Retrieval-Augmented ICL

In-context retrieval (ICR) is a retrieval-based form of in-context learning (ICL) in which demonstrations are retrieved from a source database based on similarity to the query, rather than sampled independently. In this work, we formulate ICR as a type of domain adaptation problem, where the source distribution $P$ of the database may differ from the target distribution $Q$ of the test query-label pair. We investigate the performance of ICR under a flexible class of distributional shifts that substantially extends prior work \citep{li2024fine,guo2025retrieval}, and establish theoretical guarantees that quantify the benefits and pitfalls of this learning paradigm. Our theory is verified by experiments on synthetic and language tasks.

Vision-language-action (VLA) models have shown strong potential as generalist robot policies, but adapting them to unseen tasks often requires costly parameter updates. Recent work such as RICL introduces in-context adaptability by retrieving expert demonstrations based on the current VLA observation and providing them as additional context at test time. The effectiveness of this adaptation therefore depends critically on the retrieval mechanism. In this work, we systematically study how different retrieval methods affect both retrieval quality and task performance within the RICL framework. Specifically, we compare four different methods: image-based retrieval, retrieval augmented with VLA's state, retrieval using features from the VLA backbone, and random retrieval. Our experiments yield three main findings. First, no retrieval method consistently dominates the others in task success, while surprisingly, random retrieval achieves a non-trivial success rate. Second, standard retrieval-quality diagnostics do not reliably reflect downstream VLA performance. Third, demonstrations from different but related tasks can provide useful transferable information. Together, these results provide an initial step toward understanding how retrieval mechanisms shape the in-context learning capability of VLA models and their downstream task performance, while highlighting the need for more careful design and evaluation of retrieval mechanisms for reliable test-time adaptation.

In-context learning (ICL) is crucial for boosting the inference performance of large language models (LLMs). However, the effectiveness of ICL in LLMs is greatly influenced by the choice of demonstration sets. Exhaustive searches over these sets are combinatorial, and existing selectors often rely on relevance or likelihood proxies to implicitly assess ICL quality. Making repeated queries to the target LLM with these strategies can incur substantial costs. This work simplifies selection by framing it as a constrained local search problem and presents local demonstration editing (LDE). Starting with an initially retrieved set of demonstrations, LDE employs a single structured edit to explore its surrounding neighborhood while balancing performance gains with search costs. Technically, LDE is reduced to a policy search problem, for which we train a small LLM, referred to as Jev-LDE. This model as the System-1 modifies the retrieved demonstration set by performing actions such as \texttt{Keep}, \texttt{Delete}, or \texttt{Replace} elements, all within a framework of reinforcement learning with verifiable rewards. At test time, Jev-LDE executes a single edit of the retrieved demonstration set, followed by one inference from the target LLM, avoiding the need for iterative context scoring or subset searches. Across standard classification benchmarks, various target LLMs with Jev-LDE as the plug-and-play module consistently improve ICL performance, and Jev-LDE shows transferability to held-out benchmarks and models without retraining. These findings indicate that the LDE approach offers an efficient and adaptable method for harnessing the ICL capabilities of target LLMs.

In-context learning (ICL) based on large language models (LLMs) has shown promising potential in alleviating performance bottlenecks caused by the limited availability of annotated data in Named Entity Recognition (NER). However, existing methods still face issues of retrieval misalignment and generation uncertainty, making their performance heavily dependent on the LLM's capabilities. As the parameter scale of LLMs decreases, their performance in few-shot settings deteriorates significantly. In this paper, we propose a novel unified retrieval-augmented framework, URA-NER, including three key components: Progressive Granularity Retrieval (PGR), Model-aware Representation Enhancement (MaRE), and Reason-aware Knowledge Verification. PGR is a two-stage retrieval mechanism that achieves stage alignment. It first retrieves demonstrations for span detection based on the query's global semantics, and then for type classification based on the specific entity context, providing fine-grained local information. Moreover, MaRE employs entity pre-recognition to guide the construction of representations, ensuring the query and demonstrations are aligned within the LLM's semantic space and attention pattern. In addition, to mitigate generation uncertainty, we propose RaKV, a closed-loop "generation-retrieval-verification" process. It explicates the LLM's reasoning paths, leverages them for the retrieval of external knowledge, and reorganizes the knowledge into verification evidence aligned with the original reasoning paths. We conduct extensive experiments on multiple low-resource NER datasets. Results demonstrate that URA-NER significantly enhances the performance of LLMs under low-resource settings, with particularly pronounced gains for smaller LLMs, achieving new state-of-the-art results on several benchmarks.

Medical image interpretation is central to diagnosis and care, yet adapting general-purpose multimodal large language models (MLLMs) often requires resource-intensive domain-specific fine-tuning. Here we introduce representation-guided in-context learning (RG-ICL), a training-free inference framework that retrieves query-aligned demonstrations using frozen encoders, without task-specific parameter updates. Across eight datasets spanning histopathology, radiology and retinal fundoscopy, RG-ICL improved classification (mean gain 20 percentage points) and visual question answering (VQA) (mean gain 13 percentage points) over no-context and conventional ICL, approaching or exceeding training-based comparators. Which cases were retrieved mattered more than how many: 6 query-aligned cases outperformed up to 32 randomly selected ones, whereas fixed or random cases often reduced accuracy below baseline. For VQA, aligning reference cases with both image content and question intent produced further gains. These findings indicate that for medical image interpretation, curating which reference cases an MLLM sees is a practical alternative to retraining it.

To answer questions outside of their pre-training data, large language models (LLMs) need access to new information, which can be presented in the context window as documents, encoded into the model's parameters, or injected as latent representations. However, each of these methods comes with different efficiency, cost, and performance trade-offs, with no single winner. We present a controlled comparison of representation-based (KV-cache based) and parametric (fine-tuning-based) adaptation methods on five knowledge-intensive benchmarks. We show that in the oracle setting, Cartridges (KV) are the most accurate injection method at nearly every storage budget, outperforming parametric methods by 10 points. Compaction (KV) matches Cartridges only at low compression rates, lagging behind the parametric methods by 10 points at rates higher than $50\times$. In the more realistic multi-document retrieval scenario, Cartridges are the only method that matches in-context learning (ICL), leading the parametric methods by 29 points and Compaction by 15 points. Nonetheless, Cartridges are also the only method, besides full fine-tuning and large MLP adapters, that suffers from catastrophic forgetting, i.e., a 6% performance degradation on control benchmarks, with 13% in coding.

We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer \emph{semantics} rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0% to 57.5% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number $k$ of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5%) at both $k=0$ and the adapter's training-time $k=1$, 188/200 (94.0%) at the packaged runtime's default $k=3$ --- and the submitted confidence-gated override adds no net accuracy on top of $k=3$, selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~\cite{zhao2024ratescore} decrease as $k$ grows, but paired sign tests on token-F1 differences are nonsignificant ($p \ge 0.29$); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0% MCQ accuracy on the development holdout and 93.20% on the organizer's official pre-evaluation, versus 29.43% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4).

In-context learning (ICL) is widely used in multimodal large language models (MLLMs) and achieves strong performance across a wide range of multimodal tasks. However, existing multimodal ICL methods often rely on surface level imitation of in-context demonstrations, making it difficult for MLLMs to align their responses with the reasoning path required by the given multimodal input. This limitation becomes more pronounced in complex multimodal tasks, thereby restricting further improvements in MLLM performance. To address this issue, we propose a new multimodal ICL framework that combines contrastive demonstration modeling with the self-refinement capability of MLLMs. Specifically, our framework reformulates each demonstration by explicitly contrasting a suboptimal response with a better response under the same input, together with a reasoning path that reveals how the response should be refined. This contrastive formulation makes the reasoning path toward the desired response more explicit and guides the MLLM beyond superficial imitation. Furthermore, because effective refinement depends on the current response, we introduce a response-conditioned retrieval mechanism to select demonstrations whose reasoning paths are more relevant to the current response. In addition, we use a lightweight alignment controller to predict response quality and determine whether further refinement is needed. Experiments on three types of multimodal tasks show that the proposed framework consistently improves MLLM performance, with particularly notable gains on visual question answering (VQA).

With hate speech being ubiquitous online, automatic detection is crucial, in particular when it comes to criminally relevant social media posts. We study a variety of retrieval-based in-context learning (RetICL) strategies for detecting defamatory offences under §§ 185-187 StGB (the subject of GermEval 2026 Subtask 4). Few-shot prompting beats zero-shot, but retrieval-based approaches offer only marginal gains over random demonstrations, and even fall behind an optimised static set of demonstrations. Providing concrete legal knowledge helps, yet model choice outweighs every other system choice. Models over-predict criminal relevance while still missing 26-57% of criminally relevant posts, suiting them for triage rather than autonomous moderation.

The Asset Administration Shell (AAS) is a cornerstone of Industry 4.0 and the Digital Product Passport, providing standardized digital representations of industrial assets. While manufacturers already maintain extensive technical product documentation, generating AAS instances from existing product datasheets remains a labor-intensive task because technical information is extracted from heterogeneous document structures and often involves company-specific terminology and conventions. In this work, we present AAS-RAIL, a retrieval-augmented information extraction (IE) approach that automatically generates Asset Administration Shells from PDF product datasheets using large language models (LLMs). Instead of relying on a fixed set of few-shot examples, the proposed retrieval-augmented in-context learning (RAIL) approach retrieves LLM-generated extraction helpers from similar Asset Administration Shells to provide instance-specific in-context learning (ICL). This enables the model to adapt its extraction behavior to company-specific naming conventions and formatting styles without fine-tuning. Our core contribution is the dynamic selection of company-specific AAS examples for each datasheet, replacing static prompting with an extraction pipeline that adapts to instances and combines semantic retrieval and structured information extraction. The proposed approach is evaluated on a collection of industrial product datasheets using a selection of open- and closed-weight LLMs. Experimental results show that RAIL consistently improves extraction quality over conventional few-shot prompting, yielding relative improvements of 30.4-52.4%. These results demonstrate that our approach provides an effective improvement for company-specific AAS generation.

Retrieval-augmented forecasting promises to adapt frozen Time Series Foundation Models (TSFMs) to new domains without fine-tuning, but recent methods typically rely on learned fusion modules, i.e., trained adapters that merge retrieved examples into the backbone's forecast, based on the assumption that frozen backbones cannot dynamically incorporate retrieved context on their own. We show this assumption is unnecessary. We introduce Align-RAG, a training-free method that applies a closed-form per-pair amplitude rescaling and integer-lag phase shift to retrieved past-future windows before they enter a frozen backbone's context. With no learned parameters, Align-RAG outperforms the state-of-the-art trained retrieval adapter on a frozen Chronos-Bolt on all seven datasets of the standard benchmark (avg -3.75% MSE), showing that the gains previously attributed to learned fusion are recoverable without any training. Align-RAG further improves zero-shot MSE on four additional frozen TSFMs with various architectures by 2.5% to 13.7% per backbone with no per-backbone tuning. To probe why alignment helps, we compare the frozen backbone's prediction shift under aligned demonstrations to the closed-form ridge prediction shift on the same pairs. We find that aligned demonstrations induce prediction shifts that track a closed-form ridge predictor on the same pairs, with a future-shuffle control ruling out a futures-averaging account. Together, these results indicate that frozen TSFMs already support dynamic in-context use of retrievals, and that closed-form alignment should be the default baseline for retrieval-augmented forecasting before any fusion module is trained. Code available at: https://github.com/masadi-99/align-rag

On-device in-context learning (ICL) relies on pre-inference retrieval to select demonstrations for useful context before downstream model inference. This retrieval must exploit task-specific information while operating over local memories under limited computation, memory, and data-exposure budgets. We propose Conditional Retrieval Alignment (CoRA), a gradient-free framework that converts a frozen encoder into a task-conditioned retriever using paired candidate inputs and outputs. CoRA selects complementary encoder layers, constructs an output-derived conditioning space from candidate memory, and aligns candidate input representations to this space through closed-form ridge regression. Low-rank factorization then produces a compact retrieval basis where candidate outputs are used only during offline index construction, whereas query-time retrieval requires only the query input and precomputed index. We show that CoRA's rank-constrained basis is the optimal low-rank compression of the output-conditioned fitted representation, and derive an exact two-pass streaming construction that avoids materializing the full fitted matrix. We further extend the framework to multimodal exemplar retrieval by incorporating visual representations into the conditioning and retrieval spaces. Experiments across ten textual datasets and four multimodal benchmarks with Llama-3.2-1B, MobileLLM-Pro, OpenFlamingo-3B, and Qwen3.5-2B, as well as end-to-end Raspberry Pi~5 deployment demonstrate that CoRA supports effective task-conditioned retrieval without retriever fine-tuning, backpropagation, or target-model calls.

Human Activity Recognition (HAR) from wearable sensors supports applications in healthcare, rehabilitation, fitness tracking, and smart environments. Yet, existing deep learning approaches require dataset-specific training, large labeled corpora, and repeated adaptation to new sensor settings or activity taxonomies. Retrieval-Augmented Generation for Human Activity Recognition (RAG-HAR) addresses this by framing HAR as a training-free, retrieval-augmented task, in which statistical descriptions of sensor windows are used to retrieve similar labeled examples that guide LLM-based classification. We introduce RAG-HAR+, a retrieval-first and cost-optimized extension that strengthens retrieval while reducing dependence on LLM-based inference. RAG-HAR+ uses an offline Retrieval Designer Agent to design dataset-specific feature groups from a diverse pool of motion descriptors, enabling sensor windows to be compared using features better aligned with dataset-specific activity patterns. During inference, RAG-HAR+ uses majority voting over retrieved neighbors for samples with strong retrieval evidence and defers only uncertain cases to an LLM-based Ambiguity Resolver Agent. Across six HAR benchmarks, RAG-HAR+ maintains competitive or improved performance while reducing LLM usage, token consumption, and inference time. We further extend the RAG-HAR mobile prototype to demonstrate the practical feasibility of retrieval-first, LLM-assisted HAR in mobile sensing scenarios.

Wound image classification is often treated as a task-specific supervised learning problem, requiring substantial amounts of manually labelled data and retraining when the label space or deployment setting changes. This study evaluated whether small multimodal language models (SMLMs) can provide a training-free alternative for wound classification through retrieval-based in-context learning (ICL). Experiments used two public wound-image datasets: the Kaggle wound dataset (1469 images, 10 classes) and the Medetec dataset (560 images, 9 classes). Eleven SMLMs from the Qwen 3.5, Ministral 3, and Gemma 4 families were evaluated under zero-shot prompting and few-shot prompting with random support examples, embedding-based k-nearest-neighbour (kNN) retrieval, and kNN retrieval followed by maximal marginal relevance reranking (MMR). Retrieval-only weighted-kNN controls, support-set reduction experiments, and support-context size sweeps were used to assess the effects of retrieval, model scale, and prompt length. Query-conditioned ICL consistently outperformed zero-shot and random few-shot prompting. On the Kaggle dataset, the best result was achieved by Qwen 3.5 27B with kNN+MMR, reaching 0.872 accuracy and 0.871 F1 score. On Medetec, Qwen 3.5 27B with kNN+MMR reached 0.678 accuracy and 0.670 F1. Larger models exceeded matched weighted-kNN controls, indicating use of retrieved examples beyond nearest-neighbour voting. Retrieval-based ICL degraded modestly under support-set reduction, and most gains saturated with 8-10 support images. Retrieval-based ICL allows SMLMs to perform adaptable wound image classification without task-specific retraining. Compact retrieved contexts may support practical and privacy-conscious deployment, although performance remains dependent on model scale, retrieval strategy, and dataset difficulty.

In-Context Learning (ICL) has become a powerful mechanism for adapting Large Language Models (LLMs) to new tasks without fine-tuning. Extending this concept to Large Multimodal Models (LMMs), Multimodal In-Context Learning (M-ICL) relies on retrieving relevant examples, such as images, captions, or question-answer pairs, to guide predictions across tasks like classification, captioning, and visual question answering (VQA). Most existing approaches select in-context examples based on feature-space similarity, assuming that semantically similar samples provide the most useful context. However, our systematic analysis reveals that this assumption does not always hold: visually similar examples are not necessarily those that most effectively enhance in-context learning performance. To address this, we propose the Guided Retrieval of In-context Prompts (GRIP), a learnable vision-only retrieval framework that leverages feedback from LMMs to identify examples that truly improve model predictions. GRIP learns to distinguish beneficial from detrimental in-context examples through contrastive training, refining retrieval beyond pure similarity. Across three multimodal tasks, namely classification, captioning, and VQA, GRIP improves consistently over similarity-based retrieval on Qwen2.5-VL-7B, with its strongest gains in classification on Idefics2-8B. Moreover, we demonstrate that retrievers trained with feedback from one open LMM can be transferred to other models without retraining, including closed-source GPT-4o and Gemini, enabling scalable and cost-efficient deployment of M-ICL. Code will be published upon acceptance.

Understanding human states and interaction dynamics is a core goal of human-computer interaction (HCI). As interaction paradigms become more immersive, virtual reality (VR) has emerged as a powerful platform for studying collaborative work. In such settings, evaluating team collaboration states, including team performance and team resilience, requires continuous and reliable inference of latent team-level cognitive and affective states from multi-modal sensor data, such as speech signals. However, generating ground truth labels for these latent states remains challenging due to sensor-induced noise, contextual variability, and sparse expert annotations. Traditional self-reporting approaches provide only static and delayed measurements and are therefore insufficient for capturing dynamic team processes reflected in continuous speech data. In this work, we propose a large language model (LLM)-driven, agentic inference workflow for automated emotion-related synthetic ground truth generation from streaming speech data in multi-user VR environments. Leveraging the generalization capabilities of LLMs, we use In-Context Learning (ICL) with few-shot demonstrations of paired audio-based samples and their corresponding transcriptions. ICL tends to achieve task adaptation comparable to model fine-tuning while circumventing the computational overhead of parameter updates. To construct informative and robust in-context prompts, we adopt a retrieval-based selection strategy that dynamically identifies relevant audio demonstrations based on similarity in the acoustic feature space.

Wireless localization is a fundamental capability of sixth-generation (6G) networks. Conventional model-based methods require accurate modeling of the propagation environment and degrade in complex multipath and non-line-of-sight scenarios, while learning-based methods couple model parameters tightly to the training scene, requiring costly retraining whenever the base station (BS) configuration or propagation environment changes. In this paper, we propose RA-LWLM, a retrieval-augmented in-context localization framework that achieves training-free cross-scene adaptation by externalizing scene-specific information into a per-scene fingerprint database rather than encoding it in model weights. The framework consists of three components: a frozen wireless foundation model (FM) encoder that maps raw channel state information into a scene-agnostic representation; a retrieval module that selects the most informative references from the per-scene database via similarity search in the representation space; and a transformer-based in-context learning (ICL) module that fuses the query with the retrieved references to predict the user equipment (UE) position. To accommodate varying retrieval quality and propagation complexity across queries, the ICL module adopts a mixture-of-experts design in which experts specialize in different context sizes and are softly combined by a learnable selector. Extensive ray-tracing-based experiments across heterogeneous scenes with diverse BS configurations show that RA-LWLM achieves nearly identical accuracy on seen and unseen scenes without any per-scene retraining, substantially outperforming end-to-end and FM-based baselines. These results validate the proposed retrieval-augmented in-context paradigm as a scalable solution for cross-scene localization in 6G networks.

Hallucinations in medical large language models (LLMs) pose serious risks for clinical decision support, particularly when models must reason over complex electronic health records (EHRs). However, existing benchmarks often lack a realistic clinical context and provide limited insight into how hallucinations can be mitigated in practice. We introduce Med-HEAL, a framework for systematically identifying, analyzing, and mitigating hallucinations in medical LLMs using clinically grounded data. Building on the EHRNoteQA benchmark derived from MIMIC-IV discharge summaries, we construct a hallucination dataset by evaluating BioMistral-7B on open-ended clinical question answering tasks. Model outputs are labeled through a dual evaluation pipeline that combines LLM-as-a-Judge assessment (GPT-4o) with human auditing by medical student reviewers, producing correctness judgments and annotations of reasoning errors via a custom web-based evaluation system. We then leverage this dataset to investigate mitigation strategies: a self-critique pipeline, in which the test model reviews its own answers to detect potential errors and regenerates responses for flagged cases, and retrieval-augmented in-context learning (RA-ICL), which exposes the model to hallucinated and corrected examples. Experiments across five open-source LLMs-BioMistral, Llama-3.1, DeepSeek, Qwen2.5, and Qwen3, show that the self-critique strategy improves accuracy for three of five models (p < 0.05) without requiring parameter updates. Med-HEAL provides both a reusable hallucination dataset and a practical framework for studying and mitigating hallucinations in medical LLMs, supporting safer deployment of AI systems in clinical environments. Our code and data are publicly available at https://github.com/yimingliao-blad/med-heal.git.

In-context learning has recently been linked to implicit gradient descent in linear self-attention models, suggesting that context can induce a forward-pass update. Retrieval-augmented generation (RAG) also relies on context, but retrieved documents are usually treated as static evidence rather than signals for adaptation. We study RAG as an in-context optimization process. First, we show that one linear self-attention layer can implement one gradient-descent step on a unified linearized RAG objective covering both projection-based and dot-product retrieval interfaces. This gives an exact regime where retrieval-augmented prediction and in-context optimization coincide. We use this result not as a literal model of LLM computation, but as a guide for adapting the interaction between queries and retrieved evidence. We then test the boundary of this correspondence: it remains stable under controlled linear extensions, but becomes feature-distribution dependent under nonlinear architectures. Finally, we turn this view into a lightweight method for frozen RAG LLMs. The method keeps the retriever and backbone fixed, and predicts a context-conditioned update to a generator-side evidence-use interface. Across seven QA benchmarks, two retrievers, and two frozen LLM backbones, this forward-only update improves a shared-interface baseline, transfers to held-out tasks, and approaches test-time gradient adaptation at much lower per-query cost.

We show that remotely hosted applications employing in-context learning when augmented with a retrieval function to select in-context examples can be vulnerable to membership-inference attacks even when the service provider and users are separate parties. We propose two black-box membership inference attacks that exploit query text prefixes to distinguish member from non-member inputs. The first attack uses a reference model to estimate an otherwise unavailable loss metric. The second attack improves upon it by eliminating the reference model and instead computing a membership statistic through a simple but novel weighted-averaging scheme. Our comprehensive empirical evaluations consider a stricter case in which the adversary has a paraphrased version of the text in the queries and show that our attacks can exhibit stronger resilience to paraphrasing and outperform three prior attacks in many cases with small number of prefixes. We also adapt an existing ensemble prompting defense to our setting, demonstrating that it substantially mitigates the privacy leakage caused by our second attack.

Span-level skill extraction from job advertisements underpins candidate-job matching and labor-market analytics, yet generative large language models (LLMs) often yield malformed spans, boundary drift, and hallucinations, especially with long-tail terms and cross-domain shift. We present SRICL, an LLM-centric framework that combines semantic retrieval (SR), in-context learning (ICL), and supervised fine-tuning (SFT) with a deterministic verifier. SR pulls in-domain annotated sentences and definitions from ESCO to form format-constrained prompts that stabilize boundaries and handle coordination. SFT aligns output behavior, while the verifier enforces pairing, non-overlap, and BIO legality with minimal retries. On six public span-labeled corpora of job-ad sentences across sectors and languages, SRICL achieves substantial STRICT-F1 improvements over GPT-3.5 prompting baselines and sharply reduces invalid tags and hallucinated spans, enabling dependable sentence-level deployment in low-resource, multi-domain settings.

Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content -- distinct from its semantic relevance -- can independently distort the model's attention distribution. We identify and formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2-3x more attention per token than semantically equivalent natural-language text ($\hat{o}$(KG) $\approx$ 0.70 vs. $\hat{o}$(neutral) $\approx$ 0.25), compressing demonstration attention by up to 42% -- regardless of whether the triples are relevant or noise. We develop a formal framework decomposing attention scores into semantic and structural components (Eq. 2), derive a compression bound (Proposition 1) connecting token-level format bias to demonstration attention loss, and show that the structural term governs how much attention is diverted while the semantic term governs whether this helps or hurts. This decoupling reveals two orthogonal axes for improving retrieval-augmented ICL: optimising retrieval quality (semantic axis) and reducing format-driven attention capture (structural axis). Empirically, across two model families (Mistral-7B, LLaMA-3-8B) and three QA benchmarks, we observe that source-task alignment dominates: task-matched BM25 retrieval achieves 58-62% on HotpotQA vs. ConceptNet's 25-27%, a >30 pp gap that dwarfs all gating strategies ($\leq$2 pp). We derive five structure-aware mitigation strategies from the framework, ranging from zero-cost prompt modifications to training-time regularisation; format flattening (S3) is validated by both accuracy and attention-level evidence from a verbalized-triple control, while structural dispersal (S1) yields mixed results that illuminate the challenges of format-level intervention.

Recent work has demonstrated the potential of large language models (LLMs) for program optimization, a key challenge in programming languages. We propose a blackbox adaptation method called Retrieval Augmented Search (RAS) that performs beam search over candidate optimizations; at each step, it retrieves in-context examples from a given training dataset of slow-fast program pairs to guide the LLM. Critically, we find that performing contextual retrieval based on an LLM-generated natural language description significantly outperforms retrieval based on the source code. We also propose AEGIS, a method for improving interpretability by decomposing training examples into ''atomic edits'' that are significantly more incremental in nature. We show that RAS performs up to 2.06$\times$ better than prior state-of-the-art blackbox adaptation strategies on optimizing C++ programs, and that AEGIS performs up to 1.37$\times$ better while making significantly smaller edits. We also show that using RAS improves the mean runtime percentile of Python programs by 10.27 compared to baselines.