The study investigates how different selection rules affect the performance of adaptive halting in looped language models, specifically comparing a trained stability probe with Ouro's native exit gate on a frozen Ouro-1.4B model. In a pre‑registered confirmatory experiment, the probe saved more of the four‑loop budget than the native gate, achieving a gross macro saving of +0.053 across three benchmarks, but a secondary point‑only analysis reversed this advantage. Additional comparisons showed no clear benefit over an untrained KL‑divergence baseline, and the probe performed poorly on an out‑of‑distribution benchmark, highlighting the importance of jointly evaluating halting signals and selection procedures.
arXiv:2606. 28471v1 Announce Type: new Abstract: Model capability is the central variable in LLM pre-training, yet is never observed directly: data shapes it prospectively, while evaluation reveals it only retrospectively, compressing samples, prompts, decoding, and scoring rules into one noisy score.
The paper investigates how self‑improving large language models (LLMs) can suffer from the winner’s curse when selecting better candidates based on noisy evaluation sets. By modeling correlated errors and empirically testing Qwen models that rewrite their own instructions, the authors find that most proposals after the first are harmful and that the reported gains on the selection set often overstate true held‑out performance. The study shows that larger selection sets can improve held‑out gains on some tasks (e.g., TREC) but not others (e.g., GSM8K), and that common acceptance rules do not outperform simple greedy selection across entire runs.
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.
The paper evaluates how well typed decision models, which output probability distributions, actually provide reliable guarantees. By comparing open-weight Laya and Jev 1.13 across 190 evaluations, the study finds that point‑estimate thresholds frequently break promised accuracy, while a Beta lower‑bound rule and Learn‑then‑Test approach maintain guarantees more consistently. The work also explores calibration across languages, the impact of noise, data requirements for fine‑tuning, and the limitations of confidence in predicting outcomes or providing faithful explanations.
arXiv:2609.38699v1 Announce Type: new Abstract: A cumulative token cap can fall inside a mathematical derivation, forcing a test-time controller to choose between stopping at the cap (strict) and all...
The paper investigates whether large language models’ reasoning traces truly contain early, informative signals or merely reflect budget and difficulty confounds. Using a restart‑controlled truncation probe, the authors compare continuation success rates against from‑scratch restart curves across 178 problem‑model pairs, finding that only one case shows prefix‑limited success and that continuing a model’s own prefix generally outperforms restarting. A difficulty‑controlled test and two generation‑free analyses reveal that early internal signals do not carry outcome information beyond a problem‑difficulty baseline, underscoring the need for proper counterfactual controls.
arXiv:2606. 30852v1 Announce Type: new Abstract: Reasoning models spend different amounts of useful computation across instances, but it remains unclear when a learned stopping rule improves over simple confidence or convergence thresholds.
The study compares two ways of obtaining predictions from language models fine‑tuned on customer behavior: scoring answer tokens directly versus generating a written rationale and then scoring the resulting answer. Across 13 model‑domain cells covering four retail tasks, scored readouts consistently rank outcomes more accurately than generated readouts, with an AUC improvement ranging from 1.5 to 14.5 points. The authors also find that a third readout—eliciting a probability before any verdict—improves calibration but only when outcome rates are represented in training, and they recommend using generated rationales for interpretability while relying on scored heads for ranking.
The paper investigates how fine‑tuned language agents can selectively withdraw actions when evidence changes or authority is revoked. Using a synthetic NAQD‑Env environment, the authors compare an original curriculum with an enriched one across multiple system‑prompt framings and evaluate four LoRA checkpoints of Qwen2.5‑3B‑Instruct against instruction‑tuned baselines. Results show that fine‑tuned checkpoints achieve higher decision agreement and withdrawal recall, reduce incorrect withdrawals, and narrow prohibited combinations, though they also omit valid event reports and exhibit mixed preservation of authorized actions.
StrictPRMBench is a replayable evaluation benchmark designed to audit fixed‑score PRM aggregation rules. It standardizes mathematical reasoning pools, scorer evidence, and deletion interventions so that different reducers confront identical selection problems. Across seven checkpoints and two benchmarks, the benchmark reveals that product aggregation yields higher deletion‑induced selection instability than geometric mean, with statistically significant differences in most cells.
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 challenges the assumption that large language models (LLMs) produce deterministic safety responses by examining how random seeds and temperature settings affect refusal decisions. Across four instruction‑tuned models and 876 harmful prompts, 18‑28% of prompts flipped between refusal and compliance depending on sampling configuration, with higher temperatures reducing decision stability. The authors introduce a Safety Stability Index (SSI) and recommend multi‑sample evaluation protocols that account for stochastic variation rather than relying on single‑shot tests.