From the arXiv new submissions of the past 24 to 48 hours in cs.AI, cs.CL, and cs.LG, this digest selects four preprints that demonstrate how efficiency, interpretability, and safety behavior of today’s AI systems are tested in very different ways: a significantly faster table foundation model, an inference engine that runs a 35-billion-parameter model on ordinary consumer hardware, a methodological critique of a widely used interpretability tool, and a benchmark that tests whether top models actively seek safety evidence before making risky decisions. Curation was based on transparent methodology in the abstract, concrete core results, and thematic diversity rather than fivefold repetition of the same subfield.
Models and Efficiency
TabPFN-3.5: a significantly faster table foundation model
A 47-member author team led by Benjamin Jäger – including Yann LeCun and Bernhard Schölkopf – presents the new generation of the table foundation model from Prior Labs with TabPFN-3.5, which SAP acquired for over one billion euros last summer, as reported in an earlier article on the acquisition. According to the authors, the model sets a new record on the TabArena benchmark and extends capabilities to practical cases such as non-i.i.d. data, temporal or grouped data splits, mixed tables with text and images, and very wide tables with many columns. The faster variant TabPFN-3.5-Fast runs up to three times faster than its predecessor TabPFN-3 while largely retaining accuracy gains, while the more elaborate thinking variant TabPFN-3.5-Thinking works up to twelve times faster than its predecessor thanks to improved inference-time computation methods. This matters because table data is far more widespread in companies than plain text, and a more powerful yet faster base model would have direct practical consequences for data science teams.
A 35-billion-parameter model runs on a single consumer graphics card
Yu Lin, Yiming Wang, and three other authors address a central obstacle to running large mixture-of-experts models (MoE, models that activate only part of their parameters per request) locally with Edge0: even if only a few expert modules are needed per token, their weights must sit fully in memory, and naive reloading from disk fails because the appropriate experts are only known after the previous layer has already been computed. The authors solve this with a “pre-router” that predicts the next layer’s routing decision one token in advance and takes this prediction directly as the actual routing, supplemented by a lean correction component that compensates for quality losses from 4-bit quantization. According to the authors, on a single graphics card with 24 gigabytes of memory the system runs a 35-billion-parameter MoE model at 20 tokens per second using only 3 gigabytes of active memory, staying on average just a few points behind the uncompressed baseline model across five public benchmarks; framework, checkpoints, and adapters are open source. This matters because it could make large MoE models practically usable for the first time on ordinary consumer hardware instead of data centers – an efficiency problem that increasingly arises with open MoE models like the 295-billion-parameter model Tencent Hy3 even outside the major providers.
Interpretability and Safety
When probes only show what can be read – not what actually works
Devesh Tiwari, Camille Davis, and four other authors present a fundamental methodological critique of linear probes in their work, a widely used tool for reading safety-relevant concepts such as truthfulness from a language model’s internal activations. The authors show that the mere geometric alignment of a probe does not prove that the corresponding features actually causally drive model behavior, and instead propose a diagnostic that decomposes a truth probe into sparse autoencoder features (interpretable, individually nameable building blocks of internal model activation) and ranks them both by probe alignment and by their actual effect on model behavior. On an established truth probe for the model Gemma2-9B-Instruct, the authors report that the two rankings overlapped only weakly (about 12 percent, Spearman’s rho 0.10): features shared with the probe flipped the model output in up to 27 percent of cases, whereas equally sized, merely geometrically matching features did so in only 6 percent and random features in 1 percent. This matters because truth and deception probes are increasingly discussed as a safety tool, even though, according to this study, the most obvious evaluation method can miss the features that actually matter – a pattern that matches an earlier digest finding on truth probes in a reward-trained model, where honesty and mere rule-following also could not be cleanly separated.
Do frontier models actively seek safety evidence before they act?
Omer Tafveez investigates with the new benchmark SAFE whether language models actively seek out safety-relevant information before making a deployment decision, rather than merely reacting to warnings already present. In controlled test scenarios with varying costs, probabilities, severity, and presentation of available evidence, the four tested models GPT-5.5, o3, Claude Opus 4.8, and Claude Sonnet 4.6 showed markedly different evidence-acquisition policies according to the author: Opus inspects nearly by default, o3 is the most skip-heavy and threshold-sensitive, while the other models occupy intermediate positions. Notably, willingness to inspect tracked severity and retrieval cost strongly, while pure probability had little behavioral influence – even raising the stated likelihood of a problem from 10 to 70 percent changed inspection by at most 21 percentage points – even though the models frequently cited this probability in their rationales. This matters because deployment decisions by AI agents are increasingly made automatically, and the willingness to actively seek evidence before risky steps is arguably at least as important as how models handle already-known risks – a related tension between premature restraint and premature action was already highlighted in an earlier digest piece on agents at critical action boundaries.
All four works are unreviewed preprints published in the past one to two days; the figures cited come from the abstracts and statements of the respective author teams and have not yet been confirmed by independent review or replication. Whether Edge0’s efficiency gains carry over to even larger MoE models, whether the weak probe-causality overlap also occurs in other model families and concepts, and whether the strikingly different testing strategies of the SAFE-tested models hold up in real deployment scenarios will only be shown by independent replications and peer review.
