The paper evaluates additive activation steering in chat and agent contexts, showing that the commonly used gain ratio (Δ_agent/Δ_chat) fails to reliably indicate potency and efficacy across multiple models and dose-response cells. By replacing the gain with a location metric, dEC50 (difference in EC50 between agent and chat), the authors demonstrate a more robust, two‑sided measure that consistently captures cross‑context shifts. The study also reports several refuted and unanswered claims, emphasizing that a single operating point cannot distinguish between displacement and gain effects.
arXiv:2607. 09156v1 Announce Type: new Abstract: Additive activation steering (injecting a scaled residual-stream direction during generation) is calibrated almost entirely in single-turn chat, yet the models it targets are increasingly deployed as tool-using ReAct agents.
The paper introduces AgentDiff, a metric that quantifies how much LLM agents’ answers differ when inputs are altered by meaning‑bearing rewrites (paraphrases, synonym substitutions) versus presentation changes (reordering, formatting, distractors). Across 68 model–benchmark–scaffold combinations involving ten LLMs and over 1,500 questions, meaning‑bearing rewrites consistently produce a roughly 20‑percentage‑point higher inconsistency rate than presentation changes, a gap that persists across severity proxies and remains significant even outside the Qwen family. Trace analysis reveals that meaning‑bearing rewrites preserve the first action but reduce thought similarity from the second step onward, extending the divergence cascade—a phenomenon termed “stealth divergence.”
arXiv:2606. 01090v1 Announce Type: cross Abstract: Equivariance theory predicts that an architectural symmetry prior reduces sample complexity by a factor of |G|; this is widely cited but rarely measured as a scaling law with controls that separate the prior from its confounds.
arXiv:2607. 04510v1 Announce Type: cross Abstract: Emergent misalignment (EM) -- the broad misbehaviour a language model acquires after fine-tuning on narrow harmful data -- is mediated in Qwen2.
The paper investigates selective on‑policy distillation, where a student model is trained only on token positions chosen by a selector. It demonstrates that the commonly used shared learning rate is not neutral: performance varies significantly with the learning rate for different selectors, leading to inconsistent comparisons. The authors attribute this selector‑rate entanglement to the selection process itself and recommend reporting the full arm‑by‑rate matrix for fair evaluation.
arXiv:2607. 18553v1 Announce Type: cross Abstract: Can a language model read the quality of ongoing computation, and can an external intervention turn that readout into better outcomes?
The paper reports that counterfactual fairness audits of clinical language‑model agents are unreliable without accounting for a per‑action instability floor. By repeatedly running identical vignettes, the authors found that actions changed 8.7% of the time, with instability varying eightfold across actions. A second model confirmed a pooled floor of 6.7%, showing that any reported fairness estimate lacking this floor cannot be interpreted as evidence of disparity.
arXiv:2609.08618v1 Announce Type: new Abstract: Benchmark scores describe what a checkpoint can do now, but they do not determine how it will respond to the next training episode. We measure this mis...
arXiv:2607. 04419v3 Announce Type: replace Abstract: When evaluator-derived step rewards are pooled or compared across scoring channels, their sign is treated as transportable.
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.
arXiv:2607. 17427v1 Announce Type: new Abstract: Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models.
The paper investigates how language‑model judges can make version‑dependent errors when evaluating upgraded agents. Using 35 public coding‑agent submissions, two customer‑service agents, and over a thousand expert‑labeled trajectories, the authors show that fixed judges often reject task‑conditioned error invariance and can incorrectly approve failed patches, especially as agent capability increases. Paired audits of current outputs reduce interval width only marginally, and the study concludes that independent human patch review is still necessary.