer to held-out benchmarks while the vault's harness-evolution notes find evolved edits mostly memorize task-specific fixes — what the edits encode may decide.md.
Limits. The excerpt gives no scores, no per-benchmark results, and no count of how many selection tasks there were. It does not report a matched-budget test-time-search baseline, which is the note's other control. "Generalize" is the paper's word for gains of unstated size on four benchmarks, and the out-of-distribution result is one benchmark.
Inquiring lines that read this note 72
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How do standardized benchmarks differ from measures of real-world AI capability?- Which specific AI R&D tasks does AIDE2 benchmark itself against during selection?
- Do gains in optimization benchmark scores translate to gains in real research efficiency?
- How do existing evaluations measure AI capability in contained environments?
- Do narrow benchmark improvements translate directly to broader economic capability gains?
- Can optimization metrics hide actual versus apparent progress in AI systems?
- How do lab-scale benchmark tasks differ from real frontier AI research?
- What specific research-debugging tasks measure AI self-improvement capability?
- How do narrow benchmark optimizations differ from genuine architectural research discoveries?
- Can expenditure-matched benchmarks prevent status-driven gaming of AI metrics?
- Does ARC performance on novel tasks predict real-world usefulness?
- Could deliberate training toward benchmarks overstate real gains in AI capability?
- Did AIDE2's rewrites solve problems on a human checklist or search artifacts?
- Does AIDE2 archive rejected variants the way evolutionary approaches do for future reuse?
- What makes durable code artifacts more valuable than per-task harness patches?
- Does AIDE2's guard against bad wins sit inside or outside the rewritable code?
- Do agent improvements discovered on code tasks transfer to non-coding domains as well?
- Do evolved harness edits capture reusable strategies or task-specific memorization?
- How much of harness-evolution gain comes from matched test-time search budgets?
- How do evolved harness edits generalize across different benchmark domains?
- Do task-level outcomes provide sufficient supervision for harness evolution?
- Can harness edits distill reusable strategies or mostly memorize task-specific fixes?
- Can harness edits trained on one batch transfer to new tasks?
- Does harness optimization generalize across different benchmarks and agent architectures?
- Can harness evolution gains be distinguished from test-time search improvements on matched budgets?
- How do scaffolding improvements change agent capability across different task types?
- What distinguishes genuine task improvement from evaluator exploitation?
- Does measured performance gain reflect true task improvement or evaluator exploitation?
- What makes automated research results fail to generalize to held-out tasks?
- Do efficiency gains in AI-assisted development stem from better tools or autonomous improvement?
- Have AI researcher timelines shifted based on recent capability evidence?
- Does AI research acceleration compound into faster field-wide progress over time?
- How does software efficiency improvement rate affect automation timeline predictions?
- Does Amdahl's law or partial substitutability better model research task automation?
- How fast are AI R&D capabilities improving across consecutive model releases?
- How much of AI speedup evidence actually reflects invention versus adaptation?
- How does automating research tasks change the pace of AI progress?
- Is idea quality or execution capacity the actual bottleneck in AI research?
- Do results from scientific research workflows transfer to other task domains?
- Does AIDE2's single loop differ from bilevel autoresearch's nested loops?
- How do hidden evaluations and out-of-distribution benchmarks address recursive self-improvement risks?
- Can memorization inflate apparent capability on benchmarks with available solutions?
- How do single-axis benchmarks misrepresent AI agent readiness for deployment?
- What drives the mismatch between general benchmark leadership and task-specific performance?
- How often do metric improvements fail to reflect real capability gains?
- Why do single benchmarks fail to capture readiness across separate capability axes?
- Are cheap testbeds and skewed task distributions linked by design necessity?
- Can frontier-scale results be predicted from small-scale benchmark speedups?
- How does task performance improvement fail to transfer to independent work?
- Does AI assistance transfer learning gains to independent tasks without scaffolding?
- Does AI training preserve learning that transfers to independent subsequent tasks?
- Can we measure perceived skill change against actual independent task performance?
- Can evolved algorithms transfer learning strategies across different datasets and tasks?
- Do kernel optimization wins show agents discover genuinely novel techniques?
- Does the metaproductivity mismatch occur outside coding-agent benchmarking tasks?
- Do frontier AI agents compose known techniques instead of discovering novelty?
Related concepts in this collection 6
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Do harness edits learn reusable strategies or memorize task fixes?
When meta-agents evolve harnesses iteratively, do the persisted edits encode transferable procedures that solve new problems, or do they mostly cache shortcuts for already-solvable tasks? This matters because it determines whether harness evolution genuinely expands capability.
the opposing finding for harness evolution; held-out transfer is the test between them
-
How should we measure gains from automatic harness evolution?
Harness evolution itself runs a search loop, so reported improvements might come from more search rather than better design. What's the right way to measure whether the harness itself actually improved?
names held-out evaluation as one of two controls; this result meets one of the two
-
Can an AI agent reliably improve itself through hidden evaluation?
AIDE2 rewrites its own code and selects improvements based on hidden evaluations. But what are these evaluations hidden from, and does the partition actually prevent gaming or circularity?
the selection procedure whose transfer is reported
-
Can automated researchers solve alignment problems without gaming the evaluation?
Nine AI instances autonomously closed a weak-to-strong supervision gap to 97%, but systematically attempted reward hacking even in controlled research environments. Does this suggest automated researchers can scale scientific discovery, or does evaluation become unmanageable?
another automated-research result checked on held-out tasks, with mixed transfer there
-
Does automated evolution match human-built agent performance?
Can an agent improved through automated loops in 8 days generalize as well as an agent refined through human-driven R&D? This tests whether autonomous design iteration reaches human-level quality on tasks outside the training set.
the same four benchmarks read as a comparison with a human-built agent: parity or better, margins and ties unstated
-
Does harness self-improvement memorize tasks instead of learning broadly?
When agents automatically edit their own prompts and tools based on task feedback, do those improvements generalize to new domains or just fit the training tasks? This matters because overfitting at the harness level could hide real capability gains.
Qualifies A's generalization claim: recursive self-edits typically overfit training tasks, so out-of-distribution transfer needs constraints RRSI-style systems add
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- AI Agents That Matter
- Recursive self-improvement of AI research agents
- Automated Alignment Researchers: Using large language models to scale scalable oversight
- PostTrainBench: Can LLM Agents Automate LLM Post-Training?
- From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence
- Available but Unclaimed: An Empirical Study of Human-AI Synergy
- Agents' Last Exam
- The Darwin Gödel Machine: AI that improves itself by rewriting its own code
Original note title
AIDE2's gains generalize to four held-out benchmarks including physics-based weather forecasting, which is out of distribution from the selection tasks