The paper investigates how fine‑tuning large language models with a small number of harmful examples can erode their refusal behavior, and explores whether localizing safety‑related behavior to specific layers or directions can provide robust defenses. Experiments across six checkpoints from four model families show that harmful and benign prompts remain linearly separable after attack, and that patching clean hidden states or freezing layers up to a transition depth can restore refusal. However, attackers can bypass these defenses by spreading updates or targeting singular directions, indicating that adaptive fine‑tuning can defeat localized repairs and highlighting the need for multiple defensive checks.
By Jungseob Lee, Dongyub Jude Lee, Sugyeong Eo, Seongtae Hong, Seungyoon Lee, Heuiseok Lim
LEV is a decision model that performs reasoning by looping its layers instead of generating text, allowing it to deepen its computation on a per-question basis. Trained jointly across all loop steps, a single LEV checkpoint can answer at any depth, improving accuracy—especially on hard questions—while keeping inference fast. On benchmark tests, LEV outperforms dense and cascade models by 5.4 to 8.8 accuracy points with sub‑50 ms latency, and on‑demand depth selection saves 15–24% of computation for minimal accuracy loss.
By Dingjie Song, Zhiling Yan, Yuxuan Zhang, Lichao Sun
arXiv:2610.19768v1 Announce Type: new
Abstract: Looped transformers compute by applying one block repeatedly, so they can run for more loops than they were trained with. Recurrent reasoners are known...
By Peilong Zhou, Hongtao Zhang, Yulong Huang, Cangyuan Li, Mingyu Chen, Ying Wang
arXiv:2607. 10203v2 Announce Type: replace-cross Abstract: Adaptive-compute world models -- early-exit or mixture-of-depths predictors that spend variable depth per step -- assume depth buys better predictions and can be routed adaptively.
By Achyuthan Sivasankar
arXiv:2607. 10203v1 Announce Type: cross Abstract: Adaptive-compute world models -- early-exit or mixture-of-depths predictors that spend variable depth per step -- assume depth buys better predictions and can be routed adaptively.
By Achyuthan Sivasankar
The paper investigates whether deeper predictor blocks in adaptive‑compute world models truly improve rollout accuracy and whether a router can select the appropriate depth per step. Experiments on nine DeepMind Control tasks show that deeper predictors do not always yield better accuracy, and that a greedy oracle selecting depth per step can reduce error by up to 37% compared to fixed‑depth mixtures, yet no trained router achieves this advantage. The study also finds that supervising early exits only at the first rollout step can mitigate depth‑inversion issues observed in some tasks.
By Achyuthan Sivasankar
The study investigates why small language model agents tend to repeat a tool call that just failed. By recording the failed call and its error message in the transcript, the authors measure a negative corrective gain—agents are more likely to repeat the failed action, with a drop of about 1.03 nats per token. The problem is traced to the harness design rather than the model’s understanding of errors, and the authors show that replacing the verbatim call with a runtime-generated description of the failure can reduce this backfiring effect by 76%.
By Esmail Gumaan
arXiv:2609. 31181v1 Announce Type: new Abstract: Black-box model identification works by scoring a model's response to natural-language prompts.
By Nicol\'as Vera Z\'u\~niga
arXiv:2607. 14427v1 Announce Type: new Abstract: A depth-recurrent transformer applies a weight-tied core a variable number of times, and prior work has shown that training with a randomized recursion count yields one checkpoint usable across a range of inference depths.
By Joe Logan
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...
By Srikanth Malla, Chiho Choi, Joon Hee Choi
arXiv:2607.27836v2 Announce Type: replace
Abstract: Large language model unlearning is consistently fragile under relearn attacks. On TOFU, fine-tuning on twenty forget examples substantially recover...
By Xiangyu Yin, Jiaxu Liu, Zhen Chen, Chih-Hong Cheng
arXiv:2609.36426v1 Announce Type: cross
Abstract: A detector pretrained on a broad corpus is fine-tuned on a narrow domain, its in-domain accuracy improves, and it ships. We ask what happens meanwhil...
By Trung Minh Bui, Jongsul Moon, YoungOuk Kim, Jung-Hoon Hwang, Dongin Shin