As fine-tuning becomes the dominant paradigm for improving large language models (LLMs), understanding what changes during this process is increasingly important.Traditional benchmarking often fails to explain why one model outperforms another.In this work, we use model diffing, a mechanistic interpretability approach, to analyze the specific capability differences between Gemma-2-9b-it and a SimPO-enhanced variant.Using crosscoders, we identify and categorize latent representations that differentiate the two models.We find that SimPO acquired latent concepts predominantly enhance safety mechanisms (+32.8%),multilingual capabilities (+43.8%), and instruction-following (+151.7%),while its additional training also reduces emphasis on model self-reference (-44.1%) and hallucination management (-68.5%).Our analysis shows that model diffing can yield fine-grained insights beyond leaderboard metrics, attributing performance gaps to concrete mechanistic capabilities.This approach offers a transparent and targeted framework for comparing LLMs.

Despite their remarkable ability to capture linguistic nuances across diverse languages, questions persist regarding the degree of alignment between languages in multilingual embeddings. Drawing inspiration from research on high-dimensional representations in neural language models, we employ clustering to uncover latent concepts within multilingual models. Our analysis focuses on quantifying the alignment and overlap of these concepts across various languages within the latent space. To this end, we introduce two metrics CALIGN and COLAP aimed at quantifying these aspects, enabling a deeper exploration of multilingual embeddings. Our study encompasses three multilingual models (mT5, mBERT, and XLM-R) and three downstream tasks (Machine Translation, Named Entity Recognition, and Sentiment Analysis). Key findings from our analysis include: i) deeper layers in the network demonstrate increased cross-lingual alignment due to the presence of language-agnostic concepts, ii) fine-tuning of the models enhances alignment within the latent space, and iii) such task-specific calibration helps in explaining the emergence of zero-shot capabilities in the models.

Ahmed Abdelali, Hamdy Mubarak, Shammur Chowdhury, Maram Hasanain, Basel Mousi, Sabri Boughorbel, Samir Abdaljalil, Yassine El Kheir, Daniel Izham, Fahim Dalvi, Majd Hawasly, Nizi Nazar, Youssef Elshahawy, Ahmed Ali, Nadir Durrani, Natasa Milic-Frayling, Firoj Alam. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Despite the revolution caused by deep NLP models, they remain black boxes, necessitating research to understand their decision-making processes.A recent work by Dalvi et al. (2022) carried out representation analysis through the lens of clustering latent spaces within pretrained models (PLMs), but that approach is limited to small scale due to the high cost of running Agglomerative hierarchical clustering.This paper studies clustering algorithms in order to scale the discovery of encoded concepts in PLM representations to larger datasets and models.We propose metrics for assessing the quality of discovered latent concepts and use them to compare the studied clustering algorithms.We found that K-Means-based concept discovery significantly enhances efficiency while maintaining the quality of the obtained concepts.Furthermore, we demonstrate the practicality of this newfound efficiency by scaling latent concept discovery to LLMs and phrasal concepts.1

Autonomous vehicles use a variety of sensors and machine-learned models to predict the behavior of surrounding road users. Most of the machine-learned models in the literature focus on quantitative error metrics like the root mean square error (RMSE) to learn and report their models’ capabilities. This focus on quantitative error metrics tends to ignore the more important behavioral aspect of the models, raising the question of whether these models really predict human-like behavior. Thus, we propose to analyze the output of machine-learned models much like we would analyze human data in conventional behavioral research. We introduce quantitative metrics to demonstrate presence of three different behavioral phenomena in a naturalistic highway driving dataset: 1) The kinematics-dependence of who passes a merging point first 2) Lane change by an on-highway vehicle to accommodate an on-ramp vehicle 3) Lane changes by vehicles on the highway to avoid lead vehicle conflicts. Then, we analyze the behavior of three machine-learned models using the same metrics. Even though the models’ RMSE value differed, all the models captured the kinematic-dependent merging behavior but struggled at varying degrees to capture the more nuanced courtesy lane change and highway lane change behavior. Additionally, the collision aversion analysis during lane changes showed that the models struggled to capture the physical aspect of human driving: leaving adequate gap between the vehicles. Thus, our analysis highlighted the inadequacy of simple quantitative metrics and the need to take a broader behavioral perspective when analyzing machine-learned models of human driving predictions.

Recent road trials have shown that guaranteeing the safety of driving decisions is essential for the wider adoption of autonomous vehicle technology. One promising direction is to pose safety requirements as planning constraints in nonlinear, nonconvex optimization problems of motion synthesis. However, many implementations of this approach are limited by uncertain convergence and local optimality of the solutions achieved, affecting overall robustness. To improve upon these issues, we propose a novel two-stage optimization framework: In the first stage, we find a solution to a mixed-integer linear programming (MILP) formulation of the motion synthesis problem, the output of which initializes a second nonlinear programming (NLP) stage. The MILP stage enforces hard constraints of safety and road rule compliance generating a solution in the right subspace, while the NLP stage refines the solution within the safety bounds for feasibility and smoothness. We demonstrate the effectiveness of our framework via simulated experiments of complex urban driving scenarios, outperforming a state-of-the-art baseline in metrics of convergence, comfort, and progress.

Achieving a proper balance between planning quality, safety and efficiency is a major challenge for autonomous driving. Optimisation-based motion planners are capable of producing safe, smooth and comfortable plans, but often at the cost of runtime efficiency. On the other hand, naïvely deploying trajectories produced by efficient-to-run deep imitation learning approaches might risk compromising safety. In this paper, we present PILOT a planning framework that comprises an imitation neural network followed by an efficient optimiser that actively rectifies the network’s plan, guaranteeing fulfilment of safety and comfort requirements. The objective of the efficient optimiser is the same as the objective of an expensive-to-run optimisation-based planning system that the neural network is trained offline to imitate. This efficient optimiser provides a key layer of online protection from learning failures or deficiency in out-of-distribution situations that might compromise safety or comfort. Using a state-of-the-art, runtime-intensive optimisation-based method as the expert, we demonstrate in simulated autonomous driving experiments in CARLA that PILOT achieves a seven-fold reduction in runtime when compared to the expert it imitates without sacrificing planning quality.

Effective recognition of complex long-term activities is becoming an increasingly important task in artificial intelligence. In this paper, we propose a novel approach for building models of complex long-term activities. First, we automatically learn the hierarchical structure of activities by learning about the 'parent-child' relation of activity components from a video using the variability in annotations acquired using multiple annotators. This variability allows for extracting the inherent hierarchical structure of the activity in a video. We consolidate hierarchical structures of the same activity from different videos into a unified stochastic grammar describing the overall activity. We then describe an inference mechanism to interpret new instances of activities. We use three datasets, which have been annotated by multiple annotators, of daily activity videos to demonstrate the effectiveness of our system.

With the recent proliferation of human-oriented robotic applications in domestic and industrial scenarios, it is vital for robots to continually learn about their environments and about the humans they share their environments with. In this paper, we present a novel, online, incremental framework for unsupervised symbol grounding in real-world, human environments for autonomous robots. We demonstrate the flexibility of the framework by learning about colours, people names, usable objects and simple human activities, integrating state-of-the-art object segmentation, pose estimation, activity analysis along with a number of sensory input encodings into a continual learning framework. Natural language is grounded to the learned concepts, enabling the robot to communicate in a human-understandable way. We show, using a challenging real-world dataset of human activities as perceived by a mobile robot, that our framework is able to extract useful concepts, ground natural language descriptions to them, and, as a proof-of-concept, generate simple sentences from templates to describe people and the activities they are engaged in.

Many applications require autonomous agents to achieve quick responses to task instances drawn from a rich family of qualitatively-related tasks. We address the setting where the tasks share a state-action space and have the same qualitative objective but differ in dynamics. We adopt a transfer learning approach where common structure in previously-learnt policies, in the form of shared subtasks, is exploited to accelerate learning in subsequent ones. We use a probabilistic mixture model to describe regions in state space which are common to successful trajectories in different instances. Then, we extract policy fragments from previously-learnt policies that are specialised to these regions. These policy fragments are options, whose initiation and termination sets are automatically extracted from data by the mixture model. In novel task instances, these options are used in an SMDP learning process and option learning repeats over the resulting policy library. The utility of this method is demonstrated through experiments in a standard navigation environment and then in the RoboCup simulated soccer domain with opponent teams of different skill.