arXiv:2607. 24765v1 Announce Type: cross Abstract: Large language models (LLMs) can give different answers to the same decision problem across runs, and reverse a decision when their own prior answer returns as context.
The study investigates how language models (LLMs) handle refund or claim requests under written policies when presented with past cases that include both facts and recorded decisions. By separating these two components, the authors found that changing the recorded decisions of near‑copy past cases can increase approval rates by 30–55 percentage points, while altering the facts to support approval actually decreases approval rates, suggesting that the request appears weaker when compared to such facts. These effects were consistent across multiple models and both study‑specific and publicly paraphrased policies.
arXiv:2607. 16451v1 Announce Type: cross Abstract: Chat models sometimes commit to an answer and then produce reasoning that justifies it rather than deriving it -- even when the answer contradicts a task premise.
arXiv:2609. 11291v1 Announce Type: new Abstract: We post-train Qwen3.
The study evaluates whether reward‑trained language models can learn to keep evidence‑sufficiency boundaries without a formal correctness signal. Using a controlled knock‑in design, a boundary‑keeping recipe (SFT followed by GRPO) dramatically increased spontaneous correct refusals from 25 % to 100 %, while a scrambled‑pairing placebo collapsed performance. The results show that final boundary‑keeping is trainable without an inference‑time verifier, though active numeric checking remains negligible.
arXiv:2610.20532v1 Announce Type: cross Abstract: A multimodal agent must decide whether visual evidence is worth acquiring, which candidate or resolution to query, and when further interaction can n...
The paper evaluates how three large mixture‑of‑experts models (Alibaba, OpenAI, NVIDIA) can be fine‑tuned to reason in a low‑resource language, specifically Greek. Accuracy metrics show little change, but the authors uncover significant qualitative improvements: after supervised fine‑tuning, models reason in Greek on ~98% of items, with better grammaticality and retained general ability. Reinforcement learning with pre‑registered rewards further eliminates reasoning‑channel leaks and format skips, while the Greek‑reasoning habit remains robust to an accuracy‑only gradient.
arXiv:2609.25686v1 Announce Type: cross Abstract: Long-horizon assigned work requires an LLM agent to track the state of a task: which steps are done, blocked, cancelled, or open to repetition. Agent...
FinalityBench is an executable benchmark that tests how agents decide on shipping, re‑capturing, refunding, or waiting when a merchant’s payment processor, ledger, ERP, and bank feed receive delayed, duplicated, dropped, or reordered messages, causing contradictory beliefs about an order. The benchmark uses a hidden canonical event log and faulted delivery streams to generate system views, scoring each episode by the merchant’s terminal economic position relative to a privileged reference. It contains 321 tasks, including 45 twin pairs where all four views are identical yet the correct disposition differs, and evaluates nine programmatic policies, revealing that a ship‑on‑first‑sign policy performs best by accuracy but worst by paired loss, while a runtime‑gated irreversible‑action policy achieves 85.4% accuracy without losing money.
The paper introduces TRUST, a token‑level reward model that predicts the correctness of partial chain‑of‑thought (CoT) traces in vision‑language‑action (VLA) policies, enabling monitoring and selective steering of reasoning. On driving and manipulation VLA benchmarks, TRUST improves reasoning accuracy and reduces collision rates and trajectory errors, though its impact on overall task performance varies across tasks. The study defines two evaluation axes—correctability and actionability—to assess when CoT can serve as a runtime safety interface.
Augur is a synthetic decision laboratory that simulates how users will react to product and policy changes before they are released. It constructs a typed knowledge graph from change documents, populates a persona market, runs simulations, and produces an auditable decision memo recommending one of five actions. Using a dataset of 50 real episodes (Gold‑50), the authors evaluate the system’s five‑way release verdicts and find that evaluation design, rather than model capability, largely drives performance differences among frontier and open‑weight models.
The paper treats human interventions during LLM agent failures as labels that pinpoint the internal computation that went wrong, rather than merely training data. In a controlled environment, the authors analyze one failure type—over‑asking a clarifying question when the request is already clear—using rule‑based overseer labels validated by blind human judgment. They find that the decision to ask versus act can be causally localized across several models, that over‑asking exhibits distinct mechanistic signatures, and that steering or weight consolidation can mitigate the issue, though outcomes vary by model family.