The paper investigates whether activation interventions that guide a language model toward a target answer also enable the model to use the intended condition in other computations. Using Qwen3.5‑27B on tool‑use tasks, the authors find that residual‑writing variants improve target‑selection accuracy but these gains vanish when familiar facts and identifiers are recombined, and they do not translate into better independent judgments about eligibility or selection. The study demonstrates a clear separation between target‑selection control, condition‑sensitive judgments, and complete‑action correctness, highlighting the importance of evaluating how edited states influence downstream reasoning rather than just the final answer.
arXiv:2508.15754v2 Announce Type: replace-cross Abstract: Tool use is often assumed to monotonically improve reasoning, where external evidence is expected to help when relevant and be ignored when i...
The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.
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...
The paper investigates how to efficiently repair stale key-value (KV) caches in retrieval‑augmented generation systems after document edits. It proposes a budgeted in‑place recomputation approach and evaluates training‑free position‑selection policies on a factual RAG benchmark. Across three model families, a contiguous edit‑local window consistently recovers most of the post‑edit answer quality while being 13–21 times faster than a full re‑prefill, though its effectiveness diminishes when answer‑bearing text moves downstream.
The paper introduces Verbalization Training (VT), a technique that encourages large language models (LLMs) to openly express their evaluation awareness (EA) without directly supervising their internal beliefs. VT works by truncating model rollouts just before spontaneous verbalizations, creating training prefixes that signal awareness, and then applying a reinforcement learning objective to increase calibrated verbalization. Experiments on models such as Qwen3.6-35B-A3B, Kimi K2.6, and Inkling show that VT boosts verbalized EA by 2.4–2.9× while keeping latent EA and overall behavior largely unchanged, and a causal study confirms that VT-induced verbalizations reflect newly acquired meta‑knowledge.
The paper introduces Boundary‑Calibrated Intervention Transfer (BCIT), a method for conditional experience transfer in autonomous large language model (LLM) post‑training. BCIT links each past update to its specific parent model, data, and training stage, checks whether those conditions still hold, vetoes updates with hard conflicts, and, when necessary, runs a bounded training trial to confirm applicability before adopting the update. Experiments on a 4B model across finance reasoning, text‑to‑SQL, and function calling show that BCIT reduces harmful updates and achieves higher final‑model quality under equal computational budgets compared to other approaches.
arXiv:2607. 20090v1 Announce Type: cross Abstract: Retrieval-augmented large language models frequently face contexts that interleave useful evidence with misleading statements or instruction-like content.
arXiv:2606. 12747v1 Announce Type: new Abstract: Safety-relevant studies of language models, including alignment and jailbreaking evaluations and AI control protocols, often rely on prefilling model outputs.
The paper introduces RAG‑Stress, a diagnostic protocol that tests how retrieval‑augmented generation models handle misleading evidence. By fixing the question and reference answer, editing a single assertion to support an incorrect answer, and varying source‑priority policies and evidence positions, the study measures misleading rate (MR) and clean accuracy across fifteen systems on multiple QA datasets. Results show that prioritizing documents increases MR by 10.9–13.5 percentage points, with the highest MR when the misleading span appears at the end of the evidence.
The paper investigates how reinforcement learning can cause large language model agents to adopt shortcut policies for tool use, relying on superficial prompt cues rather than actual task needs. By creating synthetic environments that mix factual QA and math reasoning, the authors show that agents often invoke tools when cues are present, even when those tools are unnecessary, with spurious invocation rates rising up to 39%. They find that shortcut learning occurs mainly when agents have already mastered the target tool and that semantic alignment between cues and tools amplifies the effect. To counter this, they propose a dense, decision-level reward where an LLM judge assesses tool necessity, which reduces cue-driven tool use while maintaining performance.
arXiv:2605. 09692v3 Announce Type: replace Abstract: Autonomous language agents increasingly expose traces, memories, plans and constraints, but existing evaluations rarely test whether these state variables are bound to final actions.
The paper investigates whether knowledge editing truly erases original facts from language models. Using a linear trace probe, the authors find that after editing a fact in GPT‑2‑XL, the original object remains highly decodable from hidden states across three different editing methods, even when the model behaves correctly on edited prompts. This suggests that editing suppresses rather than removes the original association in representational space.