Searcharxiv⌕ Search

arXiv subjects

David I. Atkinson

Publications and source records attributed to David I. Atkinson.

3 recordsLinked to original sources

Identifying Introspection From the Inside

Large language models make claims about themselves that are both consequential and increasingly difficult to verify from behavior alone. How can we distinguish plausible confabulations from genuine introspection? In this paper, we identify mechanistic signatures of faithful self-report in a controlled setting. Using low-rank adapters, we train models to make decisions on behalf of fictitious characters, according to latent linear preference functions. We find sustained fine-tuning on an implicit decision task can lead to the emergence of accurate self-reporting of models' learned preferences, even without explicit self-report supervision. We ask two research questions about this emergent phenomenon. First: is the emergence of accurate self-reporting accompanied by a measurable structural change in the model? Weight ablations and frozen-layer experiments together indicate that preference representations shift to earlier layers over training, consistent with the hypothesis that faithful self-report requires preferences to be located where pre-existing verbalization mechanisms can access them. Second: can these structural differences distinguish faithful models from unfaithful ones? Using attribution patching, we find that faithful models exhibit significantly higher attribution similarity between the decision-making and self-report tasks -- a mechanistic signature of faithful self-report that does not require us to understand the content of the report itself. Previous work on self-report has observed behaviorally that models can be faithful or unfaithful; our work proposes that, at least in our restricted setting, it is possible to distinguish between the two patterns of computation by examining the structure of the networks themselves.

cs.CL↗

ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Models

Large Multimodal Models (LMMs) exhibit shortfalls when interpreting images and, by some measures, have poorer spatial cognition than young children or animals. Despite this, they attain high scores on many popular visual benchmarks, with headroom rapidly eroded by model progress. This creates a need for difficult benchmarks that remain relevant for longer. We introduce ZeroBench - a lightweight visual reasoning benchmark curated using adversarial filtering to be "impossible" for frontier LMMs at its original release, with initial SotA scores of 0% pass@1 and pass^5. We track progress on ZeroBench over the subsequent year, observing SotA reaching 6% pass^5 and 19% pass@5, indicating the potential longevity of the benchmark. We evaluate 46 LMMs on ZeroBench, compare performance to a human baseline, analyse strengths and weaknesses, chart a year of progress in visual capabilities, and publicly release ZeroBench at https://zerobench.github.io.

cs.CV↗

Government AI Use as a Monitoring Primitive: A Public Document Pilot Study

Governments are important actors in frontier AI governance, but many facts about their adoption and use of AI systems are difficult to observe directly. Procurement disclosures and official statements are useful, but can also be delayed, selective, and better suited to measuring formal adoption than actual day-to-day use. We propose a complementary monitoring primitive: measuring traces of language-model assistance in public government documents. The approach is lightweight, externally reproducible, and based on revealed behavior rather than stated intent. In a pilot study of ten public document streams from U.S. and PRC government-related sources, we find that, while 2021 baselines are consistently near zero, by 2026, four of our ten sources show statistically significant signs of AI-assisted writing. In our sample, the U.S. signal concentrates in publications downstream of policy work; the PRC signal concentrates closer to it. We close by discussing how this signal could complement existing instruments for monitoring government AI adoption, and where it falls short.

cs.CY↗