From the arXiv new submissions of the past 24 to 48 hours in cs.AI, cs.LG, and cs.CL, this digest selects five preprints that concretely touch training, evaluation, and self-understanding of AI systems: how compute can be allocated more sensibly during reinforcement learning, how far a fully open model can keep pace with giants, how fragile the identity of deployed agents actually is – and how error diagnosis and latent thinking can be specifically improved. Curation was based on comprehensible methodology in the abstract, concrete core results, and thematic diversity instead of fivefold repetition of the same subfield.
Training and Open Models
Reinforcement learning favors easy problems – a trick aims to fix that
Michael Noukhovitch, Hamish Ivison, Nathan Lambert, and Aaron Courville describe in their paper a “Matthew Effect” in reinforcement learning training of language models: RL training improves problems a model can already solve well far more than hard problems – standard methods, the authors argue, waste too much compute on cases that are already solved. Their countermeasure, “Never Give Up” (NGU), keeps generating new attempts for a given problem until one is correct, automatically shifting more compute toward difficult cases. On the math benchmark Deepscaler, this reportedly improves performance per unit of compute noticeably, especially on harder problems; on the coding task Manufactoria, standard GRPO fails to fully solve problems with a mix of easy and hard tests, while NGU iteratively works its way to a complete solution. This matters because compute is one of the largest cost factors in RL training of language models, and a more targeted allocation could lower these costs without sacrificing training quality – a related pattern had already emerged in an earlier chess study on how far pretraining scale predetermines what RL fine-tuning can still extract.
ZGCM-1: a fully open 7B model takes on giants
A 22-person author team led by Jiyan He and Guang Liang releases ZGCM-1, a dense 7-billion-parameter model trained from scratch that combines compact model size with active tool use instead of relying purely on memorized facts. Through progressive context scaling up to 256,000 tokens, an architecture combining layered sliding-window and full attention, and a stable FP8 optimizer, the model reportedly achieves roughly 4.2 times the training efficiency compared to the prior approach at a 16,000-token context. On demanding mathematical reasoning and agentic search tasks, ZGCM-1 reportedly remains competitive with models orders of magnitude larger, such as Qwen3-235B-A22B and GLM-5.1. The team releases not just the final weights but also intermediate checkpoints, training code, per-stage data, and complete training logs – a rare degree of openness that substantially eases external replication and follow-up research.
Agents: Identity and Error Search
When agents know their identity but don’t act on it
Zhenyu Zhao and Roy Zhao introduce PAI-Bench, a provider-neutral test that clearly separates identity recall from identity actually enacted in behavior – for instance, whether an agent can merely name an assigned role or also act consistently in accordance with it. In an audited dataset of 1,536 responses, the tested systems correctly named their parent identity in all 48 direct queries, but did so in implicit self-portraits only once. Even small changes to task phrasing had large effects: explicit form-field cues raised the joint presence of three identity identifiers from 0 to 7 out of 8 profiles, and a simple name substitution in the startup configuration raised it from 1 to 7 out of 8. On identical responses, scoring by the Claude model also came out 12.5 percentage points lower on average than scoring by Astra – a sign that not only agent behavior but also the judging model itself affects the measured strength of identity. This matters because companies increasingly rely on stable, recognizable agent personas without systematically checking whether that persona actually holds up in behavior rather than just on direct questioning – a related picture had already emerged from an earlier paper on a controllable persona subspace in language models.
Iterative re-querying improves root-cause search in agent logs by more than 40 percent
Harsh Raj, David Lee, and an eight-person further author team address, with Continual Search, a practical problem in operating AI agents: the longer an execution log gets, the more likely a one-shot LLM judge overlooks the actual root cause and settles prematurely on a plausible but wrong diagnosis. Continual Search instead has the judge keep searching across multiple rounds for still-unresolved diagnostic evidence. On the newly introduced MegaRCA-Mix benchmark, comprising 50 human-annotated failure cases from long, execution-heavy tasks, the method reportedly raises GPT-5.5’s F1 score by more than 40 percent, from 0.349 to 0.498; weaker-tier models using the method can even outperform stronger-tier models that don’t use this search strategy. This matters because companies increasingly deploy AI agents for long, multi-step tasks, and error diagnosis in the resulting logs has so far often been manual or only superficially automated – an area where an earlier digest post already presented a targeted improvement in root-cause attribution within agent teams of up to 8.27 percentage points.
Why latent reasoning states should retain more instead of forgetting
Hongyu Gu, Chang Liu, and Jingwen Fu examine, in a theoretical paper (arXiv:2609.13747), how language models using continuous thinking not translated into tokens – as in recurrent or latent reasoning methods – should store their prior intermediate computations in an internal state. The intuitive assumption of keeping only the currently most plausible alternatives and discarding older intermediate steps turns out, according to the authors, not to be optimal: under identical compute, a state that retains the full reasoning history in superposition can require a lower representational dimension than one holding only the most recent alternative – because informative, mutually consistent intermediate steps reinforce each other, while unrelated alternatives merely add as interference. The authors also prove mathematically that weighting all stored intermediate steps uniformly is more robust for future reasoning than favoring more recent or more salient steps. This matters because latent reasoning is seen as a compute-saving alternative to lengthy chain-of-thought sequences, and the design principles derived here could influence how efficiently and reliably future architectures handle limited internal memory.
All five papers are unreviewed preprints published in the past few days; the figures cited here come from the authors’ own abstracts and experiments and have not yet been confirmed through independent review or replication. Whether ZGCM-1’s efficiency gains hold up at even larger scale, whether the 40 percent improvement in root-cause search transfers to real production environments, and whether the PAI-Bench findings appear this clearly across other model families as well, will only be shown by independent replications and peer review.
