LLM Semantic Gaps, Agent Harnesses, and Robust Multimodal Prediction

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering semantic abstraction and reasoning gaps in LLMs, reinforcement learning post-training and credit assignment, agent harness design, multimodal medical prediction, code vulnerability repair evaluation, Gaussian process basis selection, GenAI policy enforcement, and optimal annotation for off-policy evaluation.

1. Semantic Abstraction for Natural Language Inference: a Methodological Framework for Discovering and Compensating Semantic Knowledge and Reasoning Gaps in Large Language Models

Authors: David Torres-Moreno, Jorge Hermosillo-Valadez | Categories: cs.CL Link: arxiv.org/abs/2609.26610

The authors propose a methodological framework for building higher-level semantic knowledge in LLMs, defining semantic compatibility and incompatibility for natural language inference and reconfiguring lexical-semantic relations to induce new reasoning pathways. Their hybrid approach discovers and compensates for semantic knowledge gaps, improving accuracy by over 10% on some models—particularly for the non-entailment class—without adding data or model size.

Takeaway: A strong argument that structured semantic scaffolding, not scale, is what closes reasoning gaps in LLMs—with an unusually candid closing claim about mimicking human cognitive flexibility.

2. PACT: From Credit Assignment to Critic Alignment

Authors: Jiayan Fu, Hang Xu, Yong Zhang, Zhaokai Luo, Yao Hu et al. | Categories: cs.LG, cs.AI Link: arxiv.org/abs/2609.26355

The paper formulates three regularity conditions—Completeness, Prefix Consistency, and Neutrality—that uniquely determine token-level credit, unifying explanations of existing RL algorithms and motivating Policy Aligned Critic Training (PACT), which uses an Actor-then-Critic update order with importance sampling correction. PACT reaches 72.87% average accuracy across four agentic math reasoning benchmarks (outperforming GRPO and PPO by 8.80 and 13.16 points) and 67.4% on SWE-bench Verified.

Takeaway: A rare theoretical grounding of token-level credit assignment that translates directly into state-of-the-art post-training results—valuable for anyone working on RL for LLMs.

3. Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

Authors: Laizhen Li, Jiarui Li, Juanjuan Zhao, Kejiang Ye, Ye Li et al. | Categories: cs.AI, cs.SE Link: arxiv.org/abs/2609.26760

Growing Harness learns the agent harness itself from a strategy-free scaffold: function-level execution traces localize failures to bounded code, an optimizer jointly repairs them, and a held-out gate rolls back harmful edits—accumulating control logic in reusable code rather than model context. Across BrowseComp-Plus and WebArena-Verified it achieves the highest mean success in five of six settings, cutting LLM calls by 76.0–91.8% and inference cost by 74.4–98.6%.

Takeaway: The “grow code, not context” framing is a compelling alternative to ever-longer prompts, and the model-scale robustness results (45% vs. 6.7% at 4B) are striking.

4. MMAP: Multimodal Missing-Aware Pretraining for Longitudinal Alzheimer’s Prediction

Authors: Fiona Kekwick, Matthew Baugh, Bernhard Kainz, Paul M. Matthews, Wenjia Bai | Categories: cs.CV, cs.LG Link: arxiv.org/abs/2609.26617

MMAP is a multimodal missing-aware alignment pretraining method that learns image-tabular representations from incomplete clinical data, combining sigmoid contrastive learning with generative reconstruction for the image encoder and a tabular foundation model for the tabular encoder. A missing token generator makes both encoders robust to missing modalities, and the learned representations outperform strong multimodal and unimodal baselines on Alzheimer’s disease stage conversion and amyloid status prediction.

Takeaway: Missing data is the norm, not the exception, in clinical AI—this work tackles it head-on with a clean pretraining design and two genuinely challenging longitudinal tasks.

5. Metrics Failure in LLM-Based Code Vulnerability Repair: An Empirical Study and a Change-Aware Screen

Authors: Om Nepal, Sushant Aryal, Oluseyi Olukola, Nick Rahimi | Categories: cs.SE, cs.AI, cs.CR Link: arxiv.org/abs/2609.26749

Through five controlled experiments over 203 vulnerable functions from Big-Vul and three open-source code LLMs, the authors show that compile rate is a scientifically unreliable metric for vulnerability repair: it barely responds to real improvements, is dominated by harness and dataset artifacts (~64% of failures not attributable to the model), flips rankings under a single compiler flag, and rewards non-repairs when optimized. They propose diff_F1, a change-aware screen that gives zero credit to no-ops and near-zero to deletion-based gaming patches, while noting its limits as a repair-quality metric.

Takeaway: A necessary corrective for a subfield that has been optimizing the wrong number—required reading for anyone benchmarking LLM-generated security patches.

6. On Basis Function Selection for Sparse Gaussian Process Regression

Authors: Marnix Van Soom, Ivan De Boi | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2609.26624

Sparse Gaussian processes conventionally truncate a fixed basis to its first M entries, but the authors argue one should instead select the M basis functions that matter for the data at hand, proposing three information-theoretic ranking criteria matching different states of knowledge (no-data, no-prior, and in-between). Across six UCI regression benchmarks and three basis families (HSGP, VFF, VISH), the no-data criterion is a safe default matching or beating truncation, while the data-aware criteria give substantial gains specifically for the widely used HSGP.

Takeaway: A small, well-executed idea with immediate practical payoff—selection beats truncation, and the no-data criterion is a free upgrade for most practitioners.

7. From Alignment to Access Control: A Framework for GenAI Policy Enforcement

Authors: Nathalie Baracaldo | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2609.26682

This paper surveys “the good, the bad and the ugly” of policy enforcement in GenAI applications, proposing a methodology to systematically analyze and dissect existing approaches to defining and enforcing policy found in the wild. Based on this principled analysis, the author offers recommendations and a call to action, arguing that inconsistent meanings of “policy” across practitioners produce confused, siloed solutions inadequate for compliance.

Takeaway: A useful conceptual map of a genuinely muddled space—especially timely as agents move from chat to acting on users’ behalf, where alignment alone is not access control.

8. Optimal Sequential Annotations for Off-Policy Evaluation

Authors: Woojin Chae, Ezinne Nwankwo, Haitong Qin, Angela Zhou | Categories: stat.ME, cs.LG, stat.ML Link: arxiv.org/abs/2609.26707

The authors show how a limited budget for ground-truth annotation can be used via doubly-robust off-policy evaluation with missing rewards, characterizing variance-optimal annotation probabilities for sequential forward-monotone protocols and providing a feasible batch-adaptive implementation. Motivated by a homelessness services nonprofit’s casenotes, the method reduces RMSE by 34–65% for housing placement and 17–68% for housing application progress at budgets of 40% or more, and by 55–62% at every budget on LMArena.

Takeaway: A rare paper that pairs rigorous sequential experimental design with a concrete social-services deployment—directly relevant as LLM-as-a-judge labels proliferate with unknown bias.

This content was generated with AI assistance. Paper information sourced from arXiv.