Chain of Recursive Thoughts: Make AI think harder by making it argue with itself

Your references show me that it is absolutely task depended. In many domains it's true that "criticizing is easier than creating".

The best example might be books and movies, where it's trivial to say the characters were shallow, but it's surprisingly hard to create deeply interesting characters.

In Software Engineering, there are similar dynamics. An LLM with a security vuln finding prompt will be able to point out places, where the generated code might be insecure.

But if you want another LLM to find a reasoning mistake in a mathematical proof it basically has to do all the reasoning work as well. In which case I doubt there will be any significant performance gains.

For better or worse this has become the defacto standard in LLM Evaluation research papers since the LLM as a Judge paper [0] came out. Its also heavily embedded into frameworks like LangChain and LlamaIndex to evaluate RAG pipelines.

[0] https://arxiv.org/abs/2306.05685

[1] https://arxiv.org/abs/2411.15594

> ...so you need a checker that actually reasons about the world (compiler, linter, SAT solver, ground-truth dataset, etc.).

Agree. What do you think about telling the LLM to also generate unit tests for the code it spits and then run all tests (including previous application unit tests).

I think this is a way to ensure some level of grounded verification:

- Does code compile?

- Do unit test pass?

AI can then consume test results to help fix their own mistakes.

I believe, what the smart AI company is trying to do, right now, in secret, is to use US, the humans, and our replies to the AIs, as training for the next generation of self-verifying-models. :)

Training on corpus data gets you to 1 order of magnitude. But training on interactive data where you can observe and adapt to the OODA-loop? So much more powerful.

At least, that's what I'd be doing if I were doing AI :)

But I just do BrowserBox

I assume everyone knows this, but the idea of generating answers and testing them, dates back decades, and has been widely used for problems where generating _the_ correct answer(s) is difficult, but where generating a bunch of potential answers--(at least) one of which is likely correct--is easier. Generate-and-test of course relies on having a test algorithm that is reliable, (relatively) fast, and memory efficient, and is most useful when an exact generate algorithm (one that generated only the correct answer(s)) is either slow or inefficient of memory use (or both).

In the case described, the generator is an LLM, and the tester (called a "verifier") is "the compiler, linter, SAT solver, ground truth dataset, etc."

And of course generate-and-test is related to trial-and-error, which has probably existed since the Paleolithic.

"letting GPT-4 critique its own answers reduces accuracy"

This is because the output, being the input, steers directly into the tree as soon as the tree is in the context window.

Would a LLM under human guidance turn out to be a good verifier ? i.e. if LLM knows the rules to verify or has enough data points (internet access, actual responses)

Of course, that only works for problems where you have a verifier.

Ah, now we know why Spain was out of electricity yesterday.

For anything semi-adversarial, I have had good results asking the AI to come up with a plan, then take the side of the opponent coming with counter play/way to defeat the plan, finally asking for a revision of the initial plan given the potential reaction from the opponent.

The final plan you obtain is generally a lot more well rounded and thought out.

I find that amusing because the technique also works when I apply it to me. Picking flaws in your plan before revisiting it actually works.

I do the same, and I have one other technique.

I will often have a few chats going for a project, but with different contexts. For example, one might be tech focused, another marketing focused, another with some context on my personal goals, etc.

So I will take the same question and feed it into the chats with differing context. It is almost like having different perspectives on the same problem. And the conclusions can often differ based on the differing contexts.

This reminds me a lot of the YT video that went over using Monte Carlo Tree Search with LLMs to maximize result quality. Link: https://www.youtube.com/watch?v=mfAV_bigdRA&ab_channel=Treli...

It seemed like a pretty good idea, though I'd guess that it would greatly increase token usage. I'd also be concerned that the LLM as a judge might struggle to grade things accurately if it wasn't also able to generate good enough answers to begin with.

Kagi’s Assistant feature makes this super easy. Just switch assistants and ask them to check the other’s work.

I do it all the time in Sillytavern in a group chat - three characters kind of resembling what you just described, and me, participating in the "conversation", them going back and forth until they're satisfied.

With a good model role playing them, works awesome.

We're there any situation that first conclusion from AI was completely changed? Can you give generally examples of situations where it changed or significantly improved overall result? It sounds cool.

Isn't this kind of another way of how Inference Time Scaling works? It will basically produce several chain of thoughts and then pursue one that has maximum reward based on an internal function?

In doing some DevOps-y type tasks recently (ansible, packer, docker, baking images with guestfish), I've found it very frustrating how much ChatGPT will confidently tell me to use flags on tools that don't exist, or hallicinate completely non-existent functions or behaviours. And then when I spend time trying what it suggests only to hit a wall and come back like wtf mate it breezily goes "oh yes so you're right, good job figuring that out! You're so close now! Your next step is to do X and Y," and then serves up the same detailed tutorial as before but with the flag or whatever it was that it had wrong subtly changed.

It definitely makes me feel like I'm dealing with an overenthusiastic intern who is throwing stuff over the wall without checking their work, and like maybe having a second bot sitting in front of the first one being like ARE YOUR SURE ABOUT THAT could really improve things.

A year or so ago I experimented with splitting a user prompt down to a set of "different AI personas" that would each try to approach the user's problem in a different way and then bubble back up with a master arbiter for consensus.

I modeled it after the concept of advisors from Civilization II. It worked reasonably well though I think it was at least somewhat limited by being constrained to a single LLM (Mistral). It also lit my computer on fire.

In theory couldnt this just be baked into a single adversarial model?

Like, just endlessly grinding tokens, then processing the output and pulling out good ideas when the endless debate generates them?

Would be interesting what it comes up with with enough time and tokens.

This is being done, and you could apply it to a lot of domains. Go for it for whatever use case you have.

These ensembles have been tested throughout AI progress. Well scaffolded larger models have historically come out ahead in both quality and speed/cost.

Perhaps this is a parricularly effective ensemble, but I would need to see real data.

Yeah, but we'll finally get definitive proof that the government's been hiding super-intelligent axolotls from us all.

A society of mind, if you will. :)

This sounds like a fun thing to set up with a quick-enough local model.

This is one of the reasons I like the massive context window in Gemini. You can do this as part of the message chain. I don't try to one shot it, just use the same idea across 3 messages.

1. Figure out a plan (it responds with the plan)

2. Point out flaws in the plan (it responds with the flaws)

3. Update the plan to address the flaws (it responds with an up to date plan)

The other things I tend to ask are "what might we be missing?", "what are the [performance|security|legal|cost] considerations?". I can often iterate on the "anything else?" kind of nudging prompts, especially guiding it on topics to consider, for a few messages. After each: update the plan to take those into consideration.

I always do "now again but put on your critical hat"

Oh I really like that. It makes me want to have it score its ideas with metrics and then keep iterating until it meets some score.

Also a little clickbaity with "my AI" and then it's all Mistral...

Check out Fast Agent! (I have no affiliation with it, just use it).

https://github.com/evalstate/fast-agent

Techniques like this have been around since GPT-3.5. There are boatloads of papers on the topic.

I have no idea why anyone thinks this is novel. I guess that speaks to the state of HN

Chatgpt shares context between chats. I wonder how that impacts it?

It seems like a good approach though. What you dont want to do is ever suggest that its wrong yourself. Usually it will just assume it is wrong.

Actually what I find impressive is when I do this and it actually pushes back to defend itself.

I am thinking the same thing! Multiple "personalities", in parallel, or in series. For example, I have approximated, in GPT, some of Gemini's ability to call out nonsense, sloppy thinking, by telling GPT to be mean! (The politeness seems to filter out much that is of great value!)

However, the result is not pleasant to read. Gemini solved this in their training, by doing it in two phases... and making the first phase private! ("Thinking.")

So I thought, what I need is a two-phase approach, where that "mean" output gets humanized a little bit. (It gets harsh to work in that way for more than short intervals.)

As a side note, I think there would be great value in a UI that allows a "group chat" of different LLM personalities. I don't know if such a thing exists, but I haven't seen it yet, although the message object format seems to have been designed with it in mind (e.g. every message has a name, to allow for multiple users and multiple AIs).

Even better if it supports multiple providers, since they have different strengths. (It's like getting a second opinion.)

I think you can do most of this already with llm-consortium (maybe needs the llm-openrouter plugin with my pr merging)

A consortium sends the same prompt to multiple models in parallel and the responses are all sent to one arbiter model which judges the model responses. The arbiter decides if more iterations are required. It can also be forced to iterate more until confidence-threshold or min-iterations.

Now, using the pr i made to llm-openrouter, you can save an alias to a model that includes lots of model options. For examples, you can do llm openrouter save -m qwen3 -o online -o temperature 0, system "research prompt" --name qwen-researcher

And now, you can build a consortium where one member is an online research specialist. You could make another uses JSON mode for entity extraction, and a third which writes a blind draft. The arbiter would then make use of all that and synthesize a good answer.

Have you tried n8n? It allows you to build flows like that - you can run the community version in a Docker container within a few minutes and share the configurations for the flows you have built very easily.

Ive felt this way when using chatgpt for a simple search. Stuff that google could handle but would just be slower, mostly from me having to manually filter.

Sometimes its the easiest way to complete a very small task but the cost difference on the backend has to be pretty damn large. The user inevitably ends up not caring whatsoever. Its just not real to them.

I caught infra people saying that's pretty much the only bottleneck in the data center right now, power and cooling. We know the AI needs to run against itself continuously, and that's just a fact.

Maybe we should assign them a practical task, like making paperclips.

Scaled up and spread out - this probably gets you pretty close to consciousness(?)

Conway's game of life, but instead of colored squares with rules, they're LLM's with some kind of weighting - all chattering back and forth with one another - bubbling up somehow to cause speach/action

What you just said is what I tried and failed to say ten minutes ago!

https://news.ycombinator.com/item?id=43835798

I once attended a talk a year ago where a techlead did just that - they had AI agents that ran a scrum team with different roles, each agent's prompt was to disagree with everyone else (or be highly critical) and present their own point of view, and then an arbiter would make the final decision. They claimed it worked for them.

Maybe. Humans form teams for a reason. Yes there are different exepriences and points of view in a human (vs. Not so much in LLM), but sometimes a different hat it all it takes. E.g. Code reviewer vs. Coder.

Seems likely to me. As long as adding more appears to help people will do it

Presumably there is some point where it levels out. And no doubt there will be a committee of AIs to determine said point.

Cause we wouldn’t want to boil the ocean…

I think the names were switched up.