The paper reports three preregistered studies evaluating a longevity‑tuned 1.7B language model’s performance when a harness is applied. Across multiple tasks, the harness sometimes improves test accuracy over a capped thinking setting, but its gains are often comparable to or only slightly better than a constant label baseline. Replication experiments confirm that the observed improvements can be attributed to specific unparsed rows and pipeline artifacts rather than a general model enhancement.
arXiv:2610.05533v2 Announce Type: replace-cross Abstract: Harness search keeps a change to the prompts, token budgets or parsers around a frozen model if the change raises a score. Such a gain can co...
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.
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.
arXiv:2608. 15286v1 Announce Type: cross Abstract: We introduce AgentRelBench, an environment-agnostic reliability instrument that computes ground-truth, severity-priced damage from database state diffs across repeated runs, with no LLM in the measurement path, demonstrated on EnterpriseOps-Gym.
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 introduces TRACE, a protocol designed to diagnose whether changes in verifier scores for large language model agents reflect actual changes in agent behavior or merely alterations in the evaluation process. TRACE works by applying targeted changes to evaluation components, running paired experiments, and re‑scoring unchanged trajectories to isolate the source of score variation. Experiments on synthetic tasks and public benchmark tasks demonstrate that seemingly significant score shifts can often be attributed to evaluation artifacts, while TRACE can also detect genuine behavioral changes.
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.
arXiv:2610.15710v1 Announce Type: new Abstract: Long-term memory for assistants needs a gate that decides when a new utterance retires an old stored claim. Benchmark labels for this step mark a pair...
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.
arXiv:2608. 20290v1 Announce Type: new Abstract: Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses.
The study investigates how inference‑time interventions and weight consolidation affect open‑ended generation in an online bin‑packing task. By iteratively generating, verifying, selecting, and consolidating with LoRA, the model’s outputs shift toward higher value, reducing excess by 1.7 points and outperforming random consolidation by 3.1 points. Across three independent runs, the mean performance remained consistent, and the best candidates converged to the classic heuristic’s level without exceeding it, while consolidation also lowered the proportion of better‑than‑classic candidates but increased their absolute number.
arXiv:2608. 03842v1 Announce Type: cross Abstract: When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate.