you you contrust the [n-plet] for such a state. ontology or something esle

In Graph you used triplets in HypoRag. What about if you want also to train the evolution of changeing in infomation in other dimentions. like time or context . in simple words what is a node path change in relation to the neigbords in time and this change.

"Can the statement 'LLMs can be highly receptive to external evidence' indicate that any method requiring input to the model for evaluating its memory is unreliable? Or is there a way to determine whether the influence of the input we provide to the model on the results of evaluating its memory can be considered negligible?"

As LLMs evolved, would hipporag become more effective?

If you have a large enough context, wouldn't a reasoning model figure the transitive relationships out just by generation of lot of tokens?

I have seen the mechanistic interpretation confirming hypothesis first time. Do you think there is a larger application for this?

Do you think LLM world model-based Planner can handle disjunctive goals and also conjunctive/ composition of goals?

Can Grokking help with the reversal curse? Knowing that A = B is the same as B = A might not be considered OOD...

Could the main reason that LLMs are receptive to new knowledge or knowledge editing be due to RLHF?

Will this hippo RAG be less useful as LLM evloves?

So the grokking happens when we overfit the training data many times?

Given past and current progress, how soon do you expect advanced agents to be able to automate most (say, >50%) of remote work? Does this seem like it's just around the corner? Or do we have more like 10 or so years until this happens?

Do standard anti-overfitting (i.e. regularization) techniques like dropout help or hurt this grokking of reasoning techniques?

are all training data just* some facts or there are also value-laden statements? How does LLM reasoning handle value-laden statements and fact/value distinction issues

are all training data still some facts or there are also value-laden statements? How does LLM reasoning handle value-laden statements and fact/value distinction issues

Given that the 'compositionality gap' does not decrease with scale, what alternative architectural changes or training paradigms (e.g., symbolic reasoning, knowledge graph integration, or curriculum learning) could help LLMs improve in multi-hop reasoning and implicit comparison tasks?

"Can the statement 'LLMs can be highly receptive to external evidence' indicate that any method requiring input to the model for evaluating its memory is unreliable? Or is there a way to determine whether the influence of the input we provide to the model on the results of evaluating its memory can be considered negligible?"

Since the model performs differently on comparison vs. composition tasks. Are there any intuitive reasons why comparison reasoning is harder/easier for Transformers?

I have two questions: (1) Can we predict which tasks or datasets will exhibit grokking? Are there specific properties that make a problem "grokkable"? (2) Are there other forms of generalization beyond grokking? How can we ensure that models generalize robustly to truly novel situations, not just variations of the training data?

I have two questions:

Could the main reason that LLMs are receptive to new knowledge or knowledge editing be due to RLHF?

how does the size of the LLM influence the reasoning capability?

best approach for episodic mentioning, and how to combine with current

Two parts of RAG at least - one is storage and one is retrieval. So, maybe two scaling questions - how much can we store, and how does retrieval scale with storage?

How do you see HippoRAG and the knowledge graph it builds in its "hippocampus" in comparison to "classical" knowledge graphs? (like Google Knowledge Graph, DBpedia, Wikidata)

will the knowledge graph be updated given new information? if yes, how?

How do you distinguish between episodic and semantic memory in the design of the long-term memory

How big can RAG store get.. in terms of size?

How did you ensure that each triplet extracted was high-quality? I have faced a lot of low-information triplets in my Knowledge Graphs in the past, so I was wondering what techniques you used to optimize the graphs.

knowledge model entity/editing (injecting other memories) - does he mean prompt injection what hackers use to alter facts in the data?

hey