Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

LLMs get lost in multi-turn conversation

arxiv.org

Post reply on HN
Print view View original

171–180 of 272 posts

Re: LLMs get lost in multi-turn conversation

#171

Post by JustFinishedBSG »

dkdbejwi383 wrote: post #161
Sharlin wrote: post #20

Earlier quoted context omitted.

This is a tired semantic argument that does not bring any insight into the discussion. A token-predictor could still be trained to predict the tokens “I’m not sure what you mean because of points x, y, and z; could you elaborate?”

How would an LLM “know” when it isn’t sure? Their baseline for truth is competent text, they don’t have a baseline for truth based on observed reality. That’s why they can be “tricked” into things like “Mr Bean is the president of the USA”

It would "know" the same way it "knows" anything else: The probability of the sequence "I don't know" would be higher than the probability of any other sequence.

Re: LLMs get lost in multi-turn conversation

#172

Post by btbuildem »

Sharlin wrote: post #7

Seems like this is an aspect of their well-known overconfidence and the inability to self-reflect and recognize they have to ask for more details because their priors are too low. If you look at the output of reasoning models, it’s clear that the idea of asking for clarification very rarely occurs to them – when they’re confused, it’s just endless speculation of what the user might have meant. This, of course, has ce…

> This, of course, has certain implications as to the wisdom of the idea of “replacing human programmers”

Ironically, working with a junior dev is a lot like this -- setting them on a task, then coming back later with dogs and flashlights to retrieve them from the deep woods they've inevitably lost themselves in by just forging ahead, making assumptions, and asking no questions.

Re: LLMs get lost in multi-turn conversation

#173

Post by jimmySixDOF »

Benjammer wrote: post #2

It's nice to see a paper that confirms what anyone who has practiced using LLM tools already knows very well, heuristically. Keeping your context clean matters, "conversations" are only a construct of product interfaces, they hurt the quality of responses from the LLM itself, and once your context is "poisoned" it will not recover, you need to start fresh with a new chat.

>"conversations" are only a construct of product interfaces

This seems to be in flux now due to RL training on multiturn eval datasets so while the context window is evergreen every time, there will be some bias towards interpreting each prompt as part of a longer conversation. Mutliturn post training is not scaled out yet in public but I think it may be the way to keep on the 'double time spent on goal every 7 months curve'

Re: LLMs get lost in multi-turn conversation

#174

Post by bentt »

Benjammer wrote: post #2

It's nice to see a paper that confirms what anyone who has practiced using LLM tools already knows very well, heuristically. Keeping your context clean matters, "conversations" are only a construct of product interfaces, they hurt the quality of responses from the LLM itself, and once your context is "poisoned" it will not recover, you need to start fresh with a new chat.

Yes even when coding and not conversing I often start new conversations where I take the current code and explain it new. This often gives better results than hammering on one conversation.

This feels like something that can be fixed with manual instructions which prompt the model to summarize and forget. This might even map appropriately to human psychology. Working Memory vs Narrative/Episodic Memory.

Re: LLMs get lost in multi-turn conversation

#175

Post by lolinder »

dragonwriter wrote: post #106
CobrastanJorji wrote: post #62

Earlier quoted context omitted.

An interesting little example of this problem is initial prompting, which is effectively just a permanent, hidden context that can't be cleared. On Twitter right now, the "Grok" bot has recently begun frequently mentioning "White Genocide," which is, y'know, odd. This is almost certainly because someone recently adjusted its prompt to tell it what its views on white genocide are meant to be, which for a perfect chatb…

> This is almost certainly because someone recently adjusted its prompt to tell it what its views on white genocide are meant to be Well, someone did something to it; whether it was training, feature boosting the way Golden Gate Claude [0] was done, adjusting the system prompt, or assuring that it's internet search for contextual information would always return material about that, or some combination of those, is ne…

Source [0]. The examples look pretty clearly like they stuck it in the context window, not trained it in. It consistently seems to structure the replies as though the user they're replying to is the one who brought up white genocide in South Africa, and it responds the way that LLMs often respond to such topics: saying that it's controversial and giving both perspectives. That's not behavior I would expect if they had done the Golden Gate Claude method, which inserted the Golden Gate Bridge a bit more fluidly into the conversation rather than seeming to address a phantom sentence that the user supposedly said.

Also, let's be honest, in a Musk company they're going to have taking the shortest possible route to accomplishing what he wanted them to.

[0] https://www.cnn.com/2025/05/14/business/grok-ai-chatbot-repl...

Re: LLMs get lost in multi-turn conversation

#176

Post by arnaudsm »

airylizard wrote: post #49

Why I came up with TSCE(Two-Step Contextual Enrichment). +30pp uplift when using GPT-35-turbo on a mix of 300 tasks. Free open framework, check the repo try it yourself https://github.com/AutomationOptimization/tsce_demo I tested this another 300 times with gpt-4.1 to remove those obtrusive "em-dashes" everyone hates. Tested a single-pass baseline vs TSCE, same exact instructions and prompt "Remove the em-dashes from…

That's a lot of kilo-watt-hours wasted for a find and replace operation.

Have you heard of text.replace("—", "-") ?

Re: LLMs get lost in multi-turn conversation

#177

Post by pseudocomposer »

Benjammer wrote: post #2

It's nice to see a paper that confirms what anyone who has practiced using LLM tools already knows very well, heuristically. Keeping your context clean matters, "conversations" are only a construct of product interfaces, they hurt the quality of responses from the LLM itself, and once your context is "poisoned" it will not recover, you need to start fresh with a new chat.

I mostly just use LLMs for autocomplete (not chat), but wouldn’t this be fixed by adding a “delete message” button/context option in LLM chat UIs?

If you delete the last message from the LLM (so now, you sent the last message), it would then generate a new response. (This would be particularly useful with high-temperature/more “randomly” configured LLMs.)

If you delete any other message, it just updates the LLM context for any future responses it sends (the real problem at hand, context cleanup).

I think seeing it work this way would also really help end users who think LLMs are “intelligent” to better understand that it’s just a big, complex autocomplete (and that’s still very useful).

Maybe this is standard already, or used in some LLM UI? If not, consider this comment as putting it in the public domain.

Now that I’m thinking about it, it seems like it might be practical to use “sub-contextual LLMs” to manage the context of your main LLM chat. Basically, if an LLM response in your chat/context is very long, you could ask the “sub-contextual LLM” to shorten/summarize that response, thus trimming down/cleaning the context for your overall conversation. (Also, more simply, an “edit message” button could do the same, just with you, the human, editing the context instead of an LLM…)

Re: LLMs get lost in multi-turn conversation

#179

Post by dr_dshiv »

pseudocomposer wrote: post #177
Benjammer wrote: post #2

It's nice to see a paper that confirms what anyone who has practiced using LLM tools already knows very well, heuristically. Keeping your context clean matters, "conversations" are only a construct of product interfaces, they hurt the quality of responses from the LLM itself, and once your context is "poisoned" it will not recover, you need to start fresh with a new chat.

I mostly just use LLMs for autocomplete (not chat), but wouldn’t this be fixed by adding a “delete message” button/context option in LLM chat UIs? If you delete the last message from the LLM (so now, you sent the last message), it would then generate a new response. (This would be particularly useful with high-temperature/more “randomly” configured LLMs.) If you delete any other message, it just updates the LLM conte…

This is how Claude’s UI used to work, in practice, where you could edit the context directly.

Re: LLMs get lost in multi-turn conversation

#180

Post by dr_dshiv »

Benjammer wrote: post #2

It's nice to see a paper that confirms what anyone who has practiced using LLM tools already knows very well, heuristically. Keeping your context clean matters, "conversations" are only a construct of product interfaces, they hurt the quality of responses from the LLM itself, and once your context is "poisoned" it will not recover, you need to start fresh with a new chat.

The #1 tip I teach is to make extensive use of the teeny-tiny mostly hidden “edit” button in ChatGPT and Claude. When you get a bad response, stop and edit to get a better one, rather than letting crap start to multiply crap.
Post reply on HN