acceptodds
Under review as a conference paper at ICLR 2027

AI Research Should Scale Insight, Not Just Compute

Abstract

Scaling compute has driven substantial progress in AI, but performance gains do not necessarily bring a corresponding increase in reusable knowledge about learning. We argue that AI research should scale insight, not just compute. We define insight as reusable, communicable knowledge with a stated scope of applicability. Systematically producing, testing, and accumulating such knowledge could deepen our understanding of learning and guide more effective training, supporting both the science of AI and AI-for-AI research. We explore this approach by actively running experiments and measuring observables to identify learning phenomena, formulate hypotheses, and test interventions. We present OPHIS (Observation–Problem–Hypothesis–Intervention–Speed-up), a conceptual framework for observation-guided research; ComfyResearch, a visual tool for constructing, reproducing, and sharing experiments; and the nanoGPT Observable Library, a dataset of observable trajectories recorded during training. Using the nanoGPT Observable Library, we observe attention-entropy patterns that differ across depth and warmup settings. Applying OPHIS to AutoResearch, we identify interventions that improve validation performance under a fixed training-step budget. We demonstrate ComfyResearch on grokking and edge-of-stability experiments and show how automated experiment generation supports the exploration of learning dynamics. We outline a research agenda for deepening human understanding of learning, accumulating and sharing insights, applying them in AutoResearch, and assessing their value under compute and time constraints.

Primary area:
interpretability and explainable AI
Keywords:
science of AI, training dynamics, observables, AutoResearch, scientific understanding

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_H96hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that AI research should scale insights, not only compute. As they claim, a lot of advances in the field come from running computational experiments to see what works best, but do not necessarily give insight into why this is the case or how these findings might or might not lead to future gains in different settings (e.g., changes in architecture, training data, or method). They provide three contributions: the nanoGPT Observable Library, a dataset of observable trajectories recorded during Transformer training; OPHIS (a framework for observation-guided research); and ComfyResearch (a graphical environment for composing and sharing experiments). The paper concludes with directions for future work (how one could share insights between humans and models, identify observables effectively, etc.).

Critical strengths
  1. I believe the idea of using automated (LLM-based) research techniques to help consolidate insights, tricks, and general practices in the community and speed up model training is relevant and timely. They also have a well-motivated story and a clear argument for why such a framework is needed. The overall framework idea seems like a good starting point for doing this, and the paper indicates good future directions in which this work could be expanded.
Critical weaknesses
  1. It is hard to understand what OPHIS is actually doing in the reported experiments from the provided text. Section 3.3 describes the loop, and Appendix A reports experiments that show some improvements, but one cannot see how these improvements are motivated or what interventions actually lead to them (and whether one can learn anything from them). Also, I find the comparisons with other methods limited, which makes the results hard to evaluate. On nanochat, OPHIS starts from Recursive’s optimized solution and is compared only with that starting point (Appendix A.1, Figure 3). On grokking, the only other method is an LLM that receives the task and training setup but not the OPHIS measurements (the LLM-only baseline, Appendix A.2, Table 1).
Questions for authors

(Clarification) Figure 5 in Appendix A.3 illustrates what you mention as “overfitting early in training under certain architectural conditions”. I don't see enough information to understand exactly what is happening here. This brings me to my broader point: could you provide a complete OPHIS example from the reported study (measured observation, suggested hypothesis, interventions and their results, and final insights and the scope of these results)? I am failing to see the “whole picture” of the pipeline working to discover some “tricks” or insights, as mentioned in the paper. In Section 3.3, you write that OPHIS “outperformed the evaluated LLM-based AutoResearch baselines”. In Appendix A, I can only find Recursive’s optimized solution, which is also the starting point of OPHIS on nanochat (A.1), and the LLM-only baseline on grokking (A.2). Which baselines does this sentence refer to, and are there any other methods included in this comparison? Also, OPHIS tested many interventions, but I don't see how these lead to any insight aside from the improvement; I cannot see enough information on this.

(Clarification) I cannot follow Table 1. Appendix A.2 defines a substantial improvement as reaching the target accuracy “clearly earlier” than the unmodified baseline, which reached 95% test accuracy at about 1,250 steps. What is the numerical cutoff for this, and what counts as a failure? I cannot get a clear idea of these criteria. Also, what does “outside the comparison setting” mean in this case?

(Suggestion) Ideally, the paper would also benefit from a further explanation of how OPHIS works (even if this is relegated to the appendix), so one can understand more of the design.

2 comments
AuthorsOpenReview ↗

Response to Reviewer H96h

Thank you for recognizing the idea as relevant and timely, and confirming that this framework is a well-motivated starting point, and for the specific questions and suggestion to explain OPHIS more fully. We have expanded Appendix A accordingly and answer each point below. Please also see General Response.

1. A complete OPHIS example: forking (Figure 5) (W1, Q1, Q3). This study arose during early, human-guided inspection of observable and loss curves, and is separate from Figure 3.

  • Observation: Training and validation losses initially decreased together, then separated sharply when previously seen text was replayed.
  • Problem: Was this tied to replay, elapsed training time, or a particular model component?
  • Hypothesis: Bigram/trigram value pathways contribute to the divergence under replay. We arrived at this candidate by comparing four data schedules with the model and optimizer fixed — single-shard replay, delayed two-shard replay, sampling with replacement, and alternating 50 new/50 replayed batches — and narrowing the observables by timing/shape filters and manual inspection. The magnitude of the n-gram table updates and their alignment with the parameter updates tracked the replay conditions.
  • Intervention and result: Disabling these pathways (the bigram and trigram value-embedding lookup tables) while retaining unigram value embeddings removed the sharp divergence. Under single-shard replay, seed 42 and 1,000 updates, final validation BPB decreased from 1.6769 to 1.0768.
  • Insight and scope: These pathways contributed to replay-related generalization failure in this configuration, motivating tests under data reuse. We have expanded Appendix A.3 with the process and scope, including how this mechanism relates to the n-gram memory modules now used in frontier models.

2. Which baselines were compared? (W1, Q1) No other methods are included in Figure 3 or Table 1. On nanochat, Recursive's optimized solution is both the re-evaluated baseline and OPHIS starting point under our 2,000-step protocol — further improvement of an already optimized solution. The 7.43 figure expresses the mean BPB difference in baseline standard-deviation units. On grokking, the 30 OPHIS-like candidates are a subset of the 76 tested LLM proposals, using similar intervention types (activation dropout, parameter-norm constraints, entropy regularization) but proposed without OPHIS measurements. These comparisons examine observation-guided intervention; they do not rank complete AutoResearch systems. We have specified these comparison units in Section 3.3.

3. Table 1 definitions. (Q2) Target attainment is the first recorded step with test accuracy of at least 95%, evaluated every 10 steps within 2,000 steps; the baseline reaches it at step 1,250.

  • Substantial improvement: at least 30 steps earlier, i.e., by step 1,220. Of the 44 LLM successes, 26 reached it before step 1,100; 1,100 is not the cutoff.
  • Small improvement: 10 or 20 steps earlier. The 47 such OPHIS outcomes are excluded from the 255 substantial improvements.
  • Failure: not reaching the target within 2,000 steps, or reaching it no earlier than step 1,250. This rule applies to both methods: 48 OPHIS failures and 32 LLM failures (10 missed targets, 15 later crossings, 7 ties).

4. "Outside the comparison setting." (Q2) We evaluate general training interventions within the implemented pipeline. Of 100 LLM proposals, 24 fell outside this scope: 12 encoded modular-arithmetic structure (e.g., circular embeddings), 9 required unsupported operations (e.g., EMA or SAM), and 3 added label-conditioned auxiliary objectives (e.g., center loss by target residue). These restrictions keep the study focused on general training interventions; they do not imply that excluded methods are invalid. Percentages use tested candidates as denominators (350, 76, 30); the 24 untested proposals are disclosed separately. We have added these definitions to Appendix A.2.

Thank you for the answers and for expanding Appendix A. The Table 1 definitions and the clarification on the baselines answer my questions, and the forking example (A.3) is the kind of example I was asking for.

But even reading it as a position paper, I agree with Reviewer CLQo that the OPHIS experiments are still too preliminary to support the main message. On nanochat, OPHIS is only compared with its own starting point, and on grokking, where there is a reference, OPHIS and the OPHIS-like LLM subset have almost the same substantial-improvement rate (72.9% vs 70.0% in Table 1). So I still don’t see what the observations add, and I will keep my score.

As a small note for next time, please highlight the changes in the revision (e.g., in red). It would have made them much easier to check.

Reviewer_CLQoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues for scaling reusable insights into learning alongside compute, and proposes OPHIS (Observation–Problem–Hypothesis–Intervention–Speed-up), a conceptual framework for guiding research and training decisions using observables beyond conventional metrics such as loss. The authors also introduce ComfyResearch, a tool for constructing and reproducing experiments, and the nanoGPT Observable Library for systematically recording training observables. They illustrate how observables such as attention entropy can reveal different training dynamics across model configurations. Finally, they evaluate OPHIS on automated research tasks, reporting improvements in training performance compared with the evaluated baselines.

Critical strengths

Strengths

  • The paper presents an interesting and well-motivated perspective on AI research, emphasizing the importance of developing reusable insights into learning rather than focusing solely on improvements in model performance.
  • The idea of systematically examining different training observables beyond standard metrics such as loss is promising. The proposed tools and initial experiments provide potentially useful starting points for studying learning dynamics.
Critical weaknesses

Weaknesses

  • Unclear practical value of the observables. While considering multiple observables is interesting, the paper does not convincingly demonstrate how they can guide research questions or training decisions. For example, Figure 2 shows that attention entropy follows a different trajectory from loss and QKV weight norms. However, observing different patterns across different measures does not, by itself, establish their significance. The differences across network depths and training configurations suggest possible research questions, but it remains unclear what additional understanding can be gained, how these observations could inform useful interventions, or whether the findings generalize beyond the examined settings.
  • Insufficient explanation of the experimental improvements. In Section 3.3, it is unclear which observables were identified as useful, what hypotheses were formulated, and how the resulting interventions led to the improvements reported in Figure 3. Although the quantitative results are encouraging, the paper would benefit from a concrete example tracing the process from an observation to a hypothesis, an intervention, and its measured effect.
  • Limited evidence for reusable insights. A central motivation of the paper is to produce reusable knowledge about learning rather than merely improve training performance. Although Section 2.1 provides a conceptual definition of insight, it remains unclear how insights are identified and validated in practice. In particular, the experiments primarily demonstrate performance improvements or interesting learning phenomena, without clearly establishing what reusable insights have been obtained or their conditions of applicability. The research agenda would also benefit from more concrete criteria for identifying, evaluating, and reusing such insights.
  • Presentation and clarity. The overall motivation is interesting and generally well explained, but the presentation is uneven. Some conceptual discussions are lengthy, whereas important experimental details, particularly the explanations of the results, are too brief. The paper would benefit from a stronger emphasis on interpreting the experiments and connecting them to the central claims, rather than extensively discussing the broader research agenda. Some technical abbreviations, such as QKV, are not explicitly introduced, while others, such as BPB, could be defined more clearly.
Questions for authors
  • Concrete demonstration of the framework: Could the authors provide a detailed example of how OPHIS improves research on a particular problem, explicitly identifying the Observation, Problem, Hypothesis, and Intervention? In particular, it would be helpful to understand how the selected observables inform the hypothesis and motivate the intervention that leads to improved performance.
  • Limitations and failure modes: Could the authors discuss the potential limitations and failure modes of the proposed framework? For example, under what conditions might the observables fail to provide useful guidance, or the resulting hypotheses and interventions fail to improve training? An example of an unsuccessful application would help clarify these limitations.
  • Scope and applicability: Could the authors clarify the intended scope of the framework? In particular, what types of problems, models, or training settings are expected to benefit from this approach, and what characteristics might limit its effectiveness or generalization to other settings?
5 comments
AuthorsOpenReview ↗

Response to Reviewer CLQo

Thank you for the detailed questions and concrete suggestions, which directly shaped our revision. We also appreciate your recognition of the motivation behind our perspective and of the observable-based approach as a promising starting point for studying learning dynamics. Please also see General Response. We address both weaknesses and questions as follows.

What the observables are for (W1). The library makes internal dynamics available for systematic investigation. Figure 2 shows attention-entropy patterns across depth and warmup settings that loss and QKV weight-norm curves do not reveal; without recording these quantities, the phenomena stay invisible in loss-only summaries. Recording them turns their dependence on depth and warmup into concrete, testable questions about learning mechanisms. The intervention studies below show the next step.

From observation to intervention (W2, Q1). We first run 2,000 training steps while recording the observables. For each observable, we compute the Pearson correlation between its baseline curve and training loss, then test a candidate regularizer in a separate 2,000-step run, comparing final validation bits per byte (BPB). The hypothesis is that pulling an observable toward its future baseline values reduces final BPB within the same budget: the regularizer penalizes the normalized squared deviation from the baseline value 100 steps ahead of the current step, with coefficient 0.01 through step 750.

A success and a failure (Q1, Q2). The sum of absolute values in the third block's attention output correlated positively with training loss (r=0.867); applying this regularizer reduced final validation BPB from 0.934417 to 0.931273. The eighth block's maximum absolute query activation correlated negatively (r=-0.657), yet the same rule increased BPB to 0.983054. Both used seed 42 and the same 2,000-step budget; an intervention fails if it does not lower final BPB relative to the baseline. The second run completed normally but failed this criterion — a concrete case where an observational association did not translate into an effective intervention. Testing and rejecting such candidates is part of OPHIS. These single-candidate runs are distinct from Figure 3's combined configuration, which uses the same attention-output statistic with coefficient 0.007 and additional optimizer changes.

Reusable insight and scope (W3, Q3). A successful intervention gives local evidence; an insight is the reusable knowledge we extract from it — knowledge that supports prediction or explanation, or guides an intervention, within a stated scope.

  • Reusable insight The early, human-guided forking study (our response to H96h, now in Appendix A.3) illustrates this: observable inspection led to a hypothesis about replay-related generalization failure, and component removal tested the contribution of n-gram pathways.
  • Scope OPHIS applies when training can be instrumented and hypotheses tested through controlled interventions, as in our Transformer and modular-addition studies. Beyond this scope the loop becomes harder to close: a signal may be too noisy to suggest a testable hypothesis, or too costly to measure to be worth acting on. How OPHIS's benefit trades off against signal quality and measurement cost is therefore left open, and we make this a research priority in Section 4.

Presentation (W4). Thank you for pointing out the uneven exposition. We have expanded Appendix A with the process and these examples, and defined QKV (query, key, and value projections) and BPB at first use.

Thank you for your reply and for updating the manuscript.

What the observables are for (W1): I understand that tracking metrics like attention entropy can surface differences across depth and warmup that are not visible in the loss curve. However, in the paper, the Figure 2 observations simply stop at raising questions, they are not tested or used to guide an intervention, so their practical value remains unclear to me.

From observation to intervention & A success and a failure (W2, Q1, Q2): Thank you for explaining how the interventions work and providing a failure case. However, computing Pearson correlations with the loss and penalizing deviations from a pre-recorded baseline curve 100 steps ahead (t+100) still looks like a heuristic trial-and-error search rather than insight-guided research. It does not explain why regularizing Block 3 attention output helps on a single seed while regularizing Block 8 query activation makes it worse, and a rule that requires a pre-recorded baseline trajectory cannot be directly transferred to new settings. Furthermore, since you note that Figure 3 combines this regularizer with additional optimizer and learning-rate changes (and candidates were only screened on a single seed), it remains unclear to me how much of the gain in Figure 3 actually comes from the observable regularizer.

Reusable insight and scope (W3, Q3): You note that OPHIS interventions only provide "local evidence" and point instead to the "forking" study (Appendix A.3) as an example of a reusable insight. However, since the forking study was a human-guided debugging analysis (manually testing data-replay schedules to see that n-gram lookup tables overfit on repeated text), it does not demonstrate that the proposed automated framework extracts reusable insights. Leaving the trade-offs between signal quality and measurement cost entirely to future work also leaves the practical scope of the approach open.

Presentation (W4): Thank you for defining QKV and BPB and expanding Appendix A. However, the main text itself remains largely unchanged: the conceptual discussion is still lengthy, while the actual methods and experimental details needed to understand Section 3.3 are only in the appendix.

Overall, your clarifications helped me better understand what was done, but they also confirm my concern that the experiments do not convincingly demonstrate the scaling of reusable insights. I will therefore keep my score.

AuthorsOpenReview ↗

Response to Reviewer CLQo's follow-up [part 1/2]

Thank you for the follow-up and for engaging with the additional experimental details. Below we questions in two parts.

[part 1/2] “the Figure 2 observations simply stop at raising questions, they are not tested or used to guide an intervention” (W1) Figure 2 demonstrates how systematic observation generates specific, testable questions about learning dynamics. Comparing attention entropy across blocks, model depths, and warmup settings exposes structure that loss and QKV weight-norm summaries do not reveal. This is the role assigned to the example in Section 3.2. A mechanism explaining this phenomenon or a performance intervention based on it would require a further study. The intervention studies in Appendix A illustrate the subsequent research steps; each example need not independently complete the entire process.

“still looks like a heuristic trial-and-error search rather than insight-guided research” (W2, Q1, Q2) OPHIS does more than record correlations: it tests whether manipulating observable trajectories improves training, distinguishing helpful from unhelpful interventions under the stated conditions. The hypothesis is concrete: moving a selected observable toward its later baseline state may improve performance within the same training budget. The workflow thus tests an intervention suggested by the measurements and records its effect, rather than treating correlation itself as the answer.

“It does not explain why regularizing Block 3 attention output helps on a single seed while regularizing Block 8 query activation makes it worse” (W2, Q1, Q2) In the reported seed-42 screening runs, the third-block attention-output regularizer reduced BPB from 0.934417 to 0.931273, whereas the eighth-block query-activation regularizer increased it to 0.983054. The same intervention rule therefore produced different outcomes depending on which observable was targeted. This supplies empirical knowledge about which trajectory manipulation helps in the tested setting. As Section 2.1 explains, such intervention knowledge need not await a complete mechanistic account of why the outcomes differ.

“a rule that requires a pre-recorded baseline trajectory cannot be directly transferred to new settings” (W2, Q1, Q2) The reference trajectory is measured for the training setting in which the intervention is applied. Applying this procedure in a new setting may require a new reference run; the method does not assume that the same numerical trajectory applies to every model. Transfer of a particular empirical finding is a separate question, explicitly addressed in Section 4.2, “Accumulating and sharing insights.”

“it remains unclear to me how much of the gain in Figure 3 actually comes from the observable regularizer” (W2, Q1, Q2) The regularizer was also tested on its own: in the reported screening run, adding it reduced validation BPB from 0.934417 to 0.931273. That test used coefficient 0.01. Figure 3 separately evaluates a configuration using coefficient 0.007 together with optimizer changes. The isolated test measures the regularizer's effect in that run; Figure 3 reports the combined configuration's performance. Appendix A.1 explicitly distinguishes these experiments. They do not estimate what fraction of the combined gain is attributable to the regularizer, and we make no such attribution.

“it does not demonstrate that the proposed automated framework extracts reusable insights” (W3, Q3) The automated intervention study addresses the same underlying concern as W2: OPHIS tests which trajectory manipulations improve training, providing a concrete route from observations to experimentally supported intervention knowledge. Forking provides another example: comparing replay schedules and disabling the implicated pathways identified their contribution to generalization failure in that configuration. In Section 3.3 we implement OPHIS through AutoResearch; the same methodology can also be carried out by human researchers. Our General Response explicitly includes both forms of research. The human-guided example complements the automated intervention study.

Reviewer_hwYHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the optimization behavior for iterative research agents, arguing agentic iterations should "scale insights beyond compute". The paper considers "insights" as reusable, communicable knowledge with a stated scope of applicability, and argues AI research should prioritize systematically producing, testing, and accumulating such knowledge so as to deepen our understanding of learning and guide more effective training. The paper proposes OPHIS (Observation–Problem–Hypothesis–Intervention–Speed-up), a conceptual framework; ComfyResearch, a visual tool for constructing, reproducing, and sharing experiments; and the nanoGPT Observable Library, a dataset of observable trajectories recorded during training. The paper provides demonstrations on how these tools may assist observation and discovery in AutoResearch. Finally, the paper presents a research agenda advocating for deepening human understanding of learning and accumulating and sharing insights in AutoResearch.

Note: the paper is submitted under "Primary Area: interpretability and explainable AI", but the construction of the submitted manuscript is most aligned with that of a position paper. The paper prioritizes claims and arguments and advocates for a proposed research agenda. The experiments in the style of pilot studies, supplemented with demos for a prototype system implementing the paper's ideas.

Critical strengths
  • The advocated scientific direction, prioritizing insights as a first-order factor in AutoResearch, is valid.

  • The presented tools look interesting. They may have their value when used in the right way.

Critical weaknesses

As noted above, the paper is submitted under "Primary Area: interpretability and explainable AI", but the construction of the submitted manuscript is most aligned with that of a position paper. The paper prioritizes claims and arguments and advocates for a proposed research agenda. The experiments are in the style of pilot studies, supplemented with demos for a prototype system implementing the paper's ideas.

For AutoResearch-style agentic research system, it is not a new idea to consider leveraging the insights accumulated during the agent's explorations and facilitating its future iterations. In most cases, the agents have full visibility to its experience which will, more or less, affect their further explorations. The ongoing research frontier is to better organze/manage the context/experience for more effective explorations.

Admittedly, the framing for the very original AutoResearch project (nanoGPT) resembles zeroth-order optimization. Yet, a wealth of forthcoming research has developed systems that improve further explorations based on accumulated experience and findings (e.g., AI Scientists, self-improving agents/recuisive self-improving (RSI) systems).

Further, a number of works incorporate more delicate components such as world models or memory systems that may enable beyond first-order optimization.

Questions for authors

What is the main contribution of this paper? Do the authors consider the submission to be original research or a position paper?

3 comments
AuthorsOpenReview ↗

Response to Reviewer hwYH

Thank you for recognizing insight as a first-order factor and the tools as interesting and potentially valuable. Please also see General Response. We address both weaknesses and questions as follows,

Paper type and contribution. (Q1, W1) This is a position paper, supported by original tools and exploratory studies.

  1. The ICLR 2027 Reviewer Guidelines explicitly recognize that submissions may pursue different research objectives, including drawing attention to new problems, and that these objectives warrant different considerations of value and impact. There is also a precedent for such contributions in the ICLR main conference: Ngo et al., The Alignment Problem from a Deep Learning Perspective (ICLR 2024), explicitly presents itself as a position paper whose primary contribution is conceptual analysis rather than novel empirical or theoretical findings.
  2. Its contribution is a new view and a research agenda: AI research should devote part of its computational budget to discovering reusable knowledge about learning, and develop ways to scale that discovery. The scope includes human research as well as AutoResearch — knowledge about learning should guide the choice of architectures, training procedures, and experiments. This concern with understanding learning also motivates the placement under interpretability and explainable AI. And the three explorations (Observable Library, OPHIS, ComfyResearch) make the agenda concrete.

W2: For AutoResearch-style agentic research system, it is not a new idea to consider leveraging the insights accumulated during the agent's explorations and facilitating its future iterations. We agree that experience reuse is valuable and not new. But using experience is not the same as extracting insight. Our question is whether the system has extracted reusable knowledge about learning — a testable relationship that researchers can inspect, check, and apply beyond the trial that produced it. The width-dependent relationships behind muTransfer are an example: they guide hyperparameter choices across model sizes, not just within one tuning run. Sharing raw records does not by itself extract such a relationship.

W3: a wealth of forthcoming research has developed systems that improve further explorations based on accumulated experience and findings (e.g., AI Scientists, self-improving agents/recuisive self-improving (RSI) systems). We appreciate the reviewer's concern regarding the novelty of our contribution. However, we respectfully believe that the possibility of similar work emerging in the future should not, by itself, constitute a weakness of the present submission. Novelty should be assessed against identifiable existing research, rather than hypothetical future publications. Indeed, the ICLR 2027 Reviewer Guidelines explicitly state that even the lack of comparisons with contemporaneous or unpublished arXiv papers cannot serve as grounds for rejection. We would be happy to discuss any specific, publicly identifiable work that substantially overlaps with our central contributions. In the absence of such evidence, we respectfully suggest evaluating the novelty and significance of our submission based on its actual contributions and the existing literature.

W4: a number of works incorporate more delicate components such as world models or memory systems that may enable beyond first-order optimization. As you mentioned, these works only "possibly" offer first-order or higher insights. Actually, the components alone do not guarantee the kind of insight we define. Here we use two examples to demonstrate:

  • Memory systems. Reflexion stores verbal reflections that guide later trials. But the reflections are guidance tied to the tasks they came from; they do not by themselves yield a tested relationship about learning. Better context management and insight discovery are related goals, but not the same goal. Actually, we think the memory system could be a useful component of future "scale insight" tool or system.
  • World models. Dreamer learns predictive dynamics for planning and control. A predictive relationship encoded in a model is not automatically an explicit account that researchers can inspect and reason with. When a world model does surface such knowledge, it advances exactly the direction we advocate.

Why the three explorations matter. The Observable Library expands systematic measurement; OPHIS connects observations to hypotheses and intervention tests; ComfyResearch lowers the cost of building and reproducing experiments. Together they explore how to make insight discovery a scalable research activity — a broader contribution than an additional context-management mechanism or a new AutoResearch optimizer.

Reviewer_hwYHOpenReview ↗

Follow-up questions

Thanks for the dedicated response and discussion.

Two follow-up questions:

  1. Do the authors consider this submission's contribution as a position paper or an original research method?

These are two very different types of contributions and submissions are reviewed under substantially different criteria.

For example, a position paper prioritizes community impact--will this paper, if published, spark meaningful discussions--regardless of the reviewer agree with the paper's opinion or not.

In comparison, a paper presenting anoriginal research method requires clear methodology, strong rigor, and thorough evaluations.

The current manuscript falls short for either criteria.

  1. As the rebuttal noted, there are a wealth of related work towards the same direction. These works are not clearly presented/effectively discussed in this work, which is a necessary first step to establish the research gap to motivate this work.

This part needs major revision if the paper wishes to present its proposed tools under original research method. Also needed are empirical comparisons with existing methods implementing similar/related ideas.

Hope it helps :)

AuthorsOpenReview ↗

Response to Reviewer hwYH's follow-up

Thank you for the follow-up and for continuing the discussion.

“Do the authors consider this submission's contribution as a position paper or an original research method?” (Q1) Our previous response explicitly answered: “This is a position paper, supported by original tools and exploratory studies.” The General Response also distinguished the central position from those explorations. Including original tools and experiments does not change the paper's primary contribution into a standalone method.

“a position paper prioritizes community impact” (Q1) Our proposal concerns research priorities: allocating part of the computational budget to discovering regularities about learning that can guide subsequent research. As explained in Sections 2.2–2.3, this includes systematically discovering learning phenomena and developing experimentally supported intervention principles for human and automated research. The General Response identifies how the three explorations support this agenda.

“The current manuscript falls short for either criteria.” (Q1) The follow-up does not explain which central argument remains unconvincing as a position after our clarification. Please identify the specific unresolved objection to the position and explain how it relates to the distinction made in our first response. Repeating the question about contribution type does not identify that objection.

“These works are not clearly presented/effectively discussed in this work” (Q2) Sections 2.2 and 2.3 discuss the science of AI and automated research, including The AI Scientist and automated training-method development. Our first response further distinguished experience reuse from extracting explicit, testable knowledge about learning. The presence of memory or a world model does not by itself establish that a system produces regularities researchers can inspect, test, and apply. Please identify the specific works and aspects of our positioning that remain insufficiently distinguished, so that we can address the overlap directly.

“Also needed are empirical comparisons with existing methods implementing similar/related ideas.” (Q2) Appendix A reports the comparisons supporting the two training studies, and our first response to H96h, point 2, specifies exactly which baselines were evaluated. These studies provide concrete explorations of the position. The paper does not claim superiority over complete AutoResearch systems. A broader system-level comparison would therefore address a different claim.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Comment

Dear Area Chair and other reviewers,

Thank you for the thoughtful reviews. Several reviews read the paper's position as being about agentic RSI/AutoResearch. However, our position is that AI research in general should scale insight, not just compute — that part of the computational budget should go toward discovering reusable knowledge about learning. OPHIS-style agentic RSI/AutoResearch is one concrete technical route toward this goal, not the position itself; the argument applies to human research as much as to automated research.

By "insight" we mean reusable, communicable knowledge with a stated scope of applicability — knowledge that, within its scope, should support qualitatively correct prediction or explanation, or help guide an intervention (Section 2.1). It need not be a complete mechanism, and it can take many forms. We especially value compact empirical or theoretical regularities that compress observations and guide research. Kepler's laws illustrate the kind of compressed knowledge we have in mind: an empirical law can guide research before a complete mechanistic explanation is available. Communicability and reusability are properties of this knowledge; the regularity itself is what provides guidance.

This position paper is supported by three complementary explorations:

  • The nanoGPT Observable Library records internal training trajectories, expanding the empirical basis for relationships that conventional performance metrics may miss.
  • OPHIS connects observations to problems, hypotheses, and intervention tests, in both human-guided and automated research.
  • ComfyResearch represents experiments as executable graphs, lowering the cost of building, modifying, and reproducing experiments so these investigations can expand.

A full technical treatment of each project is outside the scope of a position paper. Since the reviewers were concerned about the experimental details, we have added some additional experimental details to the appendix in the revised version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.