RubricsTree introduces a hierarchical taxonomy of over 100 clinically-verifiable Boolean rubrics, evolved from 4,000 real user queries via human-in-the-loop curation. It enables scalable, expert-aligned evaluation of personal health agents by dynamically routing queries to relevant rubrics and outperforms baseline methods in alignment, context sensitivity, and model performance gains of up to 66% on HealthBench.
Zone of Proximal Policy Optimization (ZPPO) integrates teacher knowledge directly into prompts rather than policy gradients. It uses Binary and Negative Candidate-included Questions to surface student failure modes and amplifies learning through a prompt replay buffer, achieving superior performance on hard questions across student scales, especially at smaller model sizes.
SkillMigrator learns reusable web skills by matching layout structures instead of element references. It stores each skill as a transferable interaction pattern with a structural sketch, enabling efficient skill transfer across sites. Compared to state-of-the-art methods, it reduces average LLM-action counts by 8-10% on WebArena and Mind2Web at matched success rates.
A new method enables large language models to learn from their own reasoning traces without external supervision. By distilling inference-time computation into lightweight, modular latent memories, the model achieves performance competitive with full training and outperforms zero-shot and raw ICL baselines on mathematical reasoning tasks, with minimal computational overhead.
Handlebars' triple-brace interpolation fails to protect against structural role injection, as HTML escaping only neutralizes angle-bracket delimiters. It leaves colon and Markdown hash delimiters intact, enabling attackers to hijack model turns. The default escaping provides no protection for most delimiter families and cannot replace a structural separation of instruction and data.
A new benchmark, COGNITIVE ATROSPHY BENCH, measures how LLMs induce cognitive decline in mental-health conversations. Built from 1,576 human-generated counseling sessions and evaluated by clinical experts, it identifies patterns like directive advice and validation that may reduce user autonomy. The tool introduces metrics such as UIRI and ARI to assess atrophy risk and track behavioral trajectories across user interactions.
TAC evaluates whether AI agents avoid animal exploitation in travel bookings. Seven frontier models all score below 64% chance level, with Claude Opus 4.7 at 53%. Adding a welfare-aware system prompt improves performance significantly, though models show no evidence of evaluation awareness in their responses.
A red-team study of Anthropic's Fable 5 and Opus 4.8 models reveals both are vulnerable to adaptive iterative attacks, with Opus 4.8 breached on 11.5% of harmful intents and Fable -5 on 6.1%. Despite robust defenses, both models generated 1,620 and 702 panel-confirmed harmful completions across all harm categories, automatically and efficiently under automated attack.
RubricsTree introduces a hierarchical taxonomy of over 100 clinically-verifiable Boolean rubrics, evolved from 4,000 real user queries via human-in-the-loop curation. It enables scalable, expert-aligned evaluation of personal health agents by dynamically routing queries to relevant rubrics and outperforms baseline methods in alignment, context degradation detection, and model performance gains of up to 66% on HealthBench.
VERITAS introduces a generator-verifier framework that enables robots to improve policies in real time without additional training. A visual verifier evaluates actions at inference time, allowing consistent performance gains through verified rollouts that serve as effective supervision for offline policy improvement. Post-training with these verified rollouts matches expert demonstrations in efficiency, without human intervention.
A study challenges the assumption that visual attention signals reliability in vision-language models. It finds near-zero correlation between spatial attention and accuracy, showing instead that self-consistency across reasoning paths is a stronger predictor of truth. Reliability is better explained by generation dynamics and internal state distributions, not visual attention patterns.
NarrativeWorldBench evaluates 21 LLMs on nine narrative-structure metrics across horizons of 10 to 200 episodes, with cross-lingual support in Hindi, Tamil, Telugu, and Marathi. N-VSSM, a latent world model using Mamba-2, achieves plot-beat F1 of at least 0.84 across all horizons with 4x lower compute than closed-frontier models and outperforms Claude Opus 4.5 in long-arc consistency and controllability in a professional writer study.
PARSE reduces prompt injection attack success from 25.4% to 15.6% on real enterprise documents across five professional domains, with statistically significant improvement (p=0.014) and 86.9% utility. It outperforms paraphrasing and uses provenance-aware sanitization to preserve factual content while routing most documents through a lightweight path.
STATEWITNESS introduces an activation explainer that audits deception in reasoning LLMs by reading hidden states and generating natural-language answers or structured reports. It achieves a 0.916 mean AUROC, outperforming existing black-box monitors and activation probes by 11.6% and 25.0% respectively, and provides query-level, schema, and evidence-level traces for human inspection.
As enterprise agent tool catalogs scale from 10 to 110 agents, routing accuracy drops 16--23 percentage points on under-specified requests. An oracle analysis identifies retrieval and confusion gaps, with embedding-based shortlisting recovering +10--11pp F1. A human-annotated study of 1,435 utterances confirms real-world recovery of +10--17pp despite lower absolute performance.
Concatenating LLM-generated features to graph neural networks systematically reduces accuracy on homophilous benchmarks, with PubMed accuracy dropping by -17.0 ± 0.3 pp. This degradation is linked to LLM-alone discriminability (Delta_sig), which correlates strongly with concatenation cost (r² = 0.38) and shows a power law relationship with feature dimension and node count (r² = 0.97), particularly in low-Delta_sig, low-node scenarios.
OPD-Evolver introduces a slow-fast co-evolution framework that enables agents to select, act on, and reuse experience through on-policy self-distillation. It outperforms existing memory and training-based methods by up to 11.5% and 5.8% respectively, and demonstrates capability to challenge large-scale models like Qwen3.5-397B-A17B and Step-3.5-Flash.
SkillMigrator learns reusable web skills by matching layout structures instead of specific element references. It stores each skill as a transferable interaction pattern (TIP) with a structural sketch, enabling efficient skill reuse across sites. Compared to state-of-the-art methods, it reduces average LLM-action counts by 8-10% on WebArena and Mind2Web at matched success rates.
MambaCount introduces a spatial sparse state space duality block to enable efficient text-guided open-vocabulary object counting. It addresses causal modeling limitations and high entropy in spatial token responses, achieving state-of-the-art results on FSC-147 with a test MAE of 12.23 while maintaining linear complexity.
SuCo introduces Minimal Sufficient CoT (MSC) as the shortest reasoning prefix adequate for correct answers. It employs a two-stage training framework—MSC-Aligned Fine-Tuning and Sufficiency-Aware Policy Optimization—to reduce reasoning length while maintaining or improving accuracy across math, code, and science tasks.