T1 is a 122B-parameter Mixture-of-Experts model trained with reinforcement learning to operate a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task's own verifier. The paper reports 64.0% resolved on Terminal-Bench 2.1 and 27.9% on Long-Horizon Terminal Bench, surpassing GPT-5.4 and GLM-5.1, with TITO and R3 reducing the training-to-inference log-probability difference from 0.021 to 0.013.
The constraint that moved from prior terminal-agent work is long-horizon training stability: the paper reports up to 300+ tool-call turns, a dense verifier-based process reward, TITO construction with drift repair at turn boundaries, and rollout routing replay to align sampled token identifiers and MoE expert choices between rollout and training. The reported log-probability difference drop from 0.021 to 0.013 is the specific stability constraint claimed to be relaxed.
Evidence is limited to a single arXiv abstract. Key claims—benchmark scores, log-probability reduction, zero token drift, and out-of-distribution transfer—are reported without full baseline scores, variance, ablations, or released artifacts. The comparative claim against GPT-5.4 and GLM-5.1 cannot be independently verified from the abstract because the baselines' numbers are omitted. The 122B total parameter count omits active parameter count, so inference cost is unknown.
The primary document announces RESCUE-BENCH, a benchmark and task formulation for relation-aware multi-party emotional support conversation. RESCUE is constructed from real couple and family interview conversations and contains 191 samples, 7,079 annotated turns, and 1,064.8 minutes of video. It defines six tasks across Relational Understanding and Relation-Sensitive Support. Experiments with ten LLMs reportedly show relatively good performance on tasks relying on local emotional or intervention cues, but difficulty with relation-intensive tasks such as relation pattern prediction, viewpoint prediction, and support strategy prediction.
The constraint that moved is the scope of emotional support conversation modeling and evaluation: from one-on-one seeker-supporter interactions and individual emotional states to multi-party settings in which evolving interpersonal relations are the object of prediction and support decisions. This shifts the evaluation target toward relation pattern, viewpoint, and support strategy prediction.
The available text is an arXiv abstract. It reports the benchmark's stated corpus size, task count, capability framing, and a qualitative summary of LLM results. It does not provide model names, metrics, splits, annotation guidelines, inter-annotator agreement, statistical tests, or human evaluation of support outcomes. Claims about model limitations are therefore directionally informative but not independently verifiable from the supplied evidence.
PlannerForge is presented as an LLM-agent framework that extends all scenario-based testing stages from scenario generation to ADS assessment and adds LLM-enhanced ADS enhancement and benchmarking. The abstract reports evaluation with 10 off-the-shelf LLMs across six tasks under five prompt conditions, best-per-task scores of 0.88–1.00, open-source 20–35B models matching commercial APIs on most tasks, end-to-end retention of 83% commercial / 78% open, and improvements over named baselines on generation, retrieval, and edit validity. At N=400, cost-tuning is reported to raise planner success from 50.4% to 70.2% and reduce collisions from 19.0% to 8.4% without domain-specific fine-tuning. The available document is an abstract, so methodological details, metric definitions, and independent validation are not present.
The proposed constraint shift is from a fragmented, modular scenario-based testing pipeline with separate tools to a unified LLM-agent framework that spans the full pipeline and adds enhancement and benchmarking. If the reported results hold, a second constraint also moves: open-source 20–35B backends can approach commercial APIs across most evaluated tasks, and cost-tuning can improve planner success and reduce collisions without domain-specific fine-tuning, potentially lowering the cost and integration barrier for scenario-based ADS validation.
Evidence is limited to a single arXiv abstract. It reports self-contained quantitative claims and named comparisons, but does not provide the scoring rubric, statistical uncertainty, experimental protocol, validator definitions, or baseline configuration. No code, data, or peer-review status is available in the supplied document. The arXivLabs boilerplate is unrelated. The claims are therefore best treated as provisional, and no independent verification is possible from the provided text.
In a single empirical study, two code models—a 100M-parameter model trained from scratch (Rainbow-Pony-100M) and a fine-tuned Qwen2.5-Coder-0.5B—were trained in both direct whole-file generation and iterative diff-based edit regimes on a shared Flutter/Dart dataset. On a held-out set of approximately 1,790 tasks per model, direct generation outperformed diff-based generation on all measured aggregate metrics, with the gap persisting after matched-ID difficulty control and on code compiling on both sides. Diff-based generation was competitive on short, spatially localized edits, with category-level wins in refactoring and error-handling/edge-case fixes, the categories with the lowest mean edit-step count; the authors term this task locality.
The operative constraint shifts from 'diff-based editing should be preferred because it mirrors developer edits and uses fewer tokens per turn' to 'diff-based editing is viable only when tasks have high spatial locality (low edit-step count), such as refactoring and error-handling fixes; otherwise direct whole-file generation is the stronger regime'.
The evidence is a single arXiv abstract. It reports four trained models, a held-out set of approximately 1,790 tasks per model, multiple metrics, a matched-ID comparison, and a compile-both restriction. The abstract does not report effect sizes, confidence intervals, exact task counts, step-budget values, or judge agreement. The mechanism claim rests on category-level aggregation of mean edit-step count, not on a controlled manipulation of locality. No external replication or full paper detail is available in the source.
The paper proposes auditing bias in hidden-state representations across related model variants by encoding sentences through similarities to anchor sentences, yielding relative representations. It reports that Representational Bias Shift ΔB correlates with output-level bias change in 15 of 18 settings, detects bias-increasing checkpoints with ROC AUC 0.65–0.99, and runs in about three minutes using 3–50× less compute than the output-level benchmarks considered.
The operative constraint moves from output-level evaluation data and judge models to a representation-level audit that compares related checkpoints through anchor-relative similarities. If the reported properties hold, bias auditing before and after fine-tuning can be performed without task-specific evaluation data, in about three minutes, and with 3–50× less compute than the output-level benchmarks considered. The remaining constraints are the anchor and attribute sets, which are not eliminated but relocated into the audit design.
Only the arXiv abstract is available as a substantive block; the remaining blocks are arXiv page furniture. All quantitative results—15/18 correlation, |r| = 0.84, ROC AUC 0.65–0.99, 3–50× compute reduction, and stability under anchor/attribute/template changes—are author-reported in that abstract. The abstract does not provide per-setting breakdowns, confidence intervals, anchor-set size, benchmark-specific compute denominators, or the three settings where correlation failed. The paper positions ΔB as complementary to output-based auditing, which is consistent with the evidence presented, but the evidence cannot independently verify the claims.