Aryo Pradipta Gema

Alexander Hägele, Aryo Pradipta Gema, Henry Sleight et al. · anthropic

As AI becomes more capable, we entrust it with more general and consequential tasks. The risks from failure grow more severe with increasing task scope. It is therefore important to understand how extremely capable AI models will fail: Will they fail by systematically pursuing goals we do not intend? Or will they fail by being a hot mess, and taking nonsensical actions that do not further any goal? We operationalize this question using a bias-variance decomposition of the errors made by AI models: An AI's \emph{incoherence} on a task is measured over test-time randomness as the fraction of its error that stems from variance rather than bias in task outcome. Across all tasks and frontier models we measure, the longer models spend reasoning and taking actions, \emph{the more incoherent} their failures become. Incoherence changes with model scale in a way that is experiment dependent. However, in several settings, larger, more capable models are more incoherent than smaller models. Consequently, scale alone seems unlikely to eliminate incoherence. Instead, as more capable AIs pursue harder tasks, requiring more sequential action and thought, our results predict failures to be accompanied by more incoherent behavior. This suggests a future where AIs sometimes cause industrial accidents (due to unpredictable misbehavior), but are less likely to exhibit consistent pursuit of a misaligned goal. This increases the relative importance of alignment research targeting reward hacking or goal misspecification.

Aryo Pradipta Gema, Pasquale Minervini, Luke Daines et al.

Adapting pretrained language models to novel domains, such as clinical applications, traditionally involves retraining their entire set of parameters. Parameter-Efficient Fine-Tuning (PEFT) techniques for fine-tuning language models significantly reduce computational requirements by selectively fine-tuning small subsets of parameters. In this study, we propose a two-step PEFT framework and evaluate it in the clinical domain. Our approach combines a specialised PEFT adapter layer designed for clinical domain adaptation with another adapter specialised for downstream tasks. We evaluate the framework on multiple clinical outcome prediction datasets, comparing it to clinically trained language models. Our framework achieves a better AUROC score averaged across all clinical downstream tasks compared to clinical language models. In particular, we observe large improvements of 4-5% AUROC in large-scale multilabel classification tasks, such as diagnoses and procedures classification. To our knowledge, this study is the first to provide an extensive empirical analysis of the interplay between PEFT techniques and domain adaptation in an important real-world domain of clinical applications.

Wai-Chung Kwan, Aryo Pradipta Gema, Joshua Ong Jun Leang et al.

Self-play can train language models without external supervision. However, existing methods require rule-checkable answers, leaving open-ended tasks dependent on curated prompts or frontier-model judges. We introduce SCOPE, a data-free self-play framework for open-ended tasks that co-evolves two policies: a Challenger that generates document-grounded tasks, and a Solver that answers them through multi-turn retrieval. A frozen copy of the initial model serves as the self-judge, which writes task-specific rubrics from the source document and grades Solver responses against them. Across three 7-8B instruction-tuned models (Qwen2.5, Qwen3, OLMo-3), SCOPE improves open-ended performance by up to +10.4 points on eight benchmarks and matches or exceeds GRPO_data trained on ~9K curated prompts. Although trained only on open-ended tasks, SCOPE also improves held-out short-form QA by up to +13.8 points on seven held-out benchmarks, surpassing GRPO_data on all three models. Ablations show that co-evolving the Challenger is necessary to keep tasks near the Solver's frontier, that gains arise from improvements in both retrieval and synthesis with the relative contribution varying by task, and that rubric generation quality is the bottleneck for self-judging.

Farooq Ahmad Wani, Alessandro Suglia, Rohit Saxena et al.

The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions reflect stable multimodal processing. In this work, we argue that this assumption is insufficient. We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness (spatial consistency of vision tokens), alongside standard label-based metrics. Applying this framework to modern VLMs across the SEEDBench, MMMU, and POPE datasets reveals three distinct failure modes. First, models frequently preserve predicted answers while undergoing substantial internal representation drift; for perturbations such as text overlays, this drift approaches the magnitude of inter-image variability, indicating that representations move to regions typically occupied by unrelated inputs despite unchanged outputs. Second, robustness does not improve with scale; larger models achieve higher accuracy but exhibit equal or greater sensitivity, consistent with sharper yet more fragile decision boundaries. Third, we find that perturbations affect tasks differently: they harm reasoning when they disrupt how models combine coarse and fine visual cues, but on the hallucination benchmarks, they can reduce false positives by making models generate more conservative answers.

Giorgos Lysandrou, Roma English Owen, Vanja Popovic et al.

There exists an invisible barrier between healthcare professionals' perception of a patient's clinical experience and the reality. This barrier may be induced by the environment that hinders patients from sharing their experiences openly with healthcare professionals. As patients are observed to discuss and exchange knowledge more candidly on social media, valuable insights can be leveraged from these platforms. However, the abundance of non-patient posts on social media necessitates filtering out such irrelevant content to distinguish the genuine voices of patients, a task we refer to as patient voice classification. In this study, we analyse the importance of linguistic characteristics in accurately classifying patient voices. Our findings underscore the essential role of linguistic and statistical text similarity analysis in identifying common patterns among patient groups. These results allude to even starker differences in the way patients express themselves at a disease level and across various therapeutic domains. Additionally, we fine-tuned a pre-trained Language Model on the combined datasets with similar linguistic patterns, resulting in a highly accurate automatic patient voice classification. Being the pioneering study on the topic, our focus on extracting authentic patient experiences from social media stands as a crucial step towards advancing healthcare standards and fostering a patient-centric approach.

Matteo Attimonelli, Alessandro De Bellis, Aryo Pradipta Gema et al.

Composed Image Retrieval (CIR) is a multimodal retrieval task where a query consists of a reference image and a textual modification, and the goal is to retrieve a target image satisfying both. In principle, strong performance on CIR benchmarks is assumed to require multimodal composition, i.e., combining complementary information from reference image and textual modification. In this work, we show that this assumption does not always hold. Across four widely used CIR benchmarks and eleven Generalist Multimodal Embedding models, a large fraction of queries can be solved using a single modality (from 32.2% to 83.6%), revealing pervasive unimodal shortcuts. Thus, high CIR performance can arise from unimodal signals rather than true multimodal composition. To better understand this issue, we perform a two-stage audit. First, we identify shortcut-solvable queries through cross-model analysis. Second, we conduct human validation on 4,741 shortcut-free queries, of which only 1,689 are well-formed, with common issues including ambiguous edits and mismatched targets. Re-evaluating models on this validated subset reveals qualitatively different behaviour: queries can no longer be solved with a single modality, and successful retrieval requires combining both inputs. While accuracy decreases, reliance on multimodal information increases. Overall, current CIR benchmarks conflate shortcut-solvable, noisy, and genuinely compositional queries, leading to an overestimation of model capability in multimodal composition.

Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong et al.

Maybe not. We identify and analyse errors in the popular Massive Multitask Language Understanding (MMLU) benchmark. Even though MMLU is widely adopted, our analysis demonstrates numerous ground truth errors that obscure the true capabilities of LLMs. For example, we find that 57% of the analysed questions in the Virology subset contain errors. To address this issue, we introduce a comprehensive framework for identifying dataset errors using a novel error annotation protocol. Then, we create MMLU-Redux, which is a subset of 5,700 manually re-annotated questions across all 57 MMLU subjects. We estimate that 6.49% of MMLU questions contain errors. Using MMLU-Redux, we demonstrate significant discrepancies with the model performance metrics that were originally reported. Our results strongly advocate for revising MMLU's error-ridden questions to enhance its future utility and reliability as a benchmark. https://huggingface.co/datasets/edinburgh-dawg/mmlu-redux-2.0.

Aryo Pradipta Gema, Dominik Grabarczyk, Wolf De Wulf et al.

Knowledge graphs are powerful tools for representing and organising complex biomedical data. Several knowledge graph embedding algorithms have been proposed to learn from and complete knowledge graphs. However, a recent study demonstrates the limited efficacy of these embedding algorithms when applied to biomedical knowledge graphs, raising the question of whether knowledge graph embeddings have limitations in biomedical settings. This study aims to apply state-of-the-art knowledge graph embedding models in the context of a recent biomedical knowledge graph, BioKG, and evaluate their performance and potential downstream uses. We achieve a three-fold improvement in terms of performance based on the HITS@10 score over previous work on the same biomedical knowledge graph. Additionally, we provide interpretable predictions through a rule-based method. We demonstrate that knowledge graph embedding models are applicable in practice by evaluating the best-performing model on four tasks that represent real-life polypharmacy situations. Results suggest that knowledge learnt from large biomedical knowledge graphs can be transferred to such downstream use cases. Our code is available at https://github.com/aryopg/biokge.

Aryo Pradipta Gema, Michał Kobiela, Achille Fraisse et al.

The SARS-CoV-2 pandemic has emphasised the importance of developing a universal vaccine that can protect against current and future variants of the virus. The present study proposes a novel conditional protein Language Model architecture, called Vaxformer, which is designed to produce natural-looking antigenicity-controlled SARS-CoV-2 spike proteins. We evaluate the generated protein sequences of the Vaxformer model using DDGun protein stability measure, netMHCpan antigenicity score, and a structure fidelity score with AlphaFold to gauge its viability for vaccine development. Our results show that Vaxformer outperforms the existing state-of-the-art Conditional Variational Autoencoder model to generate antigenicity-controlled SARS-CoV-2 spike proteins. These findings suggest promising opportunities for conditional Transformer models to expand our understanding of vaccine design and their role in mitigating global health challenges. The code used in this study is available at https://github.com/aryopg/vaxformer .

Yu Zhao, Alessio Devoto, Giwon Hong et al.

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context -- this phenomenon, known as \emph{context-memory knowledge conflicts}, can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. Analysing the internal activations of LLMs, we find that they can internally register the signals of knowledge conflict at mid-layers. Such signals allow us to detect whether a knowledge conflict occurs and use \emph{inference-time} intervention strategies to resolve it. In this work, we propose \textsc{SpARE}, a \emph{training-free} representation engineering method that uses pre-trained sparse auto-encoders (SAEs) to control the knowledge selection behaviour of LLMs. \textsc{SpARE} identifies the functional features that control the knowledge selection behaviours and applies them to edit the internal activations of LLMs at inference time. Our experimental results show that \textsc{SpARE} can effectively control the usage of either knowledge source to resolve knowledge conflict in open-domain question-answering tasks, surpassing existing representation engineering methods ($+10\%$) as well as contrastive decoding methods ($+15\%$).

Giwon Hong, Aryo Pradipta Gema, Rohit Saxena et al.

Large Language Models (LLMs) have transformed the Natural Language Processing (NLP) landscape with their remarkable ability to understand and generate human-like text. However, these models are prone to ``hallucinations'' -- outputs that do not align with factual reality or the input context. This paper introduces the Hallucinations Leaderboard, an open initiative to quantitatively measure and compare the tendency of each model to produce hallucinations. The leaderboard uses a comprehensive set of benchmarks focusing on different aspects of hallucinations, such as factuality and faithfulness, across various tasks, including question-answering, summarisation, and reading comprehension. Our analysis provides insights into the performance of different models, guiding researchers and practitioners in choosing the most reliable models for their applications.

Aryo Pradipta Gema, Alexander Hägele, Runjin Chen et al.

We construct evaluation tasks where extending the reasoning length of Large Reasoning Models (LRMs) deteriorates performance, exhibiting an inverse scaling relationship between test-time compute and accuracy. Our evaluation tasks span four categories: simple counting tasks with distractors, regression tasks with spurious features, deduction tasks with constraint tracking, and advanced AI risks. We identify five distinct failure modes when models reason for longer: 1) Claude models become increasingly distracted by irrelevant information; 2) OpenAI o-series models resist distractors but overfit to problem framings; 3) models shift from reasonable priors to spurious correlations; 4) all models show difficulties in maintaining focus on complex deductive tasks; and 5) extended reasoning may amplify concerning behaviors, with Claude Sonnet 4 showing increased expressions of self-preservation. These findings suggest that while test-time compute scaling remains promising for improving model capabilities, it may inadvertently reinforce problematic reasoning patterns. Our results demonstrate the importance of evaluating models across diverse reasoning lengths to identify and address these failure modes in LRMs.

Aryo Pradipta Gema, Chen Jin, Ahmed Abdulaal et al.

Large Language Models (LLMs) often hallucinate, producing unfaithful or factually incorrect outputs by misrepresenting the provided context or incorrectly recalling internal knowledge. Recent studies have identified specific attention heads within the Transformer architecture, known as retrieval heads, responsible for extracting relevant contextual information. We hypothesise that masking these retrieval heads can induce hallucinations and that contrasting the outputs of the base LLM and the masked LLM can reduce hallucinations. To this end, we propose Decoding by Contrasting Retrieval Heads (DeCoRe), a novel training-free decoding strategy that amplifies information found in the context and model parameters. DeCoRe mitigates potentially hallucinated responses by dynamically contrasting the outputs of the base LLM and the masked LLM, using conditional entropy as a guide. Our extensive experiments confirm that DeCoRe significantly improves performance on tasks requiring high contextual faithfulness, such as summarisation (XSum by 18.6%), instruction following (MemoTrap by 10.9%), and open-book question answering (NQ-Open by 2.4% and NQ-Swap by 5.5%).

Ne Luo, Aryo Pradipta Gema, Xuanli He et al.

Large language models (LLMs) remain prone to factual inaccuracies and computational errors, including hallucinations and mistakes in mathematical reasoning. Recent work augmented LLMs with tools to mitigate these shortcomings, but often requires curated gold tool-use demonstrations. In this paper, we investigate whether LLMs can learn to use tools without demonstrations. First, we analyse zero-shot prompting strategies to guide LLMs in tool utilisation. Second, we propose a self-training method to synthesise tool-use traces using the LLM itself. We compare supervised fine-tuning and preference fine-tuning techniques for fine-tuning the model on datasets constructed using existing Question Answering (QA) datasets, i.e., TriviaQA and GSM8K. Experiments show that tool-use enhances performance on a long-tail knowledge task: 3.7% on PopQA, which is used solely for evaluation, but leads to mixed results on other datasets, i.e., TriviaQA, GSM8K, and NQ-Open. Our findings highlight the potential and challenges of integrating external tools into LLMs without demonstrations.

Joshua Ong Jun Leang, Aryo Pradipta Gema, Shay B. Cohen

Mathematical reasoning remains a significant challenge for large language models (LLMs), despite progress in prompting techniques such as Chain-of-Thought (CoT). We present Chain of Mathematically Annotated Thought (CoMAT), which enhances reasoning through two stages: Symbolic Conversion (converting natural language queries into symbolic form) and Reasoning Execution (deriving answers from symbolic representations). CoMAT operates entirely with a single LLM and without external solvers. Across four LLMs, CoMAT outperforms traditional CoT on six out of seven benchmarks, achieving gains of 4.48% on MMLU-Redux (MATH) and 4.58% on GaoKao MCQ. In addition to improved performance, CoMAT ensures faithfulness and verifiability, offering a transparent reasoning process for complex mathematical tasks

Neel Rajani, Aryo Pradipta Gema, Seraphina Goldfarb-Tarrant et al.

Training large language models (LLMs) for reasoning via maths and code datasets has become a major new focus in LLM post-training. Two particularly popular approaches are reinforcement learning (RL) and supervised fine-tuning (SFT), but their training dynamics are poorly understood. We present a comparative analysis of RL and SFT on the same maths problems with the same model and similar hyperparameters. We find that RL yields minor in-domain gains on maths and slight degradation on knowledge-intensive benchmarks like MMLU, while both trends are more pronounced in SFT. We also analyse model parameters across checkpoints, observing that both algorithms modify query and key weights the most. Meanwhile, SFT exhibits greater updates and also affects mid-layer MLPs more, leading us to hypothesise that this may have caused the out-of-domain degradation. We therefore investigate whether freezing parts of the model during training can mitigate the reduced performance on knowledge-intensive benchmarks. However, our results are inconclusive, with benefits on GPQA:Diamond and degradation on other benchmarks. Taken together, our observations provide a preliminary indication for why RL amplifies existing capabilities, while SFT replaces old skills with new ones.

Rohit Saxena, Aryo Pradipta Gema, Pasquale Minervini

Understanding time from visual representations is a fundamental cognitive skill, yet it remains a challenge for multimodal large language models (MLLMs). In this work, we investigate the capabilities of MLLMs in interpreting time and date through analogue clocks and yearly calendars. To facilitate this, we curated a structured dataset comprising two subsets: 1) $\textit{ClockQA}$, which comprises various types of clock styles$-$standard, black-dial, no-second-hand, Roman numeral, and arrow-hand clocks$-$paired with time related questions; and 2) $\textit{CalendarQA}$, which consists of yearly calendar images with questions ranging from commonly known dates (e.g., Christmas, New Year's Day) to computationally derived ones (e.g., the 100th or 153rd day of the year). We aim to analyse how MLLMs can perform visual recognition, numerical reasoning, and temporal inference when presented with time-related visual data. Our evaluations show that despite recent advancements, reliably understanding time remains a significant challenge for MLLMs.

Aryo Pradipta Gema, Giwon Hong, Pasquale Minervini et al.

The NLI4CT task assesses Natural Language Inference systems in predicting whether hypotheses entail or contradict evidence from Clinical Trial Reports. In this study, we evaluate various Large Language Models (LLMs) with multiple strategies, including Chain-of-Thought, In-Context Learning, and Parameter-Efficient Fine-Tuning (PEFT). We propose a PEFT method to improve the consistency of LLMs by merging adapters that were fine-tuned separately using triplet and language modelling objectives. We found that merging the two PEFT adapters improves the F1 score (+0.0346) and consistency (+0.152) of the LLMs. However, our novel methods did not produce more accurate results than GPT-4 in terms of faithfulness and consistency. Averaging the three metrics, GPT-4 ranks joint-first in the competition with 0.8328. Finally, our contamination analysis with GPT-4 indicates that there was no test data leakage.

Yu Zhao, Xiaotang Du, Giwon Hong et al.

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context. Such conflicts can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. In this work, we investigate whether LLMs can identify knowledge conflicts and whether it is possible to know which source of knowledge the model will rely on by analysing the residual stream of the LLM. Through probing tasks, we find that LLMs can internally register the signal of knowledge conflict in the residual stream, which can be accurately detected by probing the intermediate model activations. This allows us to detect conflicts within the residual stream before generating the answers without modifying the input or model parameters. Moreover, we find that the residual stream shows significantly different patterns when the model relies on contextual knowledge versus parametric knowledge to resolve conflicts. This pattern can be employed to estimate the behaviour of LLMs when conflict happens and prevent unexpected answers before producing the answers. Our analysis offers insights into how LLMs internally manage knowledge conflicts and provides a foundation for developing methods to control the knowledge selection processes.

Joshua Ong Jun Leang, Zheng Zhao, Aryo Pradipta Gema et al.

Best-of-n sampling improves the accuracy of large language models (LLMs) and large reasoning models (LRMs) by generating multiple candidate solutions and selecting the one with the highest reward. The key challenge for reasoning tasks is designing a scoring function that can identify correct reasoning chains without access to ground-truth answers. We propose Probabilistic Confidence Selection And Ranking (PiCSAR): a simple, training-free method that scores each candidate generation using the joint log-likelihood of the reasoning and final answer. The joint log-likelihood of the reasoning and final answer naturally decomposes into reasoning confidence and answer confidence. PiCSAR achieves substantial gains across diverse benchmarks (+10.18 on MATH500, +9.81 on AIME2025), outperforming baselines with at least 2x fewer samples in 16 out of 20 comparisons. Our analysis reveals that correct reasoning chains exhibit significantly higher reasoning and answer confidence, justifying the effectiveness of PiCSAR.

Alexander Murphy, Mohd Sanad Zaki Rizvi, Aden Haussmann et al.

Large Language Models (LLMs) frequently produce factually inaccurate outputs - a phenomenon known as hallucination - which limits their accuracy in knowledge-intensive NLP tasks. Retrieval-augmented generation and agentic frameworks such as Reasoning and Acting (ReAct) can address this issue by giving the model access to external knowledge. However, LLMs often fail to remain faithful to retrieved information. Mitigating this is critical, especially if LLMs are required to reason about the retrieved information. Recent research has explored training-free decoding strategies to improve the faithfulness of model generations. We present a systematic analysis of how the combination of the ReAct framework and decoding strategies (i.e., DeCoRe, DoLa, and CAD) can influence the faithfulness of LLM-generated answers. Our results show that combining an agentic framework for knowledge retrieval with decoding methods that enhance faithfulness can increase accuracy on the downstream Multi-Hop Question Answering tasks. For example, we observe an F1 increase from 19.5 to 32.6 on HotpotQA when using ReAct and DoLa.

Mohammad Reza Ghasemi Madani, Aryo Pradipta Gema, Gabriele Sarti et al.

Feature attribution (FA) methods are common post-hoc approaches that explain how Large Language Models (LLMs) make predictions. Accordingly, generating faithful attributions that reflect the actual inner behavior of the model is crucial. In this paper, we introduce Noiser, a perturbation-based FA method that imposes bounded noise on each input embedding and measures the robustness of the model against partially noised input to obtain the input attributions. Additionally, we propose an answerability metric that employs an instructed judge model to assess the extent to which highly scored tokens suffice to recover the predicted output. Through a comprehensive evaluation across six LLMs and three tasks, we demonstrate that Noiser consistently outperforms existing gradient-based, attention-based, and perturbation-based FA methods in terms of both faithfulness and answerability, making it a robust and effective approach for explaining language model predictions.

Igor Shilov, Alex Cloud, Aryo Pradipta Gema et al.

Large Language Models increasingly possess capabilities that carry dual-use risks. While data filtering has emerged as a pretraining-time mitigation, it faces significant challenges: labeling whether data is harmful is expensive at scale, and given improving sample efficiency with larger models, even small amounts of mislabeled content could give rise to dangerous capabilities. To address risks associated with mislabeled harmful content, prior work proposed Gradient Routing (Cloud et al., 2024) -- a technique that localizes target knowledge into a dedicated subset of model parameters so they can later be removed. We explore an improved variant of Gradient Routing, which we call Selective GradienT Masking (SGTM), with particular focus on evaluating its robustness to label noise. SGTM zero-masks selected gradients such that target domain examples only update their dedicated parameters. We test SGTM's effectiveness in two applications: removing knowledge of one language from a model trained on a bilingual synthetic dataset, and removing biology knowledge from a model trained on English Wikipedia. In both cases SGTM provides better retain/forget trade-off in the presence of labeling errors compared to both data filtering and a previously proposed instantiation of Gradient Routing. Unlike shallow unlearning approaches that can be quickly undone through fine-tuning, SGTM exhibits strong robustness to adversarial fine-tuning, requiring seven times more fine-tuning steps to reach baseline performance on the forget set compared to a finetuning-based unlearning method (RMU). Our results suggest SGTM provides a promising pretraining-time complement to existing safety mitigations, particularly in settings where label noise is unavoidable.

Jingjie Zheng, Aryo Pradipta Gema, Giwon Hong et al.

Retrieval Augmented Generation (RAG) frameworks improve the accuracy of large language models (LLMs) by integrating external knowledge from retrieved documents, thereby overcoming the limitations of models' static intrinsic knowledge. However, these systems are susceptible to adversarial attacks that manipulate the retrieval process by introducing documents that are adversarial yet semantically similar to the query. Notably, while these adversarial documents resemble the query, they exhibit weak similarity to benign documents in the retrieval set. Thus, we propose a simple yet effective Graph-based Reranking against Adversarial Document Attacks (GRADA) framework aiming at preserving retrieval quality while significantly reducing the success of adversaries. Our study evaluates the effectiveness of our approach through experiments conducted on five LLMs: GPT-3.5-Turbo, GPT-4o, Llama3.1-8b, Llama3.1-70b, and Qwen2.5-7b. We use three datasets to assess performance, with results from the Natural Questions dataset demonstrating up to an 80% reduction in attack success rates while maintaining minimal loss in accuracy.

Matúš Falis, Aryo Pradipta Gema, Hang Dong et al.

Objective: To investigate GPT-3.5 in generating and coding medical documents with ICD-10 codes for data augmentation on low-resources labels. Materials and Methods: Employing GPT-3.5 we generated and coded 9,606 discharge summaries based on lists of ICD-10 code descriptions of patients with infrequent (generation) codes within the MIMIC-IV dataset. Combined with the baseline training set, this formed an augmented training set. Neural coding models were trained on baseline and augmented data and evaluated on a MIMIC-IV test set. We report micro- and macro-F1 scores on the full codeset, generation codes, and their families. Weak Hierarchical Confusion Matrices were employed to determine within-family and outside-of-family coding errors in the latter codesets. The coding performance of GPT-3.5 was evaluated both on prompt-guided self-generated data and real MIMIC-IV data. Clinical professionals evaluated the clinical acceptability of the generated documents. Results: Augmentation slightly hinders the overall performance of the models but improves performance for the generation candidate codes and their families, including one unseen in the baseline training data. Augmented models display lower out-of-family error rates. GPT-3.5 can identify ICD-10 codes by the prompted descriptions, but performs poorly on real data. Evaluators note the correctness of generated concepts while suffering in variety, supporting information, and narrative. Discussion and Conclusion: GPT-3.5 alone is unsuitable for ICD-10 coding. Augmentation positively affects generation code families but mainly benefits codes with existing examples. Augmentation reduces out-of-family errors. Discharge summaries generated by GPT-3.5 state prompted concepts correctly but lack variety, and authenticity in narratives. They are unsuitable for clinical practice.