Can a research engine judge the quality of its own research — with no ground truth?

[Pthahnix]

For the last few days I've been chasing one goal, and I've reached the point where I'd rather think out loud in public than keep patching the same design alone. This is half a vision statement, half a request for outside perspective. If you've thought about evaluation-without-ground-truth, self-improving agents, or peer-prediction style mechanisms, I'd genuinely like your read on it.

The vision: a self-referential flywheel

The thing I actually want isn't a one-off study. It's a loop:

Use DARE to do real research, and feed what comes out back into DARE's own evolution.

The engine does research; the research improves the engine; the improved engine does better research. A flywheel that, in principle, spins up on its own.

It's an appealing picture. But every concrete path toward it runs into the same wall, and that wall turns out to be the whole problem.

The wall: there is no ground truth for "good research"

To close the flywheel, you have to be able to tell whether a piece of research the engine produced is any good. And for open-ended research there is no answer key. You can't diff against a correct solution, because there isn't one.

I tried to make this precise, and the precise version is what keeps me up:

• All possible research outputs form a space. The genuinely good ones are some sub-distribution G.
• What an AI judge labels as good is a different region, Ĝ.
• Ĝ is too loose (bad precision): plenty of "looks polished, says nothing" outputs sneak in — fluent, well-structured, buzzword-rich, but hollow.
• Ĝ misses the frontier (bad recall): genuinely novel-good work is, by definition, unfamiliar — so a judge trained on "what good research usually looks like" scores it low and throws it out.

So the judge's boundary is simultaneously too permissive in the middle and wrong at the frontier. And here's the part I've made peace with: no single model and no single module fixes this. Frontier models share a huge fraction of their training data, so "three independent AI judges agree" mostly measures shared bias, not truth. Consensus stops being evidence exactly where you need it most.

There's real theory backing how hard this is — incentive-compatible peer-prediction (Bayesian Truth Serum and its relatives) can elicit honest judgments without ground truth, but only under independence assumptions that correlated, same-family LLMs violate; and there are impossibility results saying that with no structural assumptions, it simply can't be done. So this isn't pessimism, it's the shape of the terrain.

The idea I keep coming back to: don't label quality — control the generating condition

The breakthrough that got me unstuck (for a while) was to stop trying to judge quality and instead manufacture data whose label is the known condition that generated it.

Concretely: two agents talk to each other. One simulates a user; the other is DARE doing the research. All the quality pressure is injected externally, through the simulated user's conversational behavior — a demanding, skeptical user pushes the engine toward better work; a hands-off or absurd user lets it drift. DARE never "fakes" quality; quality grows organically out of the supervision pressure.

The crucial move: the label is not a subjective "good/bad" verdict (that's the unverifiable thing). The label is the generating condition — which persona, how much pressure. That sidesteps the no-ground-truth problem entirely. I'm not annotating truth; I'm controlling the experiment, then asking whether an evaluator can recover the conditions I set.

I've since pushed this into a small "pretraining"-style loop: a controller agent generates a ladder of personas from "genius researcher" down to "absurd contrarian," drives the two-agent conversations, scores two things with a separate model family (did the persona actually get injected? does output quality fall monotonically down the ladder?), and revises the persona generator like training a network — except the loss never gets to see the evaluator it's meant to feed. Information isolation everywhere, so I'm not just teaching the data to flatter the test.

Where I keep springing leaks (the honest part)

The architecture is recursive, and recursion is where my holes come from:

• The evaluation is self-referential — using the engine to judge the engine, an agent to simulate the user of another agent, a "training" metaphor over a text system with no actual gradient. Every layer of self-reference breeds a fresh circularity risk (does the data leak the test? does the label secretly depend on the judgment I'm trying to automate?). I plug one and the next layer reveals another.
• The core problem may be only partially solvable. My plan mitigates the no-ground-truth wall through engineering; it doesn't solve it. Which means there's no clean "done" — there's always one more "looks-good-but-hollow" pattern I haven't enumerated. That open-endedness is exhausting.
• The design has outgrown what I can hold in my head in one sitting. It's deep enough now that re-entering it costs real effort, and any corner that drifts out of alignment springs a new leak.

What I'm hoping to get out of posting this

Not code review — paradigm-level perspective. A few of the questions I keep circling:

• Is "label = generating condition" a sound way to dodge no-ground-truth, or am I just relocating the circularity somewhere I haven't looked yet?
• For separating genuinely novel-good from polished-but-hollow — is there any leverage beyond hand-enumerating failure modes and forcing the judge to check them one by one? Enumeration is adversarial and never complete.
• Has anyone made peer-prediction / BTS-style mechanisms work when the judges are known to be correlated (same model family)? Or is structural independence simply unrecoverable there?
• More broadly: is "evaluate the process (is it structurally sound?) instead of the product (is it true?)" a legitimate retreat, or a quiet way of measuring the wrong thing?

If any of this resonates — or if you think I'm fooling myself somewhere — I'd love to hear it. A focused RFC with the sharpest sub-problem lives in #21; this discussion is the place for the big-picture argument.

Thanks for reading. I'm tired, but I still think the question — can a system honestly assess the quality of its own open-ended work? — is worth getting right.

[Pthahnix]

Follow-up — framing the wall in theoretical terms.

Digging into this further, I realized the wall isn't "just engineering" — it's something more fundamental, and it's worth naming the two distinct kinds of "impossible" so the discussion stays honest:

• Mechanism-design / information-theoretic non-identifiability. This is the one that bites hardest. Incentive-compatible peer-prediction (Bayesian Truth Serum and relatives) can elicit honest judgments without ground truth — but only under independence assumptions. There's an impossibility result (Waggoner–Chen style) that with no structural assumptions at all, it can't be done. The sharp way to say it: given no independent truth signal, the observation "several judges agree" is mathematically indistinguishable from "they share one bias." That's not a bug to fix — it's an identity under the information constraint. Same-family LLM judges sit exactly in the failure region (shared priors → correlated error → the very thing the honesty proof assumes away).

• Computability / undecidability (the Rice's-theorem flavor). Judging "does this open-ended research carry real substance" is judging a semantic property, and semantic properties are undecidable in general. The precision/recall picture in the issue (Ĝ too loose in the middle, wrong at the frontier) reads to me like a statistical shadow of that undecidability.

So this is not "nobody's gotten around to building it." It's a known wall with a known shape.

Why I still think it's workable, though. The impossibility result is about the no-assumption case. The other half of the same theorem says: inject enough structure — judge independence, scale (n≥3), an external decidable anchor — and it becomes solvable again. So the theory isn't a death sentence; it's a spec for what must be engineered: independence, and at least one anchor that is externally checkable rather than a matter of opinion. The reframing I keep landing on is "the missing thing isn't a true signal — it's an independent one."

Open questions this sharpens:

1. Is there a known route to recover peer-prediction guarantees when judges are provably correlated (same model family), or is that region simply lost?
2. What's the minimal external anchor that lifts an open-ended quality judgment out of the non-identifiable region — and how far can it generalize beyond hard checkable facts (does X exist?) into softer substance?

Pointers to the relevant literature — or a correction if I've mis-stated the impossibility boundary — both very welcome.

[lw20250416-afk]

I'm not a native English speaker, so I've been relying on AI translation to follow your post. I'm not entirely sure I've fully understood your approach, so correct me if I'm off. Your scheme is genuinely novel — inferring persona from output might be an excellent fit for the problem you're tackling. But looking at it from the philosophy of science, your persona-ladder assumes that "better research" = "stronger questioning" + "stricter methodology" + "deeper reasoning." This describes what Kuhn called normal science — puzzle-solving within an accepted paradigm. It has almost nothing to do with how paradigm shifts actually happen.

If you train an AI only on pre-Newtonian data, no amount of persona pressure will produce Einstein. Getting from Aristotelian physics to Newton requires reconceptualizing what "force" means — from maintainer of motion to changer of motion. The old framework doesn't just lack the answer; it lacks the conceptual vocabulary to even formulate the question. From inside the old paradigm, the necessity of the new one is structurally invisible. This is a common recent criticism of AI-for-science efforts — I'm curious how you see it.

But philosophy works differently. Its "truth" — if that word even applies — isn't independent of conceptual frameworks. It lives precisely in the tensions between them. "Does free will exist?" has no answer independent of whether you're operating in a physicalist, legal, ethical, or phenomenological framework. Philosophical work is the activity of making these frameworks collide, exposing each other's blind spots, and mapping the space of possible conceptual schemes.

In epistemology, higher-order evidence is evidence about the reliability of your first-order evidence. "I see a red table" is first-order. "I've been drinking, so my vision may be unreliable" is higher-order. Your persona labels function in exactly this way. The first-order product is the argument generated in dialogue. The higher-order product is the record of generating conditions — it doesn't judge whether the argument is correct. It records: "This argument, under these conditions, produced in this way." You've replaced judging good research with recovering what persona level generated it. I think what you're doing operates analogously to higher-order evidence.

[Pthahnix]

First, a note: like you, I'm not a native English speaker either, so this reply is also relayed through AI translation — apologies if anything reads oddly.

Thank you for this comment — it's one of the most valuable responses I've received. Let me reply in a few parts.

On the persona-ladder

You've understood it correctly, and I fully agree: this is exactly where the design hits its ceiling. At best it produces a top performer within an existing system, not someone who breaks the paradigm.

But it's only one piece of the larger vision. "How to make an AI do genuinely groundbreaking science" — I honestly have no idea yet. So I chose to start from a smaller, more reachable goal: build a tool that can make an AI excel within an existing framework, and have it produce a sizable dataset for later analysis. I'm treating the persona-ladder as a sub-problem — not expecting it to hand me the answer, but hoping that, in the process of generating and analyzing this data, I'll stumble onto something I hadn't anticipated.

On higher-order evidence

I find this analogy very apt, and I think it pins down a kind of formalized verdict mechanism: judging how much credence the final result deserves, based on whether the research process itself conforms to sound scientific norms. I believe we're pointing at the same thing.

Before your comment, the relevant ideas in my head had been scattered and never got collected this cleanly. Here's what I'd had in mind:

The persona-ladder produces a dataset where each entry is a triple (research_persona, research_graph, research_result):

• research_persona — the persona injected at the start, i.e. the known generating condition;
• research_graph — the full research process reconstructed by combining what context-management records of the research run with the research spec produced by writing-specs, recovered as a graph/tree (roughly, a research execution trace you can inspect);
• research_result — the proposal the research designs, or the full result after running the experiment (experiments are costly in time and resources, so this may stop at the proposal stage for now).

On top of this, I'd sketched two kinds of "verdict." If I borrow ML terms to draw the analogy, the relationship between them is a bit like the distinction between process supervision and outcome supervision:

• Formal verdict — this is the process side. Think of how a mathematician uses Lean4 to formally prove a theorem outright: in domains like math or code, where correctness is machine-decidable, the signal is clean and instantly checkable (somewhat like the ideal case of a verifiable reward). But research isn't all like that, so the formal verdict also mixes in two softer signals: evaluations of the proposal and result by AI / human reviewers (what people now call LLM-as-a-judge), and a component that's something like taste. What it judges spans both the scientific rigor of the process and the proposal/result themselves.
• World verdict — handing a scientific discovery over to time to test. Like Darwin's theory of evolution or Einstein's relativity; and the flip side too — continental drift suppressed for fifty years, connectionism abandoned as unfashionable. This is more like a heavily delayed, sparse, real reward signal: in the end, time decides what they actually meant for human science and technology.

So my intuition is: higher-order evidence sounds like exactly the kind of concrete philosophical grounding that the formal verdict I want to build has been missing.

And here's where I think the hardest part lies: the formal verdict, given an AI reviewer equipped with the right methodology skills, can to some degree be approximated computationally. But that isn't enough — what I really want is to make some kind of prediction about the world verdict's outcome: in effect, to learn a proxy that gives an early read on a real signal that won't actually arrive for decades. I'm also well aware of the biggest landmine on this path — once the proxy drifts from the true target, you slide into the Goodhart / reward-hacking failure mode, where you optimize the metric and lose the thing it stood for. So my question is: does higher-order evidence have any hope of being pushed in this direction — toward supporting a prior, partially trustworthy forecast of the world verdict?

On "existing concepts freezing the AI's creativity"

I agree with this to a degree, but I don't think it's absolute.

There may be situations we haven't foreseen, where an ability an AI emergently develops under some condition actually yields something genuinely original — we just haven't discovered that latent capacity yet, nor the mechanism to elicit it. Let me draw a distinction here: whether a model possesses a capability and whether we've elicited it are two different things, and the current bottleneck may lie more in the latter. Going one level deeper, the solution might have to come from the AI model itself — architecture, training methods, and so on — which is likewise beyond me at this point.

For that reason, I tend to see this as an out-of-distribution problem rather than in-distribution interpolation. And that connects directly to your Kuhnian point: from inside the old paradigm, the new one is invisible — put in ML terms, a system that only interpolates within its training distribution will, by its nature, struggle to extrapolate beyond it. Here's an analogy: an LLM is trained standing on the graveyard of history — but isn't the human path of learning the same? We too can only learn from what our predecessors left behind. So why is it that humans, specifically, gave rise to the kind of intelligence that can challenge the reigning authority and break past existing limits? I'd very much like to find a technical means to approximate (in the function-approximation sense) that intelligence. I don't believe that, at the current level of technology, humans could truly understand and reproduce it mechanistically — but approximating it seems to me within reach.

One more thing, from my own experience using it: I run Claude, GPT and the like daily with my DARE on top, and these LLMs are already remarkably capable — yet the ceiling of any single research effort still depends on how much the human nudges them toward. And there's a clear difference between their creativity and human creativity, one that's also complementary; I can't quite articulate the feeling yet.

Finally

That's roughly the shape of what I've tried and what I'm still chewing on around this vision. Thank you again for your comment — this angle, coming in from the philosophy of science, is one I'd have a hard time growing on my own.

[lw20250416-afk]

Thanks for your reply. We're now entering territory where things get difficult, so I'll use my own thoughts to respond — I'm not sure if this is the right approach. It's always easier to tear a house down than to build one, so let me start with the critique.

I think of how AI works, simply, as prediction. Here's a situation — I don't know if it's been solved. There's a scientific regularity: under condition C, if A, then B. AI is trained on a huge amount of data in this form. When I ask the AI: under condition D, if A, then what? How should the AI answer? Two hypothetical scenarios. First, it still answers B — this could be right or wrong. Second, following some built-in requirement — say, if the condition changes, refuse to give a direct answer — then it might also refuse to answer many other cases where the regularity genuinely does hold. From my experience using AI, early models were indeed prone to errors, and things have improved a lot now. But has this problem really been solved? In fact, scientists don't handle this situation well either, which is why they need experiments to tell them the answer. So relying solely on something that looks like deductive reasoning is hard pressed to advance scientific development. What I want to illustrate with this is: testing predictions through experiment — what you call world verdict — remains the primary mode of scientific development. Given that AI currently cannot directly intervene in reality (AI-assisted protein folding research and such — I'm not sure whether AI can determine how probable its correctness is in these cases), if you only run what you call the formal verdict, it looks more like a scientific methodology distilled from the history of science — its operation within the AI's data resembles constant repetition, a loop. What I actually want to know is this: must AI predictions be adjudicated in the real world, or can they be tested within AI's own world model? This corresponds to the coherence theory and correspondence theory of truth in philosophy.

Can higher-order evidence support AI predictions to some degree? Intuitively, yes. But what needs to be discussed first is: what does it even mean for AI to make a prediction? Take AI-assisted protein folding research as an example. It seems that the AI's operational state is good evidence that its predictions have some probability of being true — but this rests on the fact that there are many experimentalists working hard to realize the AI's predictions.The basis for supporting that a prediction has a probability of being true still lies in the fact that there are ways to verify it There's a project on GitHub called BETTAFISH that uses multi-agent systems for prediction — mainly using agents to imitate human speech, which can predict public opinion trends and the like. This is also a kind of prediction. Compared to scientific prediction, this is more like predicting the future, or predicting "history." In science, under an established regularity, something that has already occurred can no longer count as prediction. But history is always different, yet has certain patterns — it can always count as prediction. And what about predicting stock market movements? These verifications are much faster, and models can undergo rapid iterative upgrades. I'm not sure if you're interested in these kinds of prediction. For more on higher-order evidence, you can look here: https://plato.stanford.edu/entries/higher-order-evidence/. Also worth noting is the research on peer disagreement — these are relatively new topics in epistemology.

Finally, I want to talk about what interests me. Although the very existence of AI accelerates the fading of anthropocentrism, I still believe AI must serve humanity in order to develop further. I think humans always understand AI's various appearances through their own categories and concepts, and then discuss whether AI can reason or has emotions. This is fine — we just need to avoid jumping to simple judgments. So I agree with what you said: AI may already possess enormous potential. Right now what I want to know is whether AI can have divergent thinking, and it was because of this question that I discovered your project. When I use AI, I feel that context — while helpful — is also a serious constraint. In most exchanges, AI drills deeper and deeper in one direction but cannot give me inspiration. I don't think merely adding requirements like critical thinking, analogical thinking, or brainstorming into the prompt can approximate the feeling of human divergent thinking. We share the same goal in trying to use AI to pursue innovation. Having a monkey spend time typing out the complete works of Shakespeare is still better than constant repetition.

My thoughts are scattered throughout. I originally planned to use AI to polish it, but I felt the AI-modified version didn't match my train of thought, so I'm presenting the original. If there are issues, we can discuss further. If you're Chinese, I can send you the original Chinese text.

[Pthahnix]

一、先说清我想让 AI 做的事

先澄清一下我说的 "预测" 具体指什么. 我设想的场景是: 在一项科学研究真正开始之前 —— 无论是人机协作还是完全由 AI 驱动 —— AI 在寻找方向, 构思 idea 的阶段, 能够预测哪一类方案更有可能真正有益于待解决的研究问题. 可以把它想象成一棵树:

graph TD
  A[Research question] --> B1[Candidate path 1]
  A --> B2[Candidate path 2]
  A --> B3[Candidate path 3]
  B3 --> C1[Sub-question 1]
  B3 --> C2[Sub-question 2]
  C1 --> D1[Sub-approach 1]
  C1 --> D2[Sub-approach 2]
  C1 --> D3[Sub-approach 3]
  C2 --> E1[Sub-approach 1]
  C2 --> E2[Sub-approach 2]

Loading

这里的一条 "完整研究方案", 对应从根问题走到底的一条路径: 有的候选路径本身就是终点 (如 B1, B2), 有的则会进一步分裂出几个都必须解决的子问题 (如 B3 分出 C1, C2, 于是需要各选一个子方案才能凑成一个完整方案). 在这棵树上,我大致区分几类落点:

• CASE-A —— 我们最想要的 比如 D2 + E1 组合出的方案, 既经得起实验检验, 又对科学进步有真正的意义.
• CASE-B —— 有价值的失败 比如 B2 由 AI 提出的, 想法新颖大胆, 但一经实验便被证伪的方案. 错了, 但错得有意义.
• CASE-C —— 需要严格回避的 比如 B1, 以及 D1 + E2 在形式上挑不出毛病, 足以包装成一篇论文, 却对科学进步毫无实质贡献的方案 —— 也就是俗称的 "水论文".
• 现实里当然远不止这三类,还会有各式各样的 CASE-XXX.

我想要的, 就是让 AI 从最开始就尽可能分辨出 "怎样才能做出 CASE-A 的工作", 像遵守严格规范一样主动回避 CASE-C, 同时又被鼓励去自发地多做尝试, 不必畏惧做出 CASE-B 的结果. 这就是我想要的大致效果.

二、为什么这是一种 "预测", 以及我真正想取缔的是什么

再说回 "预测" 这个视角. 把 LLM 的运作理解为一种预测, 确实不能说错 —— next-token prediction 本身就字面上写明了它的原理. 而它在各个领域的后续应用, 某种意义上都是把这种 "预测" 发扬光大: Copilot 代码补全预测下一段代码, AlphaFold 工作从氨基酸序列预测蛋白质结构, GraphCast 等模型预测未来的天气, world model 与 RL 预测环境的下一个状态, 机器人策略预测下一个动作... 我真正想做的东西, 也是这种思路的一种延伸: 假设存在一棵 "人类科技树", 一个对人类科学技术来说值得到来的未来, 让 AI 去预测下一步该往哪儿走, 怎么走, 从而推动这个假设中的未来真正成真.

这里有一个我认为很重要的区别: 我想要的并不只是得到一个 "为真" 的科研结果, 而是要这个结果 真的能推动人类科学的进步. 所以我所说的 AI 预测, 是贯穿整个科研流程始终的 —— 它主要预测的是 "这项研究是否正朝着推动人类科学进步的方向前进", 其次才是它 "是否可能为真". 这两者的主次之分,对我来说是个关键.

需要强调的是: 在我的构想里, 预测只是其中的一部分, 世界的验证依然必不可少, 我从没想过要取缔它. 我真正想 "取缔" 的, 是人类做研究时——在那片广阔的, 布满错误的海洋里, 最终走向那个能通过世界裁决的正确结果 —— 这一路上所绕的弯路, 以及那些毫无意义的水论文. 就像我在 DE-ANTHROPOCENTRIC.md 里写的, "科学最沉重的枷锁, 一直是我们称之为 '自己' 的那个". 这些弯路与水文, 正是那副枷锁在日常研究里的具体形状; 我想要的, 不是取代检验真理的实验, 而是替科学卸下这部分纯粹源于人类局限的内耗.

三、但预测不够 —— 形式裁决与世界裁决的分工

由于我目前能想到的计算手段都落在 形式裁决 之内, 所以我的思路是: 尽可能借助 AI 自身的能力 + 形式裁决, 让它产出更有可能通过 世界裁决 的发现.

另外要补充的是, 我们这套体系并不只有形式裁决. 我们也有专门用来真实运行实验的 experiment-executing skill, 只不过目前能覆盖的还只是 AI/ML 领域. 我始终认为, 真正的科学研究一定要经过实验和时间的检验——这一部分, 是无论前面那些 "研究过程" 类的 skill 做得再好, 都无法替代的.

换句话说,我认同实验是必不可少的, 纯粹的计算无法取代 世界裁决. 但我不愿意接受的是: 除了坐等 LLM 自身能力进步到足以胜任科学研究之外, 我们就什么都做不了. 我不希望唯一的路径是 —— 被动等待 LLM 在 形式裁决 的 benchmark 上分数越刷越高, 再靠它一点一点啃下 世界裁决. 我相信, 在这之外一定存在一些我们现在就能主动去做的事情.

你提出的那个问题 —— 符合论 vs 融贯论, 我先认真去了解学习一下, 以及你提到的 higher-order evidence 这些. 很感谢你把这些方向指给我.

四、关于发散思维和 context

我非常认同你说的 "In most exchanges, AI drills deeper and deeper in one direction but cannot give me inspiration",这个问题我也经常遇到.我同样很赞同 "I don't think merely adding requirements like critical thinking, analogical thinking, or brainstorming into the prompt can approximate the feeling of human divergent thinking" 这个判断.

至于 AI 究竟能不能有发散思维, 我倾向于认为这本身是个 需要用实验去验证 的问题, 而不是靠思辨就能定下来的. 也许我得专门构思一种方法, 探针实验那样去测一测: 在什么样的条件下, AI 会真正岔开方向, 而不只是在同一条路上越钻越深. 这是我接下来想琢磨的事情之一.

而 "context 既是助力也是牢笼" 这一点, 我把它看作一个 工程问题. 我没有系统学过任何 context engineering 的方法, 现在用的这套 context-management 机制, 完全是我自己摸索的一套 prototype. 我会在后续工作里重点优化这一块, 尽量让 context 把助力更大程度发挥, 同时把它带来的阻力压小.

五、最后

很感谢你提供的这些思路和链接 (peer disagreement, 以及 BETTAFISH 这个项目). 我会去认真看看它们, 看其中是否有能启发到我的东西. 顺带一提, 你提到了 AI 辅助的蛋白质研究 —— 我目前正在探索的方向之一, 恰好就是 AI 辅助下的生物学研究, 不过还处在很早期的 scoping 阶段.

我目前是以个人身份, 在学业之余做这个 de-anthropocentric-research-engine 项目, 所以无论时间还是资源都比较有限. 我现在能想到的起点, 就是先从一套 skill 体系做起, 来一步步实现这个愿景, 一边做一边思考, 来捕捉更多助力我实现目标的, 隐藏在我整个 de-anthropocentric-research-engine 项目维护过程中的信号.

---

1. First, what I actually want the AI to do

Let me clarify what I mean by "prediction." The scenario I have in mind is this: before a piece of scientific research even begins — whether human-AI collaboration or fully AI-driven — at the stage where the AI is searching for a direction and forming ideas, it can predict which kind of approach is more likely to genuinely benefit the research question at hand. You can picture it as a tree:

graph TD
  A[Research question] --> B1[Candidate path 1]
  A --> B2[Candidate path 2]
  A --> B3[Candidate path 3]
  B3 --> C1[Sub-question 1]
  B3 --> C2[Sub-question 2]
  C1 --> D1[Sub-approach 1]
  C1 --> D2[Sub-approach 2]
  C1 --> D3[Sub-approach 3]
  C2 --> E1[Sub-approach 1]
  C2 --> E2[Sub-approach 2]

Loading

A single "complete research plan" corresponds to one path running from the root all the way down: some candidate paths are endpoints in themselves (like B1, B2), while others split into several sub-questions that all have to be solved (like B3, which splits into C1 and C2, so you need to pick one sub-approach from each to assemble a complete plan). On this tree, I roughly distinguish a few kinds of landing points:

• CASE-A — what we want most. For example, the plan assembled from D2 + E1: it both holds up under experimental testing and carries real significance for scientific progress.
• CASE-B — a valuable failure. For example B2: a bold, novel idea proposed by the AI that gets falsified the moment it meets experiment. Wrong, but wrong in a meaningful way.
• CASE-C — what must be strictly avoided. For example B1, or D1 + E2: plans that are formally flawless and could be dressed up into a paper, yet contribute nothing of substance to scientific progress — what's commonly called a "filler paper."
• And of course reality holds far more than these three; there are all sorts of CASE-XXX.

What I want is for the AI, from the very start, to distinguish as well as it can "what it takes to produce CASE-A work," to actively avoid CASE-C as if following a strict discipline, while also being encouraged to make spontaneous attempts without fearing a CASE-B outcome. That's roughly the effect I'm after.

2. Why this is a kind of "prediction," and what I really want to eliminate

Coming back to the "prediction" lens. Understanding the workings of an LLM as a form of prediction isn't wrong at all — next-token prediction spells out its operating principle in plain words. And its downstream applications across various fields are, in a sense, all carrying this "prediction" further: Copilot's code completion predicts the next chunk of code, work like AlphaFold predicts protein structure from amino-acid sequences, models like GraphCast predict future weather, world models and RL predict the environment's next state, robot policies predict the next action... What I really want to build is an extension of this same line of thinking: assume there exists a "human tech tree," a future worth arriving at for human science and technology, and let the AI predict where to step next and how, so as to push that hypothetical future into reality.

There's a distinction here that I consider important: what I want is not merely to obtain a research result that is "true," but a result that genuinely advances human science. So the AI prediction I'm describing runs through the entire research process — what it primarily predicts is "whether this research is moving in the direction of advancing human science," and only secondarily "whether it is likely to be true." The order of priority between these two is, for me, the crux.

I should stress: in my conception, prediction is only one part of it; verification by the world remains indispensable, and I never intended to abolish it. What I really want to "eliminate" is, in the way humans do research — within that vast ocean of errors, on the way toward the correct result that finally passes the world's verdict — the detours taken along that road, and the meaningless filler papers. As I wrote in DE-ANTHROPOCENTRIC.md, "The heaviest chain on science was always the one we called ourselves." These detours and filler papers are precisely the concrete shape that chain takes in everyday research; what I want is not to replace the experiments that test truth, but to relieve science of this portion of internal friction that stems purely from human limitations.

3. But prediction isn't enough — the division of labor between the formal verdict and the world verdict

Since every computational means I can currently think of falls within the formal verdict, my approach is this: lean as much as possible on the AI's own capabilities + the formal verdict to make it produce findings that are more likely to pass the world verdict.

I should also add that our system isn't only the formal verdict. We do have an experiment-executing skill dedicated to actually running experiments — it's just that for now its coverage is limited to the AI/ML domain. I've always held that real scientific research must pass through the test of experiment and time — and this part is something that no amount of polish on the upstream "research-process" skills can ever replace.

In other words, I agree that experiment is indispensable, and that pure computation cannot replace the world verdict. But what I refuse to accept is that, short of waiting around for LLM capabilities to advance to the point of being able to do science on their own, there's nothing we can do. I don't want the only path to be passively waiting for LLMs to score higher and higher on formal-verdict benchmarks, and then chip away at the world verdict bit by bit. I believe there must be something, beyond that, we can actively do right now.

The question you raised — correspondence theory vs. coherence theory — I'll go and properly learn about first, along with the higher-order evidence you mentioned. Thank you for pointing me toward these directions.

4. On divergent thinking and context

I strongly agree with what you said: "In most exchanges, AI drills deeper and deeper in one direction but cannot give me inspiration." I run into this problem often too. I equally agree with the judgment that "I don't think merely adding requirements like critical thinking, analogical thinking, or brainstorming into the prompt can approximate the feeling of human divergent thinking."

As for whether AI can actually have divergent thinking, I'm inclined to think this is itself something that needs to be verified by experiment, rather than settled by speculation. Perhaps I'll have to devise a method — a probe-like experiment — to test it: under what conditions will the AI genuinely branch off in a new direction, instead of just drilling deeper along the same path. This is one of the things I want to think about next.

And the point that "context is both an aid and a cage" — I see this as an engineering problem. I've never systematically studied any context-engineering methods; the context-management mechanism I'm using now is entirely a prototype I worked out on my own. I'll make optimizing this a focus of my later work, trying to let context exert its help to a greater degree while shrinking the resistance it brings.

5. Finally

Thank you for the ideas and links you offered (peer disagreement, and the BETTAFISH project). I'll go and read through them carefully, to see whether there's anything in them that might inspire me. As an aside — you mentioned AI-assisted protein research, and one of the directions I'm currently exploring happens to be AI-assisted biological research, though it's still at a very early scoping stage.

For now I'm doing this as an individual, working on the de-anthropocentric-research-engine project in whatever time my studies leave me, so both time and resources are fairly limited. The starting point I can think of is to begin with a skill system, and realize this vision step by step — building and reflecting as I go, so as to capture more of the signals, hidden throughout the maintenance of this whole de-anthropocentric-research-engine project, that can help me reach my goal.

[lw20250416-afk]

我本科学材料科学，研究生学科学哲学，在我看来，你想做的事更贴近科学哲学，如果你想了解科学哲学家对科学研究发现中的理性逻辑的探索，可以看看王巍的《科学哲学问题研究》的第二章和第九章，当然它们没有真的覆盖掉你的东西，你关于人格标签的系列想法非常超前，我也很期待你后面的工作。

最近我依然会探索AI能不能拥有发散思维，我的路线更贴近先明白什么是发散思维，我们用这个词指什么，概念的含义是更替的，我对心理学没有敌意，但是我相信日后AI的发散思维的含义一定会取代人们对发散思维的心理学上的理解。夸张的想，让AI能类似人类科学家一样有灵感的工作，让每个地方能根据当地的情况部署属于自己的科学家AI，这实是令人晕眩的图景。

I did my undergraduate work in materials science and my graduate studies in philosophy of science. From where I stand, what you're trying to do is closer to philosophy of science. If you're interested in how philosophers of science have explored the logic of rationality in scientific discovery, you might look at Wang Wei's Studies in Philosophy of Science, chapters 2 and 9. They don't really cover what you're doing, of course — your ideas about persona labels are genuinely ahead of the curve, and I'm very much looking forward to where you take this.

I'll keep exploring whether AI can have divergent thinking. My approach leans more toward first understanding what we mean by divergent thinking — what we're pointing at with this term. Concepts shift their meanings over time. I have no hostility toward psychology, but I believe that someday the meaning of divergent thinking as realized in AI will replace the psychological understanding people hold today. To put it dramatically: giving AI the capacity to work with something like a scientist's inspiration, and deploying a dedicated scientist-AI tailored to local conditions everywhere — that's a dizzying picture.