Today’s selection from the arXiv new submissions of the past days connects four questions that rarely arise together: how reliable alleged explanations for AI decisions really are, whether negotiation skills can be specifically trained in small models, how once stored factual knowledge can be specifically deleted, and how AI systems can research more efficiently themselves. The curation was done based on substance and dispersion: each paper provides a comprehensible method with robust numbers in the abstract, and no two works address the same subfield. Since arXiv does not publish new listings on weekends, the selection mostly comes from the most recent available issue from Thursday and Friday.
Explainability Evidence for AI: Disagreement Even on Acceptable Method Choices
Ajay Pravin Mahale examines in Explanation Multiplicity whether mechanistic interpretability – the search for the internal circuits that are supposed to explain a specific AI decision – delivers robust results when analysts make equally acceptable methodological choices. The small language model GPT-2 is tested on the task of indirect object identification, with a grid of seven independently varied analysis degrees of freedom, generating 15,840 specifications and 7,561 evaluable explanation statements. The author reports that derived explanation statements flipped from true to false in 73.2 percent of all specification pairs depending on the chosen method combination; even when only the evaluation metric is standardized, the flip rate remains at 59.4 percent, and the average overlap between two identified circuits as explanations is only 4 percent. This matters because interpretability findings are increasingly intended to serve as evidence documents for regulatory authorities – the study shows that the same model can provide almost arbitrarily different “explanations” depending on the acceptable analysis choice.
SocialRL: A 4-Billion Model Negotiates Like the GPT-5 Family
Wenyue Hua and a team that includes researchers from Microsoft Research demonstrate with SocialRL that strategic negotiation skills can be specifically trained in small language models. Their training procedure combines reinforcement learning with an optional theory-of-mind framework – an explicit modeling of the presumed goals of the opposing side – and is tested on a 4-billion-parameter model across six negotiation domains, from simple haggling games to salary negotiations and scheduling. The authors report that the trained model achieves or exceeds the GPT-5 model family in each domain, closing 73 to 122 percent of the gap between the baseline model and reference systems; before training, buyer agents only opened below the actual target value in 3 percent of cases, afterward in 78 percent. On average across all environments, the trained model achieves a utility value of 0.627, placing it ahead of GPT-4.1 (0.625), GPT-5.1 (0.619), and GPT-5.2 (0.613). This matters because it shows that no frontier model is apparently needed for negotiation-strong AI assistants, but targeted training of a small, cost-effective model is sufficient.
AdaPop: Popular Facts Are Harder to Delete – and Now Get Their Own Treatment
Anna Borisiuk, Andrey Savchenko, Alexander Panchenko, and Elena Tutubalina address with AdaPop a well-known problem in machine unlearning: popular, often repeated facts are harder to completely remove from a language model than rare ones. Their method combines local token confidence with an exponent dependent on the popularity of a fact and an automatic adjustment, tested on three model families across two benchmarks. The authors report that AdaPop reduces the re-emergence of actually deleted content in reformulated queries by about five times and in specifically adversarially reformulated queries by about 1.6 times compared to comparison methods, while the model capabilities that are supposed to be retained hardly deteriorate. A previous digest issue had already shown that quantization does not reliably delete memorized training data despite reduced storage requirements – AdaPop approaches the related problem from the other side: targeted deletion, which often fails for the most popular, persistent facts. This matters because companies may increasingly be legally required to remove personal or copyright-protected training data from an already trained model upon request.
Research Agents Save Computing Time: Prediction Models Instead of Full Test Runs
A large author team led by Thomas Simon Foster presents with AI Research Preference Models a way to alleviate the most expensive bottleneck of autonomous AI research agents: a candidate experiment can be designed in minutes, but its complete evaluation can take hours to days of GPU time. The presented preference models (RPMs) predict in advance which candidate solutions deserve a complete, costly evaluation without having to perform that evaluation themselves. In tests on the AIRS-Bench benchmark, the best selection approach improved the average normalized score from 0.684 to 0.729, achieving the typical result level after 24 hours already after about 15 hours with less than two-thirds of the otherwise necessary execution budget, with best values on two of the benchmark sub-tasks. A previous digest contribution had already shown that top agents master the technique of unpublished research questions but fail at actual scientific progress – the new work addresses this more deeply: not at the research idea itself, but at the question of which ideas actually deserve the scarce computing budget. This matters because rising computing costs are becoming the limiting factor for autonomous AI research, and every saving in evaluation directly means more tested ideas per computing budget.
All four works are not independently peer-reviewed preprints – the reported numbers come from the experiments of the respective author teams and have not yet been externally replicated. This is especially true for the explainability study, which has so far demonstrated its findings only on a single, comparatively small model and a single task, as well as for SocialRL, whose negotiation domains are simulated games rather than real business deals. Whether the findings can be confirmed on further models, tasks, and in independent replications remains to be seen.
