arXiv:2608.29461v1 Announce Type: new Abstract: Sandbagging models strategically underperform on evaluations while retaining the capabilities being measured. The evaluations that guide frontier-model...
arXiv:2609.25602v1 Announce Type: new Abstract: In language models, the choice between believing the prompt and believing the weights is made by a handful of identifiable attention heads. Instruction...
arXiv:2607. 25063v1 Announce Type: new Abstract: Developers judge a model checkpoint by how it behaves.
arXiv:2607. 05904v1 Announce Type: new Abstract: Training a language model against its own reference-free judgments (the premise of self-rewarding, self-play, and LLM-as-a-judge pipelines) assumes a model's verdict on a shown answer tracks correctness.
The paper introduces Drift-Constrained Optimization (DCO), a framework that treats behavioral drift during fine‑tuning of instruction models as a bounded constraint rather than an uncontrolled side effect. By defining a drift budget, the authors reformulate fine‑tuning as a direction‑selection problem, showing that choosing different update directions can qualitatively change outcomes. Experiments on Qwen3 models demonstrate that carefully selected directions improve scientific reasoning and multilingual translation while preserving reasoning capabilities and general performance.
arXiv:2607. 01480v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR), along with recent selfdistillation variants such as SDPO, evaluates each rollout against a verifier and updates the policy from that episode-level signal.
arXiv:2608.20873v1 Announce Type: new Abstract: Every way of teaching a deployed language model something new -- full fine-tuning, adapter merging, model editing -- replaces the released checkpoint,...
arXiv:2609.06934v1 Announce Type: cross Abstract: Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023b), fine-tuning attacks (Qi...
The paper investigates whether fine‑tuning a language model erases previously embedded activation steering interventions that suppress refusals and encourage brevity. Across five instruction‑tuned models (3B–14B) subjected to non‑adversarial supervised fine‑tuning (SFT) and reinforcement learning from human feedback (RLHF), the authors find that the steering’s behavioural effect degrades when the fine‑tuning objective conflicts with the targeted behaviour, yet the underlying weight edits remain largely unchanged. Mechanistically, the steering vectors survive with minimal alteration, but functionally the steering is vulnerable and must be re‑validated after downstream training.
arXiv:2608. 05810v1 Announce Type: new Abstract: Self-evolving agents accumulate capability by distilling reusable skills from their execution trajectories, but we find this process is not monotonic: past a critical pool size, newly added skills degrade performance instead of improving it.
arXiv:2607. 09709v1 Announce Type: new Abstract: Post-training a code generator against a learned judge can optimize proxy features that raise the score without improving the artifact.
Safety alignment in large language models can be fragile under fine-tuning, as even benign task adaptation may increase harmful compliance. Existing defenses mainly follow two directions: they either intervene during or after fine-tuning through retraining or weight modification, which can be costly and may hurt task performance, or they use model-agnostic safety classifiers, which may miss failures specific to a given fine-tuned checkpoint.