Model Editing with Graph-Based External Memory

Large language models (LLMs) have revolutionized natural language processing, yet their practical utility is often limited by persistent issues of hallucinations and outdated parametric knowledge. Although post-training model editing offers a pathway for dynamic updates, existing methods frequently suffer from overfitting and catastrophic forgetting. To tackle these challenges, we propose a novel framework that leverages hyperbolic geometry and graph neural networks for precise and stable model edits. We introduce HYPE (HYperbolic Parameter Editing), which comprises three key components: (i) Hyperbolic Graph Construction, which uses Poincar\'e embeddings to represent knowledge triples in hyperbolic space, preserving hierarchical relationships and preventing unintended side effects by ensuring that edits to parent concepts do not inadvertently affect child concepts; (ii) M\"obius-Transformed Updates, which apply hyperbolic addition to propagate edits while maintaining structural consistency within the hyperbolic manifold, unlike conventional Euclidean updates that distort relational distances; and (iii) Dual Stabilization, which combines gradient masking and periodic GNN parameter resetting to prevent catastrophic forgetting by focusing updates on critical parameters and preserving long-term knowledge. Experiments on CounterFact, CounterFact+, and MQuAKE with GPT-J and GPT2-XL demonstrate that HYPE significantly enhances edit stability, factual accuracy, and multi-hop reasoning.

Exploring the Innovation Opportunities for Pre-trained Models

Innovators transform the world by understanding where services are successfully meeting customers' needs and then using this knowledge to identify failsafe opportunities for innovation. Pre-trained models have changed the AI innovation landscape, making it faster and easier to create new AI products and services. Understanding where pre-trained models are successful is critical for supporting AI innovation. Unfortunately, the hype cycle surrounding pre-trained models makes it hard to know where AI can really be successful. To address this, we investigated pre-trained model applications developed by HCI researchers as a proxy for commercially successful applications. The research applications demonstrate technical capabilities, address real user needs, and avoid ethical challenges. Using an artifact analysis approach, we categorized capabilities, opportunity domains, data types, and emerging interaction design patterns, uncovering some of the opportunity space for innovation with pre-trained models.

RL in Name Only? Analyzing the Structural Assumptions in RL post-training for LLMs

Reinforcement learning-based post-training of large language models (LLMs) has recently gained attention, particularly following the release of DeepSeek R1, which applied GRPO for fine-tuning. Amid the growing hype around improved reasoning abilities attributed to RL post-training, we critically examine the formulation and assumptions underlying these methods. We start by highlighting the popular structural assumptions made in modeling LLM training as a Markov Decision Process (MDP), and show how they lead to a degenerate MDP that doesn't quite need the RL/GRPO apparatus. The two critical structural assumptions include (1) making the MDP states be just a concatenation of the actions-with states becoming the context window and the actions becoming the tokens in LLMs and (2) splitting the reward of a state-action trajectory uniformly across the trajectory. Through a comprehensive analysis, we demonstrate that these simplifying assumptions make the approach effectively equivalent to an outcome-driven supervised learning. Our experiments on benchmarks including GSM8K and Countdown using Qwen-2.5 base models show that iterative supervised fine-tuning, incorporating both positive and negative samples, achieves performance comparable to GRPO-based training. We will also argue that the structural assumptions indirectly incentivize the RL to generate longer sequences of intermediate tokens-which in turn feeds into the narrative of "RL generating longer thinking traces." While RL may well be a very useful technique for improving the reasoning abilities of LLMs, our analysis shows that the simplistic structural assumptions made in modeling the underlying MDP render the popular LLM RL frameworks and their interpretations questionable.

AI agents may be worth the hype but not the resources (yet): An initial exploration of machine translation quality and costs in three language pairs in the legal and news domains

Large language models (LLMs) and multi-agent orchestration are touted as the next leap in machine translation (MT), but their benefits relative to conventional neural MT (NMT) remain unclear. This paper offers an empirical reality check. We benchmark five paradigms, Google Translate (strong NMT baseline), GPT-4o (general-purpose LLM), o1-preview (reasoning-enhanced LLM), and two GPT-4o-powered agentic workflows (sequential three-stage and iterative refinement), on test data drawn from a legal contract and news prose in three English-source pairs: Spanish, Catalan and Turkish. Automatic evaluation is performed with COMET, BLEU, chrF2 and TER; human evaluation is conducted with expert ratings of adequacy and fluency; efficiency with total input-plus-output token counts mapped to April 2025 pricing. Automatic scores still favour the mature NMT system, which ranks first in seven of twelve metric-language combinations; o1-preview ties or places second in most remaining cases, while both multi-agent workflows trail. Human evaluation reverses part of this narrative: o1-preview produces the most adequate and fluent output in five of six comparisons, and the iterative agent edges ahead once, indicating that reasoning layers capture semantic nuance undervalued by surface metrics. Yet these qualitative gains carry steep costs. The sequential agent consumes roughly five times, and the iterative agent fifteen times, the tokens used by NMT or single-pass LLMs. We advocate multidimensional, cost-aware evaluation protocols and highlight research directions that could tip the balance: leaner coordination strategies, selective agent activation, and hybrid pipelines combining single-pass LLMs with targeted agent intervention.

Beyond the Hype: Embeddings vs. Prompting for Multiclass Classification Tasks

Are traditional classification approaches irrelevant in this era of AI hype? We show that there are multiclass classification problems where predictive models holistically outperform LLM prompt-based frameworks. Given text and images from home-service project descriptions provided by Thumbtack customers, we build embeddings-based softmax models that predict the professional category (e.g., handyman, bathroom remodeling) associated with each problem description. We then compare against prompts that ask state-of-the-art LLM models to solve the same problem. We find that the embeddings approach outperforms the best LLM prompts in terms of accuracy, calibration, latency, and financial cost. In particular, the embeddings approach has 49.5% higher accuracy than the prompting approach, and its superiority is consistent across text-only, image-only, and text-image problem descriptions. Furthermore, it yields well-calibrated probabilities, which we later use as confidence signals to provide contextualized user experience during deployment. On the contrary, prompting scores are overly uninformative. Finally, the embeddings approach is 14 and 81 times faster than prompting in processing images and text respectively, while under realistic deployment assumptions, it can be up to 10 times cheaper. Based on these results, we deployed a variation of the embeddings approach, and through A/B testing we observed performance consistent with our offline analysis. Our study shows that for multiclass classification problems that can leverage proprietary datasets, an embeddings-based approach may yield unequivocally better results. Hence, scientists, practitioners, engineers, and business leaders can use our study to go beyond the hype and consider appropriate predictive models for their classification use cases.

What the F*ck Is Artificial General Intelligence?

Artificial general intelligence (AGI) is an established field of research. Yet Melanie Mitchell and others have questioned if the term still has meaning. AGI has been subject to so much hype and speculation it has become something of a Rorschach test. Mitchell points out that the debate will only be settled through long term, scientific investigation. To that end here is a short, accessible and provocative overview of AGI. I compare definitions of intelligence, settling on intelligence in terms of adaptation and AGI as an artificial scientist. Taking my queue from Sutton's Bitter Lesson I describe two foundational tools used to build adaptive systems: search and approximation. I compare pros, cons, hybrids and architectures like o3, AlphaGo, AERA, NARS and Hyperon. I then discuss overall meta-approaches to making systems behave more intelligently. I divide them into scale-maxing, simp-maxing, w-maxing based on the Bitter Lesson, Ockham's and Bennett's Razors. These maximise resources, simplicity of form, and the weakness of constraints on functionality. I discuss examples including AIXI, the free energy principle and The Embiggening of language models. I conclude that though scale-maxed approximation dominates, AGI will be a fusion of tools and meta-approaches. The Embiggening was enabled by improvements in hardware. Now the bottlenecks are sample and energy efficiency.

Bridging Evolutionary Multiobjective Optimization and GPU Acceleration via Tensorization

Evolutionary multiobjective optimization (EMO) has made significant strides over the past two decades. However, as problem scales and complexities increase, traditional EMO algorithms face substantial performance limitations due to insufficient parallelism and scalability. While most work has focused on algorithm design to address these challenges, little attention has been given to hardware acceleration, thereby leaving a clear gap between EMO algorithms and advanced computing devices, such as GPUs. To bridge the gap, we propose to parallelize EMO algorithms on GPUs via the tensorization methodology. By employing tensorization, the data structures and operations of EMO algorithms are transformed into concise tensor representations, which seamlessly enables automatic utilization of GPU computing. We demonstrate the effectiveness of our approach by applying it to three representative EMO algorithms: NSGA-III, MOEA/D, and HypE. To comprehensively assess our methodology, we introduce a multiobjective robot control benchmark using a GPU-accelerated physics engine. Our experiments show that the tensorized EMO algorithms achieve speedups of up to 1113x compared to their CPU-based counterparts, while maintaining solution quality and effectively scaling population sizes to hundreds of thousands. Furthermore, the tensorized EMO algorithms efficiently tackle complex multiobjective robot control tasks, producing high-quality solutions with diverse behaviors. Source codes are available at https://github.com/EMI-Group/evomo.

Guidelines For The Choice Of The Baseline in XAI Attribution Methods

Given the broad adoption of artificial intelligence, it is essential to provide evidence that AI models are reliable, trustable, and fair. To this end, the emerging field of eXplainable AI develops techniques to probe such requirements, counterbalancing the hype pushing the pervasiveness of this technology. Among the many facets of this issue, this paper focuses on baseline attribution methods, aiming at deriving a feature attribution map at the network input relying on a "neutral" stimulus usually called "baseline". The choice of the baseline is crucial as it determines the explanation of the network behavior. In this framework, this paper has the twofold goal of shedding light on the implications of the choice of the baseline and providing a simple yet effective method for identifying the best baseline for the task. To achieve this, we propose a decision boundary sampling method, since the baseline, by definition, lies on the decision boundary, which naturally becomes the search domain. Experiments are performed on synthetic examples and validated relying on state-of-the-art methods. Despite being limited to the experimental scope, this contribution is relevant as it offers clear guidelines and a simple proxy for baseline selection, reducing ambiguity and enhancing deep models' reliability and trust.

The Imitation Game According To Turing

The current cycle of hype and anxiety concerning the benefits and risks to human society of Artificial Intelligence is fuelled, not only by the increasing use of generative AI and other AI tools by the general public, but also by claims made on behalf of such technology by popularizers and scientists. In particular, recent studies have claimed that Large Language Models (LLMs) can pass the Turing Test-a goal for AI since the 1950s-and therefore can "think". Large-scale impacts on society have been predicted as a result. Upon detailed examination, however, none of these studies has faithfully applied Turing's original instructions. Consequently, we conducted a rigorous Turing Test with GPT-4-Turbo that adhered closely to Turing's instructions for a three-player imitation game. We followed established scientific standards where Turing's instructions were ambiguous or missing. For example, we performed a Computer-Imitates-Human Game (CIHG) without constraining the time duration and conducted a Man-Imitates-Woman Game (MIWG) as a benchmark. All but one participant correctly identified the LLM, showing that one of today's most advanced LLMs is unable to pass a rigorous Turing Test. We conclude that recent extravagant claims for such models are unsupported, and do not warrant either optimism or concern about the social impact of thinking machines.

Beyond the Hype: Benchmarking LLM-Evolved Heuristics for Bin Packing

Coupling Large Language Models (LLMs) with Evolutionary Algorithms has recently shown significant promise as a technique to design new heuristics that outperform existing methods, particularly in the field of combinatorial optimisation. An escalating arms race is both rapidly producing new heuristics and improving the efficiency of the processes evolving them. However, driven by the desire to quickly demonstrate the superiority of new approaches, evaluation of the new heuristics produced for a specific domain is often cursory: testing on very few datasets in which instances all belong to a specific class from the domain, and on few instances per class. Taking bin-packing as an example, to the best of our knowledge we conduct the first rigorous benchmarking study of new LLM-generated heuristics, comparing them to well-known existing heuristics across a large suite of benchmark instances using three performance metrics. For each heuristic, we then evolve new instances won by the heuristic and perform an instance space analysis to understand where in the feature space each heuristic performs well. We show that most of the LLM heuristics do not generalise well when evaluated across a broad range of benchmarks in contrast to existing simple heuristics, and suggest that any gains from generating very specialist heuristics that only work in small areas of the instance space need to be weighed carefully against the considerable cost of generating these heuristics.