Temporal Logic in AI: From Linear Time to 8-Dimensions?

Ajouté : 2026-05-06

[00:00:00]Hello community. So great that you are back. Let's talk about temporal artificial intelligence. Now you see here we have here something that a human formulated. Let's say this is my construct here. Any is going to look at it is crush it is say hey I cannot use this human input. No I have to analyze here the temporal rewrite here the timestamps the time stamp dependency the time stamp duration. I have to reorder everything from the story from the query the human asked me. I have to change this completely and then I have here a review here of the time dependencies and then I can understand here the intent of the human and then I can provide an answer. And of course we have here a central artificial intelligence for all of this. And you might say why so simple? Well in my last video that was for our channel members only where I showed you that China invented here a temporal AI a complete new architecture I got a feedback though that was a little bit heavy. So today today we're going to enjoy it. We're going to say hey what is this simplest thing for temporal artificial intelligence and also I got some question back here when I tested the new myo 2.5 pro from shyomi and there was also the question hey why can we not handle this time dependencies in a simple matter like a skill MD file you know just give here the ei the instruction here a workflow a template and then it is just here a temporal instruction set why is this not Now you know temporal reasoning we do a temporal reasoning for quite some times and here I think from October 2024 the real interesting study here LLM can learn temporal reasoning here from Georgia Technology and Cisco research and they showed you you know what we do simply a texttotemporal graph transition so the eye itself trying to understand what does the human want from me builds a temporal graph from the complete story from the background and then does a temporal graph reasoning on the temporal graph itself. Or if you see this, this is from 2024, an event-based abduction learning for hardware time sensitive question answering. And here we have that we have an eventbased event extraction here and then an abductive reasoning. So we have here a time and a location interwoven graph structure. Then we try to separate here the complexities here with any system and then we have some simple lower complexity subsequences.

[00:02:32]But you know I think that all of those methods and trying to solve this here is and was to modularize the cognitive labor of an intelligence. This means to reduce the high complexity into some multiple lower complexity elements until even a small LLM is able to solve it.

[00:02:53]Now today's study claims to go the step further. So we have here April 27, 2026 adapt time enabling adaptive temporal reasoning in LLMs. And this is fun university, city university of Hong Kong, Tensson Jarvis Lab and Westlake University. And I say hey LM's ability to handle temporal information is rather limited. No and to address this limitation existing approaches often involve external tools. as I showed you, let's build a graph lefts have some numerical solver to do this and they argue well this leads to a poor generalization.

[00:03:32]So interestingly they say the harnessing that we built around the core of an LLM is not really so helpful. What about we use here the LLM itself and the authors address this let's call it an inefficiency by introducing here an autonomous adaptive temporal reasoning framework. Now at this point I ask myself what is what is so different about the temporal dimension if we have logical reasoning and we argue about a certain I don't know about mathematics or physics or chemistry or whatever temporal seems to be such a critical element here and they argue no instead of a uniform pipeline we adapt now a dynamic routing mechanism and they have the idea I would say standard to act have an LLM here as the central orchestration that acts as an internal cognitive planner to evaluate the structural complexity of the input context of this what I want from the LLM and the query I formulate the task I formulate for the LLM and then the LLM is now here this autonomous adaptive temporal reasoning framework implementation and I thought wow this is this sounds impressive so let's have a look so what they do is more or less have three models you I talked about the modularization of the cognitive labor either of a human being or of an AI and what they do is they reformulate it. So they decompose my complex question that I have or my task that I have here into sequential sub questions reduce the complexity. You know this is a standard procedure and then comes the I think the most intelligent part where they want to rewrite this. So they want to transform an implicit narrative text into an explicit chronological timeline.

[00:05:26]Uh separate here the timestamps, the timestamps duration. Put everything on a on a singular time axis and try to understand the temporal dependencies between the events and the person, the places, everything that's going on in our story and our task. And then of course we have to have no self-verification. another agent or the identical LLM looks back and does a self-verification and everything is good and everything is beautiful except yeah it depends of course on the complexity and the the solvability here of the LLM that acts now as a self-verification machine let's have a closer look they argue that this shifts now the paradigm that I told you as killmd file or maybe some yeah template structure from a static external tool dependent reasoning to a dynamic internally guided metacognitive process. I kind of like this idea because you remember we had this beautiful visualization where we had this huge AI harness this flat ring around here the central core our LLM and in this harness we had everything we had our graph rag our classical rag our database everything yeah and now they say yeah but let's bring it back to the LLM let's do not put it somewhere in a memory file or anything else let's try that the LLM can understand it itself. So I have to smile. They now argue this has the benefit of a toolfree autonomy.

[00:07:01]So they now have the goal to eliminate tool use bring back everything to the primary intelligence of an LLM the core of an agent. And they say by eliminating the reliance on external deterministic interpreters like a python code execution for the temporal alignment the researcher or any researcher can deploy models in isolated environment without designing complex brittle APIs for tool use and therefore so on. So the idea is simple. We just have to make here the formulation the task so simple. We have to break it down in so little micro tasks that then the LLM doesn't need any tool use or any advanced tool whatever it can do it here with its internal reasoning.

[00:07:50]Now I might for a second interrupt here this presentation and in my own words say but you know what you can apply this to everything. Time is just an axis.

[00:07:59]Time is just a dimension here. So if you have something else, let's say we have a spatial complexity or a particular causal complexity or a legal reasoning task, you can do more or less the same with this methodology. No, just take a step back and look at this. No, the implementation requires initializing a meta prompt. This is our planner EI that executes some conditional logic based on semantic density of the input context.

[00:08:28]And then the researcher just define here a discrete maybe a one-dimensional action space where we have this reformulate rewrite and review. I'm going to show you an example and then prompt a planet to generate a specific operational trajectory prior to the final generation.

[00:08:46]I mean this idea you can apply to almost everything. This is not limited to a temporal reasoning structure. do this here for for legal reasoning or whatever you like. And you see this dimensions, we already talked about this here when we talked about skills here and we talked about sparse autoenccorder and the highing value skills that define here the perfect axis and the perfect minimal representation along the tangent vectors here. You can do the same here.

[00:09:17]Just select one of these. You have a 4,000 dimensional space. Select one, two, three, four time dimensions and go on this run on this linear dimensions or a hyperbolic dimension whatever you prefer.

[00:09:33]So you see the idea has some general ability but let's come back to the paper and the algorithm here formalizes here this adapt time operates here over a tupil and we have C is our context Q is our question and R is of course A is the answer and the framework is governed here by this algorithm you see at the right hand side which but look at it this is just a state machine orchestrated by an LLM planner this is an old old friend of us this is nothing special this is just a particular sequence that we run through. But now it gets interesting because if you look here at the result and they did a lot of testing here and they compared to the proportion of each operation in different data set.

[00:10:14]If the different data set have a different complexity and dependency and so on and here you see for four different data sets the amount of reformulate can be extreme different given the particular data set.

[00:10:28]Interesting a rewrite is happening in 88 86% of the cases and then the review is also here highly dependent on the particular task and on the particular data set complexity and the orders tell us for a clearly structured input a time Q&A.

[00:10:45]The planner frequently selects the one or reformulate here to isolate sube events. Invests a lot of energy and time here just to crystallize out each sub element, each subevent and all the dependencies between them. But for some high ambiguity data set, it relies here almost nothing on reformulate but then heavily on rewrite and heavily on review. No, because these are kind of open-ended high complexity tasks. So, okay, it is not a skill MD file because if you want the LLM decides here, the planning LLM here on the intensity of the three steps here, how much time how much effort it invests in reformulate or in review where I would say rewrite is almost all the time here almost max.

[00:11:37]Okay.

[00:11:39]Now they say the rewrite module normalizes here of the natural language input contain implicit temporal markers.

[00:11:46]No like during his presidency. So when did it started? When did it end? What are the exact date? Just rewrite everything with all the temporal information that we have. Build new sentences, build new sequences where you have the maximum temporal information, even if it's redundant, inserted into the sentences. And the rewrite model normalizes all of this during his pregnancy in definition into a monotonically ordered explicitly temporal representation.

[00:12:19]So what is now really interested that the text to time normalization this rewrite is crucial. They did a lot of ablation test and if they leave out this rewrite and you see this is important. Yeah.

[00:12:32]then the system is not able to really understand my task. What I want from this system and they say extracting here some scattered implicit chronological data and mapping it to explicit calendar anchor timelines significantly bounds the hallucination probabilities during the multihop reasoning. So the more information, the more temporal information you provide, the very little nitty-gritty detail on on some temporal information provide it to the LLM so that it is absolutely clear that the LLM is not missing data, is not missing relations and therefore starts to produce hallucination.

[00:13:11]But what is really interesting and I had to smile that the inherent reasoning capabilities of our LLM of the core of our agent surpasses now an agent with external tool use and this I found absolutely fascinating. However, I have to say look at the or the kind of state-of-the-art foundational models that they go deepseek version 3. Okay, when they wrote the study, version 4 was just was not out yet and they go with a Q and 38B.

[00:13:41]And they argue that those model contain the latent parameter necessary for complex chronological execution, rendering therefore the previous reliance on Python dictionaries or handcrafted matching functions or some skill MD files for temporal reasoning obsolete.

[00:14:00]You have a lot of uh empirical data in the study. Have a look at the study. It is quite interesting to read. I would say it is not groundbreaking. It is not that you see a firework of innovation.

[00:14:14]It is a methodology an implementation here of an idea that we have here and yeah everything again is fed back into the LLM. But the question that remains open how do we continue to train the ability of the LLM to go to the next level of complexity? I think this is more or less not yet answered.

[00:14:38]Okay. Short summary, we have here the reasoning pipeline. So you have here the story that is given, the context that is given. Then we have a particular question as you see here. And then we have your reasoning pipeline. Now it starts here with an LLM planner that reformulates everything. And you have here a question decomposition. And if you have this particular study from Terren Cooper here, you have now the sub questions and sub uh complexities that the LLM is now trying to answer. Then again we have the LLM plan as the central intelligence no that starts now to rewrite everything on a chronological timeline insert all the temporal information that is available. So from the 1960s to the 1970s this person was at this location for this particular issue and then from the 1970s to the 1980s and you get the idea with the rewrite. Then again we have the planner and if the planner checks again if it has all the information or we have to loop back and if not we start the review process and yeah you're familiar with the review process. Now just to show you here the template I think this is so simple. Look at this. It comes back the skill markdown file back. Now the reformulate template break the question down into several simple sub question and answer each sub question. and then return to the final answer or the rewrite template here. In the context of story and question, generate a timeline for what the question concerns and answer each sub question. Then return to the final answer or for the review template. In the context of story and question, after obtaining the answer, given the support sentences in the original story and check if the answer is correct. If yes, return to the answer again. If no, think again and return to the right answer. And here you have the adapt time template now in its whole beauty. And now tell me, is this not a skill MD file? I mean, even a very simple, yes, of course. But is this really such a a paradigm shift? Maybe I don't see it. So if you see it and you disagree with me, please leave a comment. If this is groundbreaking, I have not yet noticed it. Let me show you an example.

[00:16:53]They give you here, I think it's a panic C here, an example. from Mikail and this is here a a case study regarding the education here of Mikail here in January 1736. Now the information we have he studied in Moscow from 1730 to 1735 briefly attended the Kiev Academy in 1735 and in 1736 he was awarded a scholarship at a St. Petersburg Academy.

[00:17:16]Now the orus show us that a traditional chain of sort but please notice they go with tiny little 7B and 8B models and even here a llama three model. I would not regard this as state-ofthe-art LM but okay the artists decided to go with this to tell us that traditional chain of sword fails because the text does not explicitly state his location in the exact month of January 1736.

[00:17:43]I would argue if you take here GBD 5.5 it is able to figure out this questionnaire but okay we go with the authors and they argue now that their new methodology here with the rewrites and the review is now explicitly a sequence building here the sequence so from 7030 to 1735 Moscow from 70 35 just KF and then 1736 St. Petersburg checks here that 1736 he was awarded scholarship and therefore the key of period was strictly bounded to 1735.

[00:18:20]H okay. If you have enough temporal information in the complexity of the context given to you or the parametric knowledge of the LLM has all the temporal details of the most important event in the history of human history.

[00:18:41]Well maybe um yeah okay results here we have it now. So we have here our if you want here in this pink underline here they go with your llama 38B with adapt time then a Q1 38B adapt time and a deepse version three adapt time and here you have all the numerical values and you see yes beautiful everything is here better 5.5% 4.3 percentage points great but just think about the limitation I do appreciate that you do not want to outsource here the intelligence and the temporal logic to the hardness of the eye. I think this is great. But if now everything depends just on u stoastic LLM just be aware that the reliance on the LLM intrinsic probability distribution for the planner function introduces therefore a degree of a stoastic instability in the system. Huh?

[00:19:40]And therefore the LLM planner behavior may exhibit absolute randomness across all the different instances plus hallucination. Let's be honest. Yeah.

[00:19:50]And therefore also the orchestras indicate and have to say this is beautiful that they indicate this uh it would benefit from some reinforcement learning from human feedback from some alignment. Yeah. that we give them some hint how a human would prefer here a temporal alignment thereby stabilizing here the conditional probabilities without requiring some rigid rulebased horistic that we have to introduce yeah so hard-coded rule-based juristic I'm not a fan of this but you see we are we are heavily in the development process of all of this and yeah by forcing now the model to unroll here the quant chronological sequent into really explicitly defined discrete states. Adapt time integrates here with really distinct time points into the reasoning process including here all the logical dependencies. But you know if you look this at a multi-dimensional perspective the temporal dimension you can see this here just this one dimension no in spacetime or if you have 4,000 dimension yeah define a dimension as a temporal dimension for the temporal development of your complete system dynamics. No this is not explicitly temporal but in anyway I wanted to show you this.

[00:21:09]I think it's nice. Another point of critique would be that I would argue that the world model is missing in this particular perspective. Yeah, because our central EI planner no the knowledge it has a parametric knowledge but the understanding of a physical world of some other world or if you go with some historical data need some rack connectivity maybe some graph rack or hyperre connectivity here to all the libraries in the world to really understand each and every detail here in the Russian language in the English language in the whatever so you see all the the the little details are missing here and I think we have to add this but anyway idea I have a human query here and I'm absolutely unlogic and inconsistent and just a perfect chaos here in my definition of the task that I want the eye to perform. Yeah, it looks at here this chaos here. Says my goodness, what a complex data streams starts to reformulate have a complete data decomposition of the complexity of my intent and of my specific task is ordering. Now as you see here on the timeline here all the information all the context I provide to the eye all the dependencies that I mention maybe I also mention here the explicit dependencies is building here this timeline here from t0 to t + 20 is looking here at the temporal sequencing with the rewrite at the sequential alignment then of the events of the person of the location of the whatever is happening here and then having of course here uh verification here of this of is this really the correct temporal line that we want to examine that we want to reason about.

[00:22:58]Just think about my example of an elevator starting at floor zero and going to floor 50. all the dependencies, all the button presses that send you back and forth, maybe also in time and half here, all the other local dependencies and energy optimization or the token minimization and so on. And then come to a solution and then feed it back here and enable here maybe your central temporal hub to further understand here my human query. Maybe build an optimized temporal line, have an even better understanding. go to a world m have some graph rack get some additional external data from the internet of whatever data you have available and then build here this temporal AI model.

[00:23:48]Okay, this is it for today. I hope you had a little bit fun. I told you not one mathematical formula. It is simple as hell. It is beautiful. It is easy to explain. I hope you enjoyed it. You had a little bit of fun. It would be great to see you in my next