Subject area
16 records across all collections
Preference-based fine-tuning is expensive to study because reward models are large. We show that a reward model distilled to 1% of the size of the original recovers 94% of its ranking accuracy on held-out preference pairs, and that policies trained against the small model transfer to the large one with a modest loss. The result lowers the cost of reward-modelling experiments for groups with limited compute.
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.
Marcus Tan, Sofia Lindqvist, Lena Hartwell
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.
A practical and rigorous guide to measuring what agents actually do. The book develops a measurement framework for tool-using language agents, covers reliability, cost and safety metrics, and shows how to design benchmarks, report variance and avoid the most common evaluation mistakes. Worked examples and code accompany every chapter.
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.
Agents that retry failed tool calls recover from transient errors but can also compound mistakes and waste resources. We analyse 18,000 recorded agent runs and show that retries help for network and rate-limit errors but hurt for malformed arguments. We propose a retry budget that depends on the error class and show that it improves success by 6 points at equal cost.
Some tool failures return plausible but wrong output and raise no error. We detect such silent failures by comparing the distribution of tool outputs against a rolling reference and flagging shifts. On three deployed tools the detector finds 71% of injected faults with a 2% false-alarm rate.
Priya Venkataraman, Daniel Okafor
An introductory textbook on adapting large pretrained models cheaply. Starting from the linear algebra of low-rank updates, it covers adapters, prompt tuning, quantisation-aware training and deployment, with exercises and solutions for each chapter.
Slides of the opening keynote at the 1st Workshop on Reliable Tool-Using Agents, 12 June 2026.
Kofi Mensah, Marcus Tan, Ingrid Novak
Existing agent benchmarks measure whether a task is solved once, not whether it is solved reliably. CTM-Bench contains 1,200 tool-use tasks, each with five perturbed variants (rephrased instructions, reordered tool lists, injected API errors, latency spikes and truncated outputs) and reports pass^k, the probability that an agent succeeds on all k independent attempts. We document collection, annotation and licensing following datasheet guidelines, evaluate six open models, and find that pass^1 overstates reliability: the best model falls from 71% pass^1 to 38% pass^5. The benchmark, scoring code and the datasheet are released under CC BY 4.0.
Marcus Tan, Kofi Mensah, Lena Hartwell
1,200 tool-use tasks, each with five perturbed variants, together with the evaluation harness, a datasheet and the annotation guidelines. Accompanies the article “CTM-Bench: A Benchmark for Tool-Use Reliability in Language Agents”.
Kofi Mensah, Ingrid Novak, Lena Hartwell
As language-model agents move from demonstrations to deployments, the way they are evaluated determines which claims can be trusted. We review how agents are evaluated in the literature, organising practice along four axes: the task environment, the success criterion, the treatment of variance, and the reporting of cost. We identify recurring weaknesses: success is often judged by another model without validation, runs are rarely repeated, cost and latency are reported in under a third of studies, and environment versions are seldom pinned. We close with a checklist for evaluation reports and open problems, including measurement of safety-relevant side effects.
Priya Venkataraman, Daniel Okafor
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.
Lena Hartwell, Tomás Ibarra, Marcus Tan
Language agents that act through external tools are limited by the size of their prompt: listing hundreds of tool descriptions crowds out the task and degrades accuracy. We study tool selection as a retrieval problem and propose a two-stage selector that (i) retrieves a short list of candidate tools using an embedding of the evolving task state and (ii) lets the model re-rank candidates with a single constrained call. Across a synthetic suite of 4,000 tools and three long-horizon task families, the selector keeps task success within 2.1 points of an oracle that sees only the correct tools, while reducing prompt length by 87%. Ablations show that conditioning retrieval on the task state, rather than on the initial instruction alone, accounts for most of the gain. We discuss failure cases, including tools with overlapping descriptions, and release the evaluation harness.
Source code for the two-stage tool selector and the evaluation harness used in “Retrieval-Augmented Tool Selection for Long-Horizon Language Agents”. Includes configuration files to reproduce every table.
A registry of 4,000 synthetic tools with descriptions, argument schemas and three task families, used to study tool selection at scale. Also includes a small registry of 600 real API descriptions.
