Dimitar Kazakov

Tianda Sun, Dimitar Kazakov

We test the standard RLVR tool-use recipe -- GRPO on Qwen2.5-7B-Instruct -- on a deliberately minimal knowledge-graph tool API: four Freebase navigation verbs over Complex WebQuestions. Under a self-verifiable retrieval reward, the policy's tool-grounded answer rate climbs from $3.8\%$ to $9.6\%$ over 250 steps, then collapses to $0\%$ within a single 50-step window -- a \emph{peak-then-collapse} pattern replicated across four seeds. Across seven reward designs, we find four recurring failure modes: adding denser or more targeted proxy rewards shifts the failure mode rather than eliminating it. We argue that a key difference from Python interpreters, web search, and JSON APIs is interface feedback: their failures often leak natural-language signal the model saw in pretraining. A Python traceback names the failing line; an empty Freebase result \texttt{[]} does not. Stripping away that surface exposes a degradation regime that same-family reward redesigns do not fix. A direct oracle ablation rules out relation selection: injecting gold relations at every retrieval call lifts exact-match accuracy by only $+0.20$~pp, and $95.4\%$ of retrieval-dependent errors are retrieval-composition failures rather than answer-extraction failures. As a mitigation, one-iteration self-distillation reaches $40.0\%$ EM at 7B and is capacity-invariant: doubling capacity to 14B improves EM by only $0.25$~pp, and initialization barely matters -- the ceiling appears interface-bound within the 7B--14B range tested.

Tianda Sun, Dimitar Kazakov

Tool-using LLM agents produce trajectories whose calls form a directed dependency graph: earlier tool outputs supply arguments to later calls. Whether this execution structure is represented inside the model is unknown; prior structural probes have targeted static code or chain-of-thought text, not an agent's run-time call graph. A low-capacity edge probe on the residual stream of Qwen3-32B decodes the tool-call dependency graph well above both a Hewitt--Liang random-label control and a positional baseline. A counterfactual contrast between value corruption and structural perturbation indicates the signal tracks abstract topology rather than identifier values, and replicates under an independent, non-substring oracle. The non-positional component replicates on three further interactive multi-hop benchmarks and attenuates as call order alone becomes a sufficient proxy for dependency, vanishing in single-shot planning. Per-layer activation patching shifts the probe at a later, non-patched boundary, evidence that the representation propagates rather than passively reads out, though the realised tool call does not move. To our knowledge this is the first structural probe of an LLM agent's runtime tool-call dependency graph. Our claims concern representation, not behavioural control, and span two model families and one primary domain.

Tianda Sun, Dimitar Kazakov

Large language models (LLMs) are increasingly evaluated on their ability to perform multi-hop reasoning, i.e., to combine multiple pieces of information into a coherent inference. We introduce KinshipQA, a benchmark designed to probe this capability through reasoning over kinship relations. The central contribution of our work is a generative pipeline that produces, on demand, large-scale, realistic, and culture-specific genealogical data: collections of interconnected family trees that satisfy explicit marriage constraints associated with different kinship systems. This allows task difficulty, cultural assumptions, and relational depth to be systematically controlled and varied. From these genealogies, we derive textual inference tasks that require reasoning over implicit relational chains. We evaluate the resulting benchmark using six state-of-the-art LLMs, spanning both open-source and closed-source models, under a uniform zero-shot protocol with deterministic decoding. Performance is measured using exact-match and set-based metrics. Our results demonstrate that KinshipQA yields a wide spread of outcomes and exposes systematic differences in multi-hop reasoning across models and cultural settings.

Charmaine Barker, Daniel Bethell, Dimitar Kazakov

Mitigating bias in automated decision-making systems, particularly in deep learning models, is a critical challenge due to nuanced definitions of fairness, dataset-specific biases, and the inherent trade-off between fairness and accuracy. To address these issues, we introduce FairVIC, an innovative approach that enhances fairness in neural networks by integrating variance, invariance, and covariance terms into the loss function during training. Unlike methods that rely on predefined fairness criteria, FairVIC abstracts fairness concepts to minimise dependency on protected characteristics. We evaluate FairVIC against comparable bias mitigation techniques on benchmark datasets, considering both group and individual fairness, and conduct an ablation study on the accuracy-fairness trade-off. FairVIC demonstrates significant improvements ($\approx70\%$) in fairness across all tested metrics without compromising accuracy, thus offering a robust, generalisable solution for fair deep learning across diverse tasks and datasets.

Charmaine Barker, Dimitar Kazakov

The presence of specific linguistic signals particular to a certain sub-group can become highly salient to language models during training. In automated decision-making settings, this may lead to biased outcomes when models rely on cues that correlate with protected characteristics. We investigate whether prompting ChatGPT to rewrite text using simplification, neutralisation, localisation, and formalisation can reduce demographic signals while preserving meaning. Experimental results show a statistically significant drop in location classification accuracy across multiple models after transformation, suggesting reduced reliance on group-specific language. At the same time, sentiment analysis and rating prediction tasks confirm that the core meaning of the reviews remains greatly intact. These results suggest that prompt-based rewriting offers a practical and generalisable approach for mitigating bias in text classification.

Mudita Sharma, Manuel Lopez-Ibanez, Dimitar Kazakov

We conduct an exhaustive survey of adaptive selection of operators (AOS) in Evolutionary Algorithms (EAs). We simplified the AOS structure by adding more components to the framework to built upon the existing categorisation of AOS methods. In addition to simplifying, we looked at the commonality among AOS methods from literature to generalise them. Each component is presented with a number of alternative choices, each represented with a formula. We make three sets of comparisons. First, the methods from literature are tested on the BBOB test bed with their default hyper parameters. Second, the hyper parameters of these methods are tuned using an offline configurator known as IRACE. Third, for a given set of problems, we use IRACE to select the best combination of components and tune their hyper parameters.

Mudita Sharma, Alexandros Komninos, Manuel Lopez Ibanez et al.

Adaptive Operator Selection (AOS) is an approach that controls discrete parameters of an Evolutionary Algorithm (EA) during the run. In this paper, we propose an AOS method based on Double Deep Q-Learning (DDQN), a Deep Reinforcement Learning method, to control the mutation strategies of Differential Evolution (DE). The application of DDQN to DE requires two phases. First, a neural network is trained offline by collecting data about the DE state and the benefit (reward) of applying each mutation strategy during multiple runs of DE tackling benchmark functions. We define the DE state as the combination of 99 different features and we analyze three alternative reward functions. Second, when DDQN is applied as a parameter controller within DE to a different test set of benchmark functions, DDQN uses the trained neural network to predict which mutation strategy should be applied to each parent at each generation according to the DE state. Benchmark functions for training and testing are taken from the CEC2005 benchmark with dimensions 10 and 30. We compare the results of the proposed DE-DDQN algorithm to several baseline DE algorithms using no online selection, random selection and other AOS methods, and also to the two winners of the CEC2005 competition. The results show that DE-DDQN outperforms the non-adaptive methods for all functions in the test set; while its results are comparable with the last two algorithms.

Matthew Butler, Dimitar Kazakov

In time series analysis research there is a strong interest in discrete representations of real valued data streams. One approach that emerged over a decade ago and is still considered state-of-the-art is the Symbolic Aggregate Approximation algorithm. This discretization algorithm was the first symbolic approach that mapped a real-valued time series to a symbolic representation that was guaranteed to lower-bound Euclidean distance. The interest of this paper concerns the SAX assumption of data being highly Gaussian and the use of the standard normal curve to choose partitions to discretize the data. Though not necessarily, but generally, and certainly in its canonical form, the SAX approach chooses partitions on the standard normal curve that would produce an equal probability for each symbol in a finite alphabet to occur. This procedure is generally valid as a time series is normalized before the rest of the SAX algorithm is applied. However there exists a caveat to this assumption of equi-probability due to the intermediate step of Piecewise Aggregate Approximation (PAA). What we will show in this paper is that when PAA is applied the distribution of the data is indeed altered, resulting in a shrinking standard deviation that is proportional to the number of points used to create a segment of the PAA representation and the degree of auto-correlation within the series. Data that exhibits statistically significant auto-correlation is less affected by this shrinking distribution. As the standard deviation of the data contracts, the mean remains the same, however the distribution is no longer standard normal and therefore the partitions based on the standard normal curve are no longer valid for the assumption of equal probability.