Structured post-training pruning of transformers requires selecting complete functional units whose suppression causes limited degradation. We formulate structured-unit selection for language and vision transformers as a damage-aware multi-armed bandit problem under a fixed candidate-evaluation budget. Attention heads and MLP channel groups are temporarily masked on calibration batches. Paired damage is the masked loss minus the base loss on the same batch, reducing batch-to-batch variation. A smooth bounded reward drives either a UCB-style policy or fractional-Beta Thompson Sampling, and the final mask is constructed sequentially by adding one unit at each step. The selected units are functionally zeroed in the original dense checkpoint; therefore, the reported parameter effects represent effective structural suppression rather than physical compression or measured speedup. Experiments on WikiText-2, LAMBADA, and Imagenette cover GPT-2, OPT, Pythia, Qwen2.5, SmolLM2, ViT-B/16, DeiT-Tiny, and Swin-Tiny, with comparisons against random, magnitude, static-saliency, and budgeted-greedy selection. Across five seeds, the bandit methods usually reduce degradation relative to budgeted greedy in the paired language-model comparisons. Of 28 comparisons highlighted in the paper, 23 bootstrap confidence intervals exclude zero and 11 paired tests have p < 0.05; six have q < 0.05 after Benjamini-Hochberg correction across the full family of 116 dataset-wise tests. Matched-evaluation results for ViT-B/16 and Swin-Tiny indicate that their gains are not explained solely by a larger candidate-evaluation budget.
We introduce AutoFyn, an agent harness inspired by the Expert Iteration algorithm, adapting a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights. Each round begins from a fresh model session, and durable information is reintroduced only through explicit interfaces such as persistent memory files, reports, and repository state. Within a round, an orchestrator explores, plans and builds many alternative approaches with specialized agents, while a task-grounded verifier verifies the work and supplies an objective reward for measuring progress. This reward is distilled back into the persistent state, which updates the effective policy for the next round. In this technical report, we formalize this loop and describe its persistent state and verification interfaces. We then demonstrate its use in three domains, namely olympiad mathematics, data science, and cybersecurity. On the six fresh problems of the 2026 International Mathematical Olympiad, every model with room to improve scores higher under AutoFyn than in its provider's own coding agent. AutoFyn also built the top-ranked agent on the Spider 2.0 dbt benchmark, and has produced $16$ maintainer-confirmed vulnerability advisories in Next.js, MetaMask, pnpm, Warp, LiteLLM, Langflow, and Open WebUI.
Adib Hasan, Daniel Schaffield, Akashnil Dutta et al.· 0 citations
Long-term memory for LLM agents is evaluated today by conversational recall benchmarks (LoCoMo, LongMemEval), which measure question answering over dialogue history, not whether remembered facts change what a tool-using agent does. We present MERIT (Memory Evaluation for Realistic Instrumented Tasks), a benchmark and harness that measures the marginal utility of memory for task-executing agents under explicit cost accounting. MERIT provides episodic tool-use tasks in three domains whose dependence on earlier-episode facts is verified by an automated leak check; a difficulty ladder ending in updated-fact recall; controlled memory corruption; and full token and dollar metering of every memory operation. Across 23,440 scored episodes ($42.57), a two-generation pilot on gpt-4.1-mini and a preregistered 3-model x 3-seed grid (GPT-4.1, Claude Haiku 4.5; memory side held fixed), memory lifts dependent-task success from a leak-verified floor of 0.00 to 0.55-1.00. On updated facts, embedding retrieval collapses unpredictably (0.30-0.95 across models; max seed gap 0.45), and agents act on a correctly retrieved value only 55% of the time, while update-on-write stores (a structured fact store and, notably, LLM summarization) remain at 0.70-1.00; the hybrid is worse than the fact store alone. A latest-generation spot-check (Claude Sonnet 5, gated on a clean full-replay control) reproduces the pattern. Swapping a memory's implementation moves task success by up to 60 points, and full replay is never economical: the best condition per domain delivers 2.7-3.9x its marginal utility per dollar. We release the benchmark, harness, and all traces.
Current evaluation methods for large language models are coarse-grained and decoupled from generation, producing generic explanations that fail to provide actionable feedback for model improvement. We propose CriticGen, a fine-grained, generation-aware evaluation framework that turns evaluation into actionable control for answer improvement. CriticGen first generates sample-specific evaluation dimensions and scoring criteria under high-level categories such as subjective, objective, and self-derived constraints. These criteria then serve as a dynamic rubric for jointly producing a score, a reason, an executable refinement suggestion, and a refined answer. This rubric-conditioned refinement process enables models to diagnose flaws and perform targeted answer improvement. Experimental results show that fine-grained evaluation should be both instance-specific and actionable. CriticGen induces higher-quality rubrics, improving relevance/coverage from 3.33/4.03 to 3.97/4.24. CriticGen also achieves the best score correlations, with 0.9556 Pearson and 0.9560 Spearman, and raises the F1 of criterion-grounded reasons and executable suggestions from 0.6369/0.5994 to 0.7554/0.7900. Crucially, its feedback translates into reliable answer improvement, improving 73.17% of answers with a 93.28% non-degradation rate.
Huifang Du, Zecheng Zuo, Sen Wang et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Previous AI alignment efforts have focused primarily on first-order social norms -- teaching models what is socially acceptable or unacceptable (e.g., `do not steal'). However, social intelligence depends not only on norm recognition, but also on anticipating who will enforce it and how (e.g., public shame or even imprisonment). These second-order expectations, known as metanorms, govern how people respond when social rules are broken. We introduce a novel framework for evaluating metanorm reasoning in Large Language Models (LLMs) along two dimensions: emotional appraisal and behavioral response, and propose new classification tasks, namely, predicting self-regulation in violators, and other-regulation in observers. We release a multi-perspective dataset, NormReact, of 450 norm violation scenarios, hand-annotated for emotions and behavioral responses across norm violators' gender and observers' social closeness. Current LLMs portray a harsher social world: across six models, they overpredict negative sanctions where humans would expect inaction, and alignment with human judgments deteriorates as social distance increases. These findings suggest that AI systems in norm-sensitive domains from conflict mediation to policy simulation, may risk producing a distorted picture of social regulation: one that over-represents punishment and under-represents the tolerance, restraint, and relational calibration that characterize actual norm enforcement in real world.
Sunny Rai, Jinyi Kuang, Reyhan Jamalova et al.· 0 citations
Modern High Performance Computing (HPC) workloads exchange datasets at high velocity to enable data-driven science. HPC systems employ hardware and software I/O accelerators to enable efficient data exchange. However, domain scientists and software engineers must manually request specific hardware accelerators and manually integrate software accelerators into their workloads, significantly hindering their adoption. Recently, scientists have proposed I/O intents, defined as “why”, “what”, and “how” users or applications perform certain I/O operations, that can be used to automatically configure storage systems for the user. We designed an intent-driven scheduler plugin called IFlux that uses the workload’s I/O intents to assist the HPC system scheduler to automatically and dynamically allocate hardware and software accelerators. In this work, IFlux advances the current state-of-the-art in three key aspects. First, IFlux automatically and efficiently maps different I/O intents to various hardware and software accelerators with a throughput of 3.5 M intents per second with a scheduler overhead of 0.02%. Second, IFlux provides a scheduler workflow that uses a scheduler’s job specification enhanced with intents to allocate the required storage accelerators, deploy necessary middleware software, and enable workloads to use the allocated hardware and deployed software transparently with a runtime overhead of 0.01%. Finally, IFlux speeds up popular benchmarks such as the IOR and DLIO Benchmarks, which represent six classes of workloads from simulation, data analytics, and artificial intelligence, by up to 147× for specific use cases. In conclusion, IFlux enhances existing HPC schedulers with an intent-driven approach to speed up large-scale HPC workloads such as the 1000 Genomes workflow and Megatron Deepspeed by up to 3.12× on an institutional cluster.
Hariharan Devarajan, Vanessa V. Sochat, Daniel Milroy et al.· IEEE Transactions on Paralle...· 0 citations
