Self-evolving scientific agents capable of conquering the hard tail of formal mathematics require Compositional Learning Behaviours (CLBs) – the capacity to ground and recombine novel symbolic structures in context, beyond mere recombination of prelearned atoms. We propose S2B-LM, an adaptation of the Symbolic Behaviour Benchmark that removes numerical processing as a confound and adds chain-of-thought scaffolding to elicit rather than merely probe latent CLB competency. Cross-evaluating ten Lean 4 theorem provers on CLB competency (adj-ZSCT) and miniF2F whole-proof performance, exact permutation tests establish a hierarchical necessity structure: search-heavy models cover the tractable bulk without detectable CLBs, yet every model breaking into the Olympiad-level tier (miniF2F >75%) is among the five highest CLB scorers (p=0.004). After ruling out model scale as a confound, our results show that CLB competency is necessary but not sufficient for the hard tail of formal mathematical verification.

Despite emerging research on Language Models (LM), few approaches analyse the invertibility of LMs. That is, given a LM and a desirable target output sequence of tokens, determining what input prompts would yield the target output remains an open problem. We formulate this problem as a classical gradient-based optimisation. First, we propose a simple algorithm to achieve end-to-end differentiability of a given (frozen) LM and then find optimised prompts via gradient descent. Our central insight is to view LMs as functions operating on sequences of distributions over tokens (rather than the traditional view as functions on sequences of tokens). Our experiments and ablations demonstrate that our DLM-powered inversion can reliably and efficiently optimise prompts of lengths 10 and 80 for targets of length 20, for several white-box LMs (out-of-the-box).

Inverse problems are central to a wide range of fields, including healthcare, climate science, and agriculture. They involve the estimation of inputs, typically via iterative optimization, to some known forward model so that it produces a desired outcome. Despite considerable development in the explainability and interpretability of forward models, the iterative optimization of inverse problems remains largely cryptic to domain experts. We propose a methodology to produce explanations, from traces produced by an optimizer, that are interpretable by humans at the abstraction of the domain. The central idea in our approach is to instrument a differentiable simulator so that it emits natural language events during its forward and backward passes. In a post-process, we use a Language Model to create an explanation from the list of events. We demonstrate the effectiveness of our approach with an illustrative optimization problem and an example involving the training of a neural network.

We present an interactive and explainable automated coaching assistant called CueTip for a variant of pool/billiards. CueTip’s novelty lies in its combination of three features: a natural-language interface, an ability to perform contextual, physics-aware reasoning, and that its explanations are rooted in a set of predetermined guidelines developed by domain experts. We instrument a physics simulator so that it generates event traces in natural language alongside traditional state traces. Event traces lend themselves to interpretation by language models, which serve as the interface to our assistant. We design and train a neural adaptor that decouples tactical choices made by CueTip from its interactivity and explainability allowing it to be reconfigured to mimic any pool playing agent. Our experiments show that CueTip enables contextual query-based assistance and explanations while maintaining the strength of the agent in terms of win rate (improving it in some situations). The explanations generated by CueTip are physically-aware and grounded in the expert rules and are therefore more reliable.

Instruction-following from prompts in Natural Languages (NLs) is an important benchmark for Human-AI collaboration. Training Embodied AI agents for instruction-following with Reinforcement Learning (RL) poses a strong exploration challenge. Previous works have shown that NL-based state abstractions can help address the exploitation versus exploration trade-off in RL. However, NLs descriptions are not always readily available and are expensive to collect. We therefore propose to use the Emergent Communication paradigm, where artificial agents are free to learn an emergent language (EL) via referential games, to bridge this gap. ELs constitute cheap and readily-available abstractions, as they are the result of an unsupervised learning approach. In this paper, we investigate (i) how EL-based state abstractions compare to NL-based ones for RL in hard-exploration, procedurally-generated environments, and (ii) how properties of the referential games used to learn ELs impact the quality of the RL exploration and learning. Results indicate that the EL-guided agent, namely EReLELA, achieves similar performance as its NL-based counterparts without its limitations. Our work shows that Embodied RL agents can leverage unsupervised emergent abstractions to greatly improve their exploration skills in sparse reward settings, thus opening new research avenues between Embodied AI and Emergent Communication.

Natural languages are powerful tools wielded by human beings to communicate information. Among their desirable properties, compositionality has been the main focus in the context of referential games and variants, as it promises to enable greater systematicity to the agents which would wield it. The concept of disentanglement has been shown to be of paramount importance to learned representations that generalise well in deep learning, and is thought to be a necessary condition to enable systematicity. Thus, this paper investigates how do compositionality at the level of the emerging languages, disentanglement at the level of the learned representations, and systematicity relate to each other in the context of visual referential games. Firstly, we find that visual referential games that are based on the Obverter architecture outperforms state-of-the-art unsupervised learning approach in terms of many major disentanglement metrics. Secondly, we expand the previously proposed Positional Disentanglement (PosDis) metric for compositionality to (re-)incorporate some concerns pertaining to informativeness and completeness features found in the Mutual Information Gap (MIG) disentanglement metric it stems from. This extension allows for further discrimination between the different kind of compositional languages that emerge in the context of Obverter-based referential games, in a way that neither the referential game accuracy nor previous metrics were able to capture. Finally we investigate whether the resulting (emergent) systematicity, as measured by zero-shot compositional learning tests, correlates with any of the disentanglement and compositionality metrics proposed so far. Throughout the training process, statically significant correlation coefficients can be found both positive and negative depending on the moment of the measure.

Natural language instruction following is paramount to enable collaboration between artificial agents and human beings. Natural language-conditioned reinforcement learning (RL) agents have shown how natural languages' properties, such as compositionality, can provide a strong inductive bias to learn complex policies. Previous architectures like HIGhER combine the benefit of language-conditioning with Hindsight Experience Replay (HER) to deal with sparse rewards environments. Yet, like HER, HIGhER relies on an oracle predicate function to provide a feedback signal highlighting which linguistic description is valid for which state. This reliance on an oracle limits its application. Additionally, HIGhER only leverages the linguistic information contained in successful RL trajectories, thus hurting its final performance and data-efficiency. Without early successful trajectories, HIGhER is no better than DQN upon which it is built. In this paper, we propose the Emergent Textual Hindsight Experience Replay (ETHER) agent, which builds on HIGhER and addresses both of its limitations by means of (i) a discriminative visual referential game, commonly studied in the subfield of Emergent Communication (EC), used here as an unsupervised auxiliary task and (ii) a semantic grounding scheme to align the emergent language with the natural language of the instruction-following benchmark. We show that the referential game’s agents make an artificial language emerge that is aligned with the natural-like language used to describe goals in the BabyAI benchmark and that it is expressive enough so as to also describe unsuccessful RL trajectories and thus provide feedback to the RL agent to leverage the linguistic, structured information contained in all trajectories. Our work shows that EC is a viable unsupervised auxiliary task for RL and provides missing pieces to make HER more widely applicable.

Human beings use compositionality to generalise from past experiences to novel experiences. We assume a separation of our experiences into fundamental atomic components that can be recombined in novel ways to support our ability to engage with novel experiences. We frame this as the ability to learn to generalise compositionally, and we will refer to behaviours making use of this ability as compositional learning behaviours (CLBs). A central problem to learning CLBs is the resolution of a binding problem (BP). While it is another feat of intelligence that human beings perform with ease, it is not the case for state-of-the-art artificial agents. Thus, in order to build artificial agents able to collaborate with human beings, we propose to develop a novel benchmark to investigate agents' abilities to exhibit CLBs by solving a domain-agnostic version of the BP. We take inspiration from the language emergence and grounding framework of referential games and propose a meta-learning extension of referential games, entitled Meta-Referential Games, and use this framework to build our benchmark, the Symbolic Behaviour Benchmark (S2B). We provide baseline results and error analysis showing that our benchmark is a compelling challenge that we hope will spur the research community towards developing more capable artificial agents.

The drivers of compositionality in artificial languages that emerge when two (or more) agents play a non-visual referential game has been previously investigated using approaches based on the REINFORCE algorithm and the (Neural) Iterated Learning Model. Following the more recent introduction of the extit{Straight-Through Gumbel-Softmax} (ST-GS) approach, this paper investigates to what extent the drivers of compositionality identified so far in the field apply in the ST-GS context and to what extent do they translate into (emergent) systematic generalisation abilities, when playing a visual referential game. Compositionality and the generalisation abilities of the emergent languages are assessed using topographic similarity and zero-shot compositional tests. Firstly, we provide evidence that the test-train split strategy significantly impacts the zero-shot compositional tests when dealing with visual stimuli, whilst it does not when dealing with symbolic ones. Secondly, empirical evidence shows that using the ST-GS approach with small batch sizes and an overcomplete communication channel improves compositionality in the emerging languages. Nevertheless, while shown robust with symbolic stimuli, the effect of the batch size is not so clear-cut when dealing with visual stimuli. Our results also show that not all overcomplete communication channels are created equal. Indeed, while increasing the maximum sentence length is found to be beneficial to further both compositionality and generalisation abilities, increasing the vocabulary size is found detrimental. Finally, a lack of correlation between the language compositionality at training-time and the agents' generalisation abilities is observed in the context of discriminative referential games with visual stimuli. This is similar to previous observations in the field using the generative variant with symbolic stimuli.

Natural languages are powerful tools wielded by human beings to communicate information and co-operate towards common goals. Their values lie in some main properties like compositionality, hierarchy and recurrent syntax, which computational linguists have been researching the emergence of in artificial languages induced by language games. Only relatively recently, the AI community has started to investigate language emergence and grounding working towards better human-machine interfaces. For instance, interactive/conversational AI assistants that are able to relate their vision to the ongoing conversation. This paper provides two contributions to this research field. Firstly, a nomenclature is proposed to understand the main initiatives in studying language emergence and grounding, accounting for the variations in assumptions and constraints. Secondly, a PyTorch based deep learning framework is introduced, entitled ReferentialGym, which is dedicated to furthering the exploration of language emergence and grounding. By providing baseline implementations of major algorithms and metrics, in addition to many different features and approaches, ReferentialGym attempts to ease the entry barrier to the field and provide the community with common implementations.