Róbert Csordás

Borja Ibarz, Vitaly Kurin, George Papamakarios et al. · deepmind

The cornerstone of neural algorithmic reasoning is the ability to solve algorithmic tasks, especially in a way that generalises out of distribution. While recent years have seen a surge in methodological improvements in this area, they mostly focused on building specialist models. Specialist models are capable of learning to neurally execute either only one algorithm or a collection of algorithms with identical control-flow backbone. Here, instead, we focus on constructing a generalist neural algorithmic learner -- a single graph neural network processor capable of learning to execute a wide range of algorithms, such as sorting, searching, dynamic programming, path-finding and geometry. We leverage the CLRS benchmark to empirically show that, much like recent successes in the domain of perception, generalist algorithmic learners can be built by "incorporating" knowledge. That is, it is possible to effectively learn algorithms in a multi-task manner, so long as we can learn to execute them well in a single-task regime. Motivated by this, we present a series of improvements to the input representation, training regime and processor architecture over CLRS, improving average single-task performance by over 20% from prior art. We then conduct a thorough ablation of multi-task learners leveraging these improvements. Our results demonstrate a generalist learner that effectively incorporates knowledge captured by specialist models.

Róbert Csordás, Kazuki Irie, Jürgen Schmidhuber

How to reduce compute and memory requirements of neural networks (NNs) without sacrificing performance? Many recent works use sparse Mixtures of Experts (MoEs) to build resource-efficient large language models (LMs). Here we introduce several novel perspectives on MoEs, presenting a general framework that unifies various methods to approximate two-layer NNs (e.g., feedforward blocks of Transformers), including product-key memories (PKMs). Leveraging insights from this framework, we propose methods to improve both MoEs and PKMs. Unlike prior work that compares MoEs with dense baselines under the compute-equal condition, our evaluation condition is parameter-equal, which is crucial to properly evaluate LMs. We show that our MoEs are competitive with the dense Transformer-XL on both the WikiText-103 and enwiki8 datasets at two different scales, while being much more resource efficient. This demonstrates that MoEs are relevant not only to extremely large LMs but also to any-scale resource-efficient LMs. Our code is public.

Róbert Csordás, Kazuki Irie, Jürgen Schmidhuber

Well-designed diagnostic tasks have played a key role in studying the failure of neural nets (NNs) to generalize systematically. Famous examples include SCAN and Compositional Table Lookup (CTL). Here we introduce CTL++, a new diagnostic dataset based on compositions of unary symbolic functions. While the original CTL is used to test length generalization or productivity, CTL++ is designed to test systematicity of NNs, that is, their capability to generalize to unseen compositions of known functions. CTL++ splits functions into groups and tests performance on group elements composed in a way not seen during training. We show that recent CTL-solving Transformer variants fail on CTL++. The simplicity of the task design allows for fine-grained control of task difficulty, as well as many insightful analyses. For example, we measure how much overlap between groups is needed by tested NNs for learning to compose. We also visualize how learned symbol representations in outputs of functions from different groups are compatible in case of success but not in case of failure. These results provide insights into failure cases reported on more complex compositions in the natural language domain. Our code is public.

Kazuki Irie, Róbert Csordás, Jürgen Schmidhuber

Unsupervised learning of discrete representations in neural networks (NNs) from continuous ones is essential for many modern applications. Vector Quantisation (VQ) has become popular for this, in particular in the context of generative models, such as Variational Auto-Encoders (VAEs), where the exponential moving average-based VQ (EMA-VQ) algorithm is often used. Here, we study an alternative VQ algorithm based on Kohonen's learning rule for the Self-Organising Map (KSOM; 1982). EMA-VQ is a special case of KSOM. KSOM is known to offer two potential benefits: empirically, it converges faster than EMA-VQ, and KSOM-generated discrete representations form a topological structure on the grid whose nodes are the discrete symbols, resulting in an artificial version of the brain's topographic map. We revisit these properties by using KSOM in VQ-VAEs for image processing. In our experiments, the speed-up compared to well-configured EMA-VQ is only observable at the beginning of training, but KSOM is generally much more robust, e.g., w.r.t. the choice of initialisation schemes.

Róbert Csordás, Christopher Potts, Christopher D. Manning et al.

The Linear Representation Hypothesis (LRH) states that neural networks learn to encode concepts as directions in activation space, and a strong version of the LRH states that models learn only such encodings. In this paper, we present a counterexample to this strong LRH: when trained to repeat an input token sequence, gated recurrent neural networks (RNNs) learn to represent the token at each position with a particular order of magnitude, rather than a direction. These representations have layered features that are impossible to locate in distinct linear subspaces. To show this, we train interventions to predict and manipulate tokens by learning the scaling factor corresponding to each sequence position. These interventions indicate that the smallest RNNs find only this magnitude-based solution, while larger RNNs have linear representations. These findings strongly indicate that interpretability research should not be confined by the LRH.

Kazuki Irie, Róbert Csordás, Jürgen Schmidhuber

Recent studies of the computational power of recurrent neural networks (RNNs) reveal a hierarchy of RNN architectures, given real-time and finite-precision assumptions. Here we study auto-regressive Transformers with linearised attention, a.k.a. linear Transformers (LTs) or Fast Weight Programmers (FWPs). LTs are special in the sense that they are equivalent to RNN-like sequence processors with a fixed-size state, while they can also be expressed as the now-popular self-attention networks. We show that many well-known results for the standard Transformer directly transfer to LTs/FWPs. Our formal language recognition experiments demonstrate how recently proposed FWP extensions such as recurrent FWPs and self-referential weight matrices successfully overcome certain limitations of the LT, e.g., allowing for generalisation on the parity problem. Our code is public.

Joakim Edin, Róbert Csordás, Tuukka Ruotsalo et al.

Ensuring faithful interpretability in large language models is imperative for trustworthy and reliable AI. A key obstacle is self-repair, a phenomenon where networks compensate for reduced signal in one component by amplifying others, masking the true importance of the ablated component. While prior work attributes self-repair to layer normalization and back-up components that compensate for ablated components, we identify a novel form occurring within the attention mechanism, where softmax redistribution conceals the influence of important attention scores. This leads traditional ablation and gradient-based methods to underestimate the significance of all components contributing to these attention scores. We introduce Gradient Interaction Modifications (GIM), a technique that accounts for self-repair during backpropagation. Extensive experiments across multiple large language models (Gemma 2B/9B, LLAMA 1B/3B/8B, Qwen 1.5B/3B) and diverse tasks demonstrate that GIM significantly improves faithfulness over existing circuit identification and feature attribution methods. Our work is a significant step toward better understanding the inner mechanisms of LLMs, which is crucial for improving them and ensuring their safety. Our code is available at https://github.com/JoakimEdin/gim.

Róbert Csordás, Piotr Piękos, Kazuki Irie et al.

Despite many recent works on Mixture of Experts (MoEs) for resource-efficient Transformer language models, existing methods mostly focus on MoEs for feedforward layers. Previous attempts at extending MoE to the self-attention layer fail to match the performance of the parameter-matched baseline. Our novel SwitchHead is an effective MoE method for the attention layer that successfully reduces both the compute and memory requirements, achieving wall-clock speedup, while matching the language modeling performance of the baseline Transformer. Our novel MoE mechanism allows SwitchHead to compute up to 8 times fewer attention matrices than the standard Transformer. SwitchHead can also be combined with MoE feedforward layers, resulting in fully-MoE "SwitchAll" Transformers. For our 262M parameter model trained on C4, SwitchHead matches the perplexity of standard models with only 44% compute and 27% memory usage. Zero-shot experiments on downstream tasks confirm the performance of SwitchHead, e.g., achieving more than 3.5% absolute improvements on BliMP compared to the baseline with an equal compute resource.

Róbert Csordás, Christopher D. Manning, Christopher Potts

Modern LLMs are increasingly deep, and depth correlates with performance, albeit with diminishing returns. However, do these models use their depth efficiently? Do they compose more features to create higher-order computations that are impossible in shallow models, or do they merely spread the same kinds of computation out over more layers? To address these questions, we analyze the residual stream of the Llama 3.1, Qwen 3, and OLMo 2 family of models. We find: First, comparing the output of the sublayers to the residual stream reveals that layers in the second half contribute much less than those in the first half, with a clear phase transition between the two halves. Second, skipping layers in the second half has a much smaller effect on future computations and output predictions. Third, for multihop tasks, we are unable to find evidence that models are using increased depth to compose subresults in examples involving many hops. Fourth, we seek to directly address whether deeper models are using their additional layers to perform new kinds of computation. To do this, we train linear maps from the residual stream of a shallow model to a deeper one. We find that layers with the same relative depth map best to each other, suggesting that the larger model simply spreads the same computations out over its many layers. All this evidence suggests that deeper models are not using their depth to learn new kinds of computation, but only using the greater depth to perform more fine-grained adjustments to the residual. This may help explain why increasing scale leads to diminishing returns for stacked Transformer architectures.

Julie Kallini, Shikhar Murty, Christopher D. Manning et al.

Models that rely on subword tokenization have significant drawbacks, such as sensitivity to character-level noise like spelling errors and inconsistent compression rates across different languages and scripts. While character- or byte-level models like ByT5 attempt to address these concerns, they have not gained widespread adoption -- processing raw byte streams without tokenization results in significantly longer sequence lengths, making training and inference inefficient. This work introduces MrT5 (MergeT5), a more efficient variant of ByT5 that integrates a token deletion mechanism in its encoder to dynamically shorten the input sequence length. After processing through a fixed number of encoder layers, a learned delete gate determines which tokens are to be removed and which are to be retained for subsequent layers. MrT5 effectively "merges" critical information from deleted tokens into a more compact sequence, leveraging contextual information from the remaining tokens. In continued pre-training experiments, we find that MrT5 can achieve significant gains in inference runtime with minimal effect on performance, as measured by bits-per-byte. Additionally, with multilingual training, MrT5 adapts to the orthographic characteristics of each language, learning language-specific compression rates. Furthermore, MrT5 shows comparable accuracy to ByT5 on downstream evaluations such as XNLI, TyDi QA, and character-level tasks while reducing sequence lengths by up to 75%. Our approach presents a solution to the practical limitations of existing byte-level models.

Piotr Piękos, Róbert Csordás, Jürgen Schmidhuber

Recent advances in large language models highlighted the excessive quadratic cost of self-attention. Despite the significant research efforts, subquadratic attention methods still suffer from inferior performance in practice. We hypothesize that dynamic, learned content-based sparsity can lead to more efficient attention mechanisms. We present Mixture of Sparse Attention (MoSA), a novel approach inspired by Mixture of Experts (MoE) with expert choice routing. MoSA dynamically selects tokens for each attention head, allowing arbitrary sparse attention patterns. By selecting $k$ tokens from a sequence of length $T$, MoSA reduces the computational complexity of each attention head from $O(T^2)$ to $O(k^2 + T)$. This enables using more heads within the same computational budget, allowing higher specialization. We show that among the tested sparse attention variants, MoSA is the only one that can outperform the dense baseline, sometimes with up to 27% better perplexity for an identical compute budget. MoSA can also reduce the resource usage compared to dense self-attention. Despite using torch implementation without an optimized kernel, perplexity-matched MoSA models are simultaneously faster in wall-clock time, require less memory for training, and drastically reduce the size of the KV-cache compared to the dense transformer baselines.

Aryaman Arora, Neil Rathi, Nikil Roashan Selvam et al. · stanford

State space models (SSMs) for language modelling promise an efficient and performant alternative to quadratic-attention Transformers, yet show variable performance on recalling basic information from the context. While performance on synthetic tasks like Associative Recall (AR) can point to this deficiency, behavioural metrics provide little information as to why--on a mechanistic level--certain architectures fail and others succeed. To address this, we conduct experiments on AR and find that only Transformers and Based SSM models fully succeed at AR, with Mamba a close third, whereas the other SSMs (H3, Hyena) fail. We then use causal interventions to explain why. We find that Transformers and Based learn to store key-value associations in-context using induction heads. By contrast, the SSMs compute these associations only at the last state, with only Mamba succeeding because of its short convolution component. To extend and deepen these findings, we introduce Associative Treecall (ATR), a synthetic task similar to AR based on PCFG induction. ATR introduces language-like hierarchical structure into the AR setting. We find that all architectures learn the same mechanism as they did for AR, and the same three models succeed at the task. These results reveal that architectures with similar accuracy may still have substantive differences, motivating the adoption of mechanistic evaluations.

Houjun Liu, Shikhar Murty, Christopher D. Manning et al.

Current approaches for scaling inference-time compute in transformers rely on training them to emit explicit chain-of-thought tokens before producing an answer. While these methods are powerful, they are limited because they cannot be applied during pretraining and are limited to only serially-generated, natural-language verbalization to scale inference-time compute. In this work, we propose Thoughtbubbles, a transformer variant that natively performs parallel adaptive computation in latent space by learning to fork or delete residual streams. Thus, tokens that require a large amount of computation can form a "bubble" of cloned residuals in the middle of the network for additional thinking. Crucially, this behavior is learned during pretraining with only language modeling loss. Thoughtbubbles outperforms both standard decoder LMs as well as non-adaptive parallel computation approaches on OpenWebText and peS2o perplexity and in zero-shot evaluations such as HellaSwag and LAMBADA after pretraining across 150M to 772M parameter scales. The implicit nature of our method enables adaptive computation to be learned starting at pretraining time, paving the way to unify train and test-time behavior for reasoning models.

Vincent Herrmann, Róbert Csordás, Jürgen Schmidhuber

Detecting when a neural sequence model does "interesting" computation is an open problem. The next token prediction loss is a poor indicator: Low loss can stem from trivially predictable sequences that are uninteresting, while high loss may reflect unpredictable but also irrelevant information that can be ignored by the model. We propose a better metric: measuring the model's ability to predict its own future hidden states. We show empirically that this metric -- in contrast to the next token prediction loss -- correlates with the intuitive interestingness of the task. To measure predictability, we introduce the architecture-agnostic "prediction of hidden states" (PHi) layer that serves as an information bottleneck on the main pathway of the network (e.g., the residual stream in Transformers). We propose a novel learned predictive prior that enables us to measure the novel information gained in each computation step, which serves as our metric. We show empirically that our metric predicts the description length of formal languages learned in-context, the complexity of mathematical reasoning problems, and the correctness of self-generated reasoning chains.

Mingchen Zhuge, Haozhe Liu, Francesco Faccio et al.

Both Minsky's "society of mind" and Schmidhuber's "learning to think" inspire diverse societies of large multimodal neural networks (NNs) that solve problems by interviewing each other in a "mindstorm." Recent implementations of NN-based societies of minds consist of large language models (LLMs) and other NN-based experts communicating through a natural language interface. In doing so, they overcome the limitations of single LLMs, improving multimodal zero-shot reasoning. In these natural language-based societies of mind (NLSOMs), new agents -- all communicating through the same universal symbolic language -- are easily added in a modular fashion. To demonstrate the power of NLSOMs, we assemble and experiment with several of them (having up to 129 members), leveraging mindstorms in them to solve some practical AI tasks: visual question answering, image captioning, text-to-image synthesis, 3D generation, egocentric retrieval, embodied AI, and general language-based task solving. We view this as a starting point towards much larger NLSOMs with billions of agents-some of which may be humans. And with this emergence of great societies of heterogeneous minds, many new research questions have suddenly become paramount to the future of artificial intelligence. What should be the social structure of an NLSOM? What would be the (dis)advantages of having a monarchical rather than a democratic structure? How can principles of NN economies be used to maximize the total reward of a reinforcement learning NLSOM? In this work, we identify, discuss, and try to answer some of these questions.

Anian Ruoss, Grégoire Delétang, Tim Genewein et al.

Transformers have impressive generalization capabilities on tasks with a fixed context length. However, they fail to generalize to sequences of arbitrary length, even for seemingly simple tasks such as duplicating a string. Moreover, simply training on longer sequences is inefficient due to the quadratic computation complexity of the global attention mechanism. In this work, we demonstrate that this failure mode is linked to positional encodings being out-of-distribution for longer sequences (even for relative encodings) and introduce a novel family of positional encodings that can overcome this problem. Concretely, our randomized positional encoding scheme simulates the positions of longer sequences and randomly selects an ordered subset to fit the sequence's length. Our large-scale empirical evaluation of 6000 models across 15 algorithmic reasoning tasks shows that our method allows Transformers to generalize to sequences of unseen length (increasing test accuracy by 12.0% on average).

Kazuki Irie, Róbert Csordás, Jürgen Schmidhuber

Linear layers in neural networks (NNs) trained by gradient descent can be expressed as a key-value memory system which stores all training datapoints and the initial weights, and produces outputs using unnormalised dot attention over the entire training experience. While this has been technically known since the 1960s, no prior work has effectively studied the operations of NNs in such a form, presumably due to prohibitive time and space complexities and impractical model sizes, all of them growing linearly with the number of training patterns which may get very large. However, this dual formulation offers a possibility of directly visualising how an NN makes use of training patterns at test time, by examining the corresponding attention weights. We conduct experiments on small scale supervised image classification tasks in single-task, multi-task, and continual learning settings, as well as language modelling, and discuss potentials and limits of this view for better understanding and interpreting how NNs exploit training patterns. Our code is public.

Kazuki Irie, Imanol Schlag, Róbert Csordás et al.

The weight matrix (WM) of a neural network (NN) is its program. The programs of many traditional NNs are learned through gradient descent in some error function, then remain fixed. The WM of a self-referential NN, however, can keep rapidly modifying all of itself during runtime. In principle, such NNs can meta-learn to learn, and meta-meta-learn to meta-learn to learn, and so on, in the sense of recursive self-improvement. While NN architectures potentially capable of implementing such behaviour have been proposed since the '90s, there have been few if any practical studies. Here we revisit such NNs, building upon recent successes of fast weight programmers and closely related linear Transformers. We propose a scalable self-referential WM (SRWM) that learns to use outer products and the delta update rule to modify itself. We evaluate our SRWM in supervised few-shot learning and in multi-task reinforcement learning with procedurally generated game environments. Our experiments demonstrate both practical applicability and competitive performance of the proposed SRWM. Our code is public.

Kazuki Irie, Imanol Schlag, Róbert Csordás et al.

We share our experience with the recently released WILDS benchmark, a collection of ten datasets dedicated to developing models and training strategies which are robust to domain shifts. Several experiments yield a couple of critical observations which we believe are of general interest for any future work on WILDS. Our study focuses on two datasets: iWildCam and FMoW. We show that (1) Conducting separate cross-validation for each evaluation metric is crucial for both datasets, (2) A weak correlation between validation and test performance might make model development difficult for iWildCam, (3) Minor changes in the training of hyper-parameters improve the baseline by a relatively large margin (mainly on FMoW), (4) There is a strong correlation between certain domains and certain target labels (mainly on iWildCam). To the best of our knowledge, no prior work on these datasets has reported these observations despite their obvious importance. Our code is public.

Róbert Csordás, Kazuki Irie, Jürgen Schmidhuber

Despite progress across a broad range of applications, Transformers have limited success in systematic generalization. The situation is especially frustrating in the case of algorithmic tasks, where they often fail to find intuitive solutions that route relevant information to the right node/operation at the right time in the grid represented by Transformer columns. To facilitate the learning of useful control flow, we propose two modifications to the Transformer architecture, copy gate and geometric attention. Our novel Neural Data Router (NDR) achieves 100% length generalization accuracy on the classic compositional table lookup task, as well as near-perfect accuracy on the simple arithmetic task and a new variant of ListOps testing for generalization across computational depths. NDR's attention and gating patterns tend to be interpretable as an intuitive form of neural routing. Our code is public.

Róbert Csordás, Kazuki Irie, Jürgen Schmidhuber

Recently, many datasets have been proposed to test the systematic generalization ability of neural networks. The companion baseline Transformers, typically trained with default hyper-parameters from standard tasks, are shown to fail dramatically. Here we demonstrate that by revisiting model configurations as basic as scaling of embeddings, early stopping, relative positional embedding, and Universal Transformer variants, we can drastically improve the performance of Transformers on systematic generalization. We report improvements on five popular datasets: SCAN, CFQ, PCFG, COGS, and Mathematics dataset. Our models improve accuracy from 50% to 85% on the PCFG productivity split, and from 35% to 81% on COGS. On SCAN, relative positional embedding largely mitigates the EOS decision problem (Newman et al., 2020), yielding 100% accuracy on the length split with a cutoff at 26. Importantly, performance differences between these models are typically invisible on the IID data split. This calls for proper generalization validation sets for developing neural networks that generalize systematically. We publicly release the code to reproduce our results.

Kazuki Irie, Imanol Schlag, Róbert Csordás et al.

Transformers with linearised attention (''linear Transformers'') have demonstrated the practical scalability and effectiveness of outer product-based Fast Weight Programmers (FWPs) from the '90s. However, the original FWP formulation is more general than the one of linear Transformers: a slow neural network (NN) continually reprograms the weights of a fast NN with arbitrary architecture. In existing linear Transformers, both NNs are feedforward and consist of a single layer. Here we explore new variations by adding recurrence to the slow and fast nets. We evaluate our novel recurrent FWPs (RFWPs) on two synthetic algorithmic tasks (code execution and sequential ListOps), Wikitext-103 language models, and on the Atari 2600 2D game environment. Our models exhibit properties of Transformers and RNNs. In the reinforcement learning setting, we report large improvements over LSTM in several Atari games. Our code is public.

Róbert Csordás, Sjoerd van Steenkiste, Jürgen Schmidhuber

Neural networks (NNs) whose subnetworks implement reusable functions are expected to offer numerous advantages, including compositionality through efficient recombination of functional building blocks, interpretability, preventing catastrophic interference, etc. Understanding if and how NNs are modular could provide insights into how to improve them. Current inspection methods, however, fail to link modules to their functionality. In this paper, we present a novel method based on learning binary weight masks to identify individual weights and subnets responsible for specific functions. Using this powerful tool, we contribute an extensive study of emerging modularity in NNs that covers several standard architectures and datasets. We demonstrate how common NNs fail to reuse submodules and offer new insights into the related issue of systematic generalization on language tasks.

Róbert Csordás, Jürgen Schmidhuber

The Differentiable Neural Computer (DNC) can learn algorithmic and question answering tasks. An analysis of its internal activation patterns reveals three problems: Most importantly, the lack of key-value separation makes the address distribution resulting from content-based look-up noisy and flat, since the value influences the score calculation, although only the key should. Second, DNC's de-allocation of memory results in aliasing, which is a problem for content-based look-up. Thirdly, chaining memory reads with the temporal linkage matrix exponentially degrades the quality of the address distribution. Our proposed fixes of these problems yield improved performance on arithmetic tasks, and also improve the mean error rate on the bAbI question answering dataset by 43%.