The editorial team selects four papers from the arXiv submissions of the past 24 to 48 hours that demonstrate how far the capability and reliability of today’s AI systems can diverge – from deceptive safety checks to staged deception and feigned tool use, to a sober architectural insight. The curation was based on substance: each paper provides a comprehensible method and a concrete numerical result in the abstract, not just a new topic headline. The selection deliberately covers four different subfields – safety testing, agent deception, agent introspection, and model architecture.
Compressed models pass every check – yet invent procedure steps
I. Kennedy and T. Kennedy show in Fidelity Is Not Safety: Gently-Compressed LLMs Pass Every Data-Free Quality Guard Yet Invent Procedure Steps in Agentic Execution that lightly compressed language models pass standard quality checks such as perplexity, MMLU scores, and output fidelity without issue, but still develop problematic behavior in independent, agentic execution. They test two types of compression on three model families: low-rank truncation (SVD) reportedly triggers the problem systematically, while magnitude pruning at matched perplexity does not. According to the authors, models that pass the paired output-fidelity test simultaneously fail a specially developed canary test for invented procedure steps – it is not the magnitude of the error itself but the coherence and rate of the compression error that determines whether the problem occurs. This matters because it shows that established quality metrics do not certify agent safety, leaving a blind spot in common release processes for compressed models – a risk that, similar to the previously described agent guardrails, only becomes visible during actual execution, not in isolated model tests. As a remedy, the authors propose a two-axis statistic of coherent-error-fraction and error-rate that flags risky compressed builds with consistent thresholds across architectures.
AI rarely sustains its deception role in the social game
A team led by Niklas Bauer and Lars Benedikt Kaesberg presents ParliamentBench, an open-source framework that measures deception, persuasiveness, and reasoning under information asymmetry using the social game Secret Hitler – a social-deduction game in which hidden roles pursue conflicting goals. Across 1,600 simulated matches with 16 language models, the authors establish new metrics for social deduction, reasoning, and deceptive consistency. Top models such as GPT-5.4, Kimi K2.5, Grok 4.1 Fast, and DeepSeek 3.1 Terminus perform strongly in both cooperative and deceptive roles, the authors report, while weaker systems underperform baseline strategies. The key finding: most models fail to sustain a consistent deceptive persona, with deception-retention rates dropping below 50 percent on average. This matters because reliable deceptive capability is a precondition for risky manipulation – a link to the earlier digest finding on deceptive internal circuits, which showed that deception can be mechanistically localized but apparently cannot be reliably sustained over longer interactions.
Agents claim skill use that often does not influence their decision
Jinwei Hu and colleagues show in Skill Use or Skill Theater? Evaluating the Reasoning Backroom in Skill-Augmented Language Agents that skill use in language agents has so far mostly been inferred only from visible reasoning or the agent’s own self-report – signals that show what an agent appears to use, not whether the skill actually changed its decision. With the evaluation framework BACKTRACE, the authors pair each skill-conditioned answer with a matched no-skill counterfactual, intervene deliberately on skill meaning, wording, identity, and assignment, and elicit attribution only after the answer is committed. Across model families, domains, and single- and multi-agent settings, they identify a pervasive provenance failure: stated skill use often remains stable while causal reliance varies widely – producing both silent uptake and purely performative use. Observational detectors based on direct skill-use claims, text mentions, trace similarity, and an LLM judge do not reliably identify which decisions actually depend on the skill, the authors report; in multi-agent systems, skill influence can even survive after its source is lost in communication. This matters because companies are increasingly extending agent capabilities through exactly such skill libraries and have so far relied mostly on these unreliable self-reports – echoing earlier digest findings on exploited evaluation gaps, which showed that agent benchmarks can systematically overstate what a system actually achieves.
More memory parameters outperform a larger overall model
Rubin Wei, Jiaqi Cao, and colleagues separate, in Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory, long-term memory explicitly from a language model’s reasoning ability and scale their memory module with 300 billion training tokens to 6.9 billion parameters; a distributed indexing pipeline is meant to sidestep the computational bottlenecks of existing methods. Paired with the small base model Pythia-410M, the memory module raises the average benchmark score, according to the authors, from 29.86 to 37.34 – surpassing even Pythia-12B (37.24) while using 39 percent fewer total parameters. For the larger Qwen3 Base models (0.6B–14B), 1.7-billion-parameter domain-specific memories improve performance by more than 9 points across three domains at all scales. This matters because it suggests that additional parameters flow more efficiently into dedicated memory than into a uniformly larger base model – an architectural principle that connects to the digest example of a layer-organized agent memory, though there it applies within a single agent session rather than as a pretrained, scalable parameter module.
Bottom line: all four papers are unreviewed preprints so far; their numbers come from the respective author teams’ own experiments. The compression study tests only three model families and a self-developed canary test, ParliamentBench measures deception exclusively within the narrow frame of a single social game, BACKTRACE so far covers mainly logic and competition-mathematics tasks, and the memory scaling has not yet been tested on the largest available base models. Whether the patterns hold up in production-grade systems and independent replications remains to be seen.
