A 0.6B language model consistently answers YES to 1,200 logical tests, yet its behavior shows no discrimination. Linear probes reveal the correct verdict with high AUC (0.96) and transfer to unseen structures, but a single scalar readout fails due to a saturated decision threshold offset by +4.6 σ. Adjusting this threshold restores behavior accuracy from 50 % to 81 % and improves higher‑scale models, demonstrating that miscalibrated readouts, not hidden knowledge loss, drive performance gaps.
By Gnaneswar Villuri, Hashmath Shaik, Alex Doboli
The study examines how two small instruction‑tuned language models, Qwen2.5‑1.5B and Llama‑3.2‑1B, respond to user pushback on TriviaQA. When initially correct, the models flip to a wrong answer in about 42–43% of cases, with the effectiveness of different pushback styles varying by model. Attempts to decode capitulation from the pre‑response residual stream fail under a rigorous validation protocol, revealing overfitting and a measurement hazard that underestimates capitulation by 18–24 percentage points.
By Saad Aamir, Muhammad Awais Bin Adil
arXiv:2607. 08268v1 Announce Type: new Abstract: High-volume structured extraction pays a large model's latency on every item, so distilling the task into a small on-device model is attractive: comparable output at a fraction of the time and cost.
By Vinay Kumar Chaganti
The paper investigates correctness‑gated multi‑teacher distillation, comparing a weighted arm to unfiltered distillation across eight experimental arms. While the weighted arm shows modest gains in accuracy (+0.1660) and macro‑F1 (+0.1323) and a reduction in unsafe action rate (−0.4979), it also exhibits lost label functionality, such as zero Refuted recall and over‑assignment of NotEnoughInfo. A subsequent grounding audit was inconclusive, failing to demonstrate a clear improvement in evidence grounding or overall system performance.
By Xiaofei Feng
arXiv:2609.37616v1 Announce Type: new
Abstract: Language models tend to agree with whatever a user asserts, and post-training increasingly targets this sycophancy so that models evaluate claims on th...
By Abhinav Rajeev Kumar (Lossfunk), Paras Chopra (Lossfunk)
arXiv:2607. 14111v1 Announce Type: cross Abstract: Can small language models detect and report on perturbations their own internal activations?
By Ely Hahami, Ishaan Sinha, Lavik Jain
arXiv:2607. 19956v1 Announce Type: cross Abstract: Knowledge distillation (KD) is a standard approach for compressing sequence-to-sequence models, but its per-sample effects are rarely examined.
By Dipto Sumit, Ankan Kumar Roy Srizon, Sadia Khair Rodela, Atia Haque Asha, Mourchona Afrin, Niloy Farhan, Farig Sadeque
The study evaluates large language model (LLM) graders on two computer‑science exams, testing 171 configurations of closed‑ and open‑weights models. While the best LLM configuration achieved a mean absolute error of 1.64/35—better than the 2.61/35 error between two human graders—its performance was highly sensitive to the prompt. A short "strict grader" preamble caused most open‑weight models to exceed acceptable error thresholds or stop grading entirely, whereas fine‑tuning with a single LoRA adapter restored parity with human graders and reduced sensitivity to harsh prompts.
By Ali Habibullah, Yazan Alshoibi, Mohammad Alshiekh, Salman Khan, Naeemullah Khan
arXiv:2608. 14644v1 Announce Type: new Abstract: Real-world LLM deployments increasingly rely on runtime-injected prohibitions--enterprise policies, PII redlines, tool boundaries--that vary per request and per tenant.
By Zihan Li, Feifei Li, Wenhui Que
The paper examines how large language models (LLMs) can be biased by irrelevant social contexts when evaluating teachers, using a large U.S. classroom transcript dataset. It shows that spurious contexts can shift model ratings by up to 1.48 points on a 7‑point scale and that standard mitigation methods like SFT and DPO are insufficient. The authors introduce Debiasing‑DPO, a method that combines contrastive reasoning‑augmented DPO with SFT, which reduces bias by 84% and improves predictive accuracy by 52% on Llama and Qwen Instruct models.
By Hyunji Nam, Dorottya Demszky
arXiv:2607. 19257v1 Announce Type: cross Abstract: Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade.
By Netanel Eliav
The study evaluates zero‑shot essay scoring using sub‑3B open language models that run locally on a single 8 GB consumer GPU. Four instruction‑tuned models (Qwen2.5‑0.5B, 1.5B, 3B and SmolLM2‑1.7B) were tested on all eight ASAP‑AES prompts, comparing rubric‑decomposed versus holistic prompting, different aggregation methods, and trait‑mapping strategies. Results show rubric‑decomposed prompting consistently outperforms holistic prompting, trait‑mapping is sensitive to calibration, and longer essays reduce error, yet the best local configuration (macro QWK 0.388) still falls short of human agreement and a length‑only baseline, suggesting these models are best suited for formative, human‑supervised feedback.
By Nguyen Dung Son, Dang Quang Minh, Nguyen Huu Loi, Truong Viet Vu, Nguyen Thai Anh