Search AEX CONTINUUM

Titles, abstracts, authors, keywords and DOIs across all five collections.

PreprintsResearch ArticleNot peer reviewed

Mining Failure Modes from Agent Trajectories at Scale

Version 1 · 21 October 2026 · 10.xxxx/ctm.pp.2026.0007

Agent developers inspect failed runs by hand, which does not scale beyond a few dozen trajectories. We propose a clustering approach that groups failures by the step at which the trajectory diverges from a successful reference run, and label the clusters with short natural-language descriptions. On 18,000 trajectories from three agent frameworks, the method recovers 83% of failure categories identified by manual analysis and surfaces four categories that were previously unreported.

Abstract PDF 2,065 6901 cited

PublishingResearch ArticleUnder peer review

Dense Retrieval for Code-Mixed Queries: A study of Hindi–English and Yoruba–English search

Oluwaseun Adebayo, Elias Wagner, Ravi Menon

Version 1 · 0 reviewer reports · 5 October 2026 · 10.xxxx/ctm.pb.2026.0009

Users frequently mix languages within a single query, but retrieval benchmarks rarely reflect this. We construct code-mixed query sets for two language pairs and evaluate dense retrievers under different tokenisation and fine-tuning strategies. Adapting the query encoder with synthetic code-mixed data improves recall@20 by 9 points over a multilingual baseline, while transliteration normalisation helps for romanised queries but hurts for native-script queries.

Abstract PDF 640 1880 cited

PreprintsResearch ArticleNot peer reviewed

Prompt-Injection Resilience in Browser-Embedded Assistants

Marcus Tan, Sofia Lindqvist, Lena Hartwell

Version 1 · 2 October 2026 · 10.xxxx/ctm.pp.2026.0006

Assistants embedded in web browsers read page content on the user's behalf, which exposes them to indirect prompt injection: instructions hidden in a page that the assistant mistakes for the user's. We build a test suite of 600 injected pages across seven attack styles and measure three defences: instruction-hierarchy prompting, a classifier that screens page text, and permission gating of sensitive actions. No single defence is sufficient; combining a screening classifier with gating of actions that write or send data lowers attack success from 62% to 4% while preserving 93% of benign task success.

Abstract PDF 912 2880 cited

PreprintsResearch ArticleNot peer reviewed

Tokenizer Effects on Low-Resource Language Adaptation

Oluwaseun Adebayo, Mei Tanaka

Version 1 · 12 September 2026 · 10.xxxx/ctm.pp.2026.0005

Adapting a pretrained model to a low-resource language often begins with extending the tokenizer, yet the choice of extension method is rarely compared. We evaluate five vocabulary-extension strategies on six languages and find that fertility reduction alone does not predict downstream accuracy. Strategies that initialise new embeddings from semantically related tokens consistently outperform random initialisation, at the same training budget.

Abstract PDF 1,890 6032 cited

PublishingResearch ArticleUnder peer review

Quantisation-Aware Distillation for On-Device Assistants

Daniel Okafor, Yuki Saito, Priya Venkataraman

Version 1 · 1 reviewer report · 28 August 2026 · 10.xxxx/ctm.pb.2026.0006

Running assistant-grade language models on phones requires both small models and low-precision arithmetic, yet distillation and quantisation are usually applied one after the other. We train a 1.3B-parameter student with distillation losses computed through simulated 4-bit quantisation, so that the student learns to be robust to its own rounding error. Compared with distilling first and quantising afterwards, the joint procedure recovers 41% of the accuracy lost to 4-bit quantisation, reaches 94% of the full-precision student on a suite of assistant tasks, and runs 3.1× faster on a mid-range mobile processor.

Abstract PDF 2,144 7862 cited

PublishingResearch ArticlePeer reviewed

Statistical Power in Agent Benchmarks: How Many Runs Are Enough?

Hana Kovács, Kofi Mensah, Carlos Medina

Version 2 · 3 reviewer reports · 14 August 2026 · 10.xxxx/ctm.pb.2026.0008

Most agent benchmarks report one run per task, yet agents are stochastic. We analyse run-to-run variance on four public agent benchmarks and derive the number of repeated runs needed to detect a given difference between two systems at conventional power. For differences smaller than five points, single-run comparisons have less than 40% power on all four benchmarks. We provide a sequential testing procedure that stops early when the evidence is clear, and show that it halves the cost of repeated evaluation while keeping the false-positive rate below 5%.

Abstract PDF 2,740 1,1033 cited

PublishingResearch ArticlePeer reviewed

Refusal Calibration in Instruction-Tuned Language Models: When should a model decline, and how well do current models know?

Nadia Haddad, Hana Kovács

Version 2 · 3 reviewer reports · 9 July 2026 · 10.xxxx/ctm.pb.2026.0007

Instruction-tuned models decline requests for many reasons, but little is known about how well their refusals track the actual risk of a request. We build a labelled set of 6,400 requests spanning benign, borderline and harmful categories and measure calibration of refusal decisions across eight open models. Refusal rates correlate only weakly with annotated risk for borderline requests, and over-refusal is concentrated in a handful of phrasing patterns rather than topics. A light-weight calibration step that uses the model's own confidence reduces over-refusal by 31% at no measurable cost to harmful-request refusal. We release the labelled set and the annotation guidelines.

Abstract PDF 3,180 1,2604 cited

PublishingResearch ArticlePeer reviewed

Scaling Behaviour of Parameter-Efficient Fine-Tuning Across Model Sizes

Priya Venkataraman, Daniel Okafor

Version 2 · 3 reviewer reports · 6 March 2026 · 10.xxxx/ctm.pb.2026.0002

Parameter-efficient fine-tuning (PEFT) methods adapt large models by updating a small fraction of weights, but most evidence comes from a single model size. We fit scaling curves for low-rank adaptation across six base-model sizes spanning two orders of magnitude and four downstream tasks. The optimal adapter rank grows sub-linearly with model width, and the gap between PEFT and full fine-tuning shrinks as models grow, closing to within 0.6 points at the largest size. A simple rule that sets the rank proportional to the square root of the hidden dimension recovers 96% of the best rank found by grid search. We provide fitted coefficients, discuss threats to validity, and note that results on instruction-following tasks differ from those on classification.

Abstract PDF 3,902 1,5445 cited