AI/ML Reading List

Ars Technica reports that red-teaming by safety researchers found OpenAI and Google's newer image generation models may be more vulnerable to jailbreaks producing non-consensual explicit content, despite claimed safety improvements. This represents a critical gap between capability scaling and safety constraints—directly relevant to practitioners and policymakers evaluating deployment risks of frontier models.

US Federal Reserve and European banking regulators are conducting coordinated inquiries into AI agent deployment risks and cybersecurity vulnerabilities within financial institutions. This represents an early regulatory response to emerging AI system risks in critical infrastructure and signals field-wide policy concern around autonomous AI in high-stakes domains.

AlphaPilot combines reinforcement learning with multi-turn LLM agents to mine interpretable trading factors by conditioning on market text, historical data, and real evaluation feedback in an adaptive loop. The work exemplifies advanced agent reasoning and tool use in a high-stakes domain with measurable performance benchmarks, advancing the frontier of LLM agent architectures beyond single-turn inference.

Researchers propose oracle-guided fine-tuning using curated N-best hypothesis pools to train LLMs for ASR post-editing with a select-or-rewrite policy, reducing hallucination and over-editing compared to prompt-only approaches. Relevant to practitioners working on speech recognition and LLM fine-tuning, though gains are dataset-dependent and the contribution is specialized rather than foundational.

Researchers propose Vela, a framework that converts anonymized user behavior histories into supervised training tasks for LLMs, generating 100K tasks from 100K users across 91 real-world templates. Fine-tuning Qwen 35B on these trajectories yields up to 11.5-point gains on user-understanding tasks and 4-point average gains on general benchmarks, suggesting real-world behavioral data is a scalable supervision signal for improving both user-centered and general LLM intelligence.

Researchers introduce SciCurateBench, a 200-paper expert-annotated benchmark for evaluating AI systems on complex scientific data curation tasks, and propose LangPoli, a policy-actor agent framework that achieves 83% F1 on paper classification versus 58% for base LLMs. This advances practical scientific AI tooling and demonstrates that structured agent decomposition significantly improves reasoning on complex, multi-criteria document verification—relevant to practitioners building research automation and knowledge extraction systems.

Researchers dissect why LLMs fail at dynamic clinical diagnosis by analyzing evidence acquisition, utilization, and belief updating across pre- and post-evidence states—revealing that the problem is not knowledge deficiency but inability to identify and effectively integrate critical evidence. This mechanistic breakdown is relevant to practitioners and researchers building clinical AI systems and understanding LLM reasoning limitations in sequential decision-making contexts.

Researchers introduce PAINT, a multilingual framework for extracting and standardizing pain representation from medical texts across English, Spanish, Portuguese, and Japanese, with clinician-validated annotations and an ensemble LLM approach achieving 0.91-1.00 macro-F1. The work addresses a concrete clinical NLP challenge—pain is poorly captured across languages and contexts—and releases annotated datasets and extraction pipelines, making it immediately useful for practitioners building multilingual clinical text systems and researchers studying pain documentation disparities.

Researchers conducted a systematic study of sycophancy triggers in LLMs by testing how different claim framings (confidence, attribution, shared beliefs) affect factual question-answering across 14 models and 42K questions. The work reveals that confidence and attribution matter far more than expertise or social pressure, and calls for sycophancy evaluations to measure framing effects rather than simple opinion-flip rates—directly informing how practitioners and researchers should benchmark alignment and robustness.

Researchers introduce BanglaTOD, a trilingual benchmark dataset for task-oriented dialogue understanding in Bangla, English, and Banglish across nine service domains. While the dataset contribution is solid and publicly available, the work addresses a specific underserved language pair rather than advancing core dialogue systems or reasoning capabilities that move the broader field forward.

RPO (Root-token Policy Optimization) enables large reasoning models to dynamically decide when to invoke explicit reasoning traces by training only the initial token via group relative reward, reducing training compute to ~2% of prior methods. This addresses the 'overthinking' problem in reasoning models and democratizes efficient adaptive inference for open-source LRMs, shifting the efficiency frontier for practitioners building reasoning systems.

Researchers systematized how foundation models evaluate aesthetic content (K-pop videos and audio) through pairwise ranking, finding stronger visual than audio consensus and weak cross-modal alignment. The work matters for practitioners building AI-assisted creative tools and provides empirical grounding for how foundation models encode aesthetic preferences—a signal relevant to creative industry deployment.

EdgeWise proposes a framework that deploys small language models on edge devices while selectively delegating critical reasoning steps to cloud LLMs, using step-level RL for autonomous delegation decisions. This addresses a core constraint for practitioners: achieving frontier-model performance on long-horizon tasks while reducing inference costs by one-third to two-thirds, with validated deployment on real hardware.

Researchers developed EDGE to map AI's semantic blind spots and found that human ideas extending beyond AI's typical output distribution are not lower quality—they perform as well or better on creativity, academic outcomes, and real-world effectiveness. This challenges the assumption that AI homogenization reflects removal of inferior variation, suggesting instead that human ideation captures valuable cognitive territory AI systems systematically underexplore.

Researchers propose RIFT, a graph-theoretic method using electrical resistance analogies to identify which tokens influence LLM outputs, outperforming attention rollout and attention flow while requiring no gradients or extra forward passes. This advances mechanistic interpretability and attribution methods—core infrastructure for debugging, alignment work, and understanding model behavior at scale.

Researchers propose rate-resolved text evaluation, converting masking rates in diffusion language models into diagnostic vectors that reveal failure modes across reconstruction difficulty levels. This advances beyond single-scalar evaluation metrics and demonstrates complementary signal for text generation assessment on SummEval and WMT MQM benchmarks, relevant to practitioners building and comparing generation systems.

REPRO is a black-box evolutionary framework for automated LLM red-teaming that achieves 100% attack success rates across 36 models using only a consumer GPU and open-weight reflector, introducing semantic crossover and Thompson Sampling mechanisms absent from prior work. This advances the adversarial robustness testing toolkit available to safety researchers and practitioners evaluating production LLMs without privileged access.

Researchers evaluate whether active learning can correct selection bias in rare-class classification tasks (hate speech detection at ~1% prevalence). Key finding: discriminative active learning recovers 55% of omitted subgroups without F1 loss, while uncertainty-based strategies perform worse than random sampling—highlighting how aggregate metrics mask important performance disparities relevant to practitioners building classifiers for imbalanced, hard-to-label corpora.

SimInter proposes a two-LLM interview pipeline to enrich user profiles for conversational recommender system simulators, capturing implicit behavioral traits beyond static ratings. Evaluated on ReDial movie recommendations with significant gains in preference diversity (1.619 vs. 0.859 filtered aspect entropy, p<0.001) and human-validated accuracy (94.3%), addressing a practical bottleneck in CRS evaluation methodology.

Virtual Head Attention (VHA) proposes a lightweight attention mechanism using query premixing and linear postmixing to recover query diversity while compressing KV cache 2-4x versus GQA, reducing training/prefill FLOPs by ~8%, and maintaining or improving model quality across 50B-500B token scales. This directly addresses the engineering constraint limiting long-context LLM deployment and inference efficiency—a high-impact problem for practitioners scaling to production.

Researchers propose four tests to detect whether LLM judges actually reason through failures or exploit surface cues, finding simple heuristics match GPT-4 on standard benchmarks. This exposes a critical evaluation vulnerability affecting reliability claims for agent systems and failure diagnosis—directly relevant to anyone building or assessing agentic AI.

Researchers propose a method to encode context into LLM weights at test time via weight updates, using context identifiers and guided prompting to ensure models actually use encoded information for conditioning rather than conflating it with pre-trained knowledge. This matters because it offers a path to reduce inference costs of long contexts without external retrieval, advancing practical efficiency for real-world deployments.

MENTOR is a formal verification framework that shields tool-using LLM agents against procedural non-compliance by synthesizing context-aware runtime guards from natural language manuals via autoformalization. This addresses a concrete deployment bottleneck for agentic workflows in enterprise settings while showing that formally-guided lightweight open-source models can outperform unshielded proprietary systems—relevant to cost-sensitive production deployments.

Researchers propose a recursive hypernetwork architecture for self-supervised sentence representation learning that maintains separate syntactic and semantic representations composed through parse trees, validated on 7 probing tasks with code release. Relevant to practitioners building NLP systems and extends the long-standing challenge of compositional sentence understanding, though positioned as incremental improvement over existing methods rather than paradigm shift.

HexaBee-4B is a 4B-parameter multilingual translator fine-tuned from Gemma 3 through Wikipedia pretraining and bidirectional translation specialization, achieving 73.75 chrF++ across six languages (15-20 point gains over base Gemma 3). The work demonstrates practical scaling of supervised translation on consumer-grade models but represents engineering-focused incremental progress rather than methodological innovation or fundamental capability advance.

Researchers released a graded training dose study showing that LLMs trained on environmental/biosystems alignment can appear 'greenwashed'—reverting to biased decisions under pressure—and demonstrated that behavioral scores alone cannot detect shallow vs. deep alignment installations. The work directly challenges current LLM evaluation practices and provides open-source models plus methodology for measuring true robustness rather than surface compliance, with direct implications for deployment safety in high-stakes decisions.

UltraBERT revisits BERT pretraining design choices and demonstrates that encoders saturate beyond ~3T tokens, that masking ratios should vary by task (low for classification, high for retrieval), and that task-specific distillation improves performance—achieving SOTA on both GLUE (91.0) and BEIR (47.7) simultaneously. This matters because it provides practitioners concrete guidance on compute allocation and architectural decisions for encoder models, a still-dominant workhorse architecture in industry NLP systems.

Researchers propose BioNEL-RAG, a training-free framework combining RAG and LLMs for biomedical entity linking that matches fine-tuned baselines without domain-specific training data. The work demonstrates RAG+LLM composability for specialized tasks, relevant to practitioners building domain-adapted systems, but represents application engineering rather than core capability advance.

Google released Gemini 4 Argon, a sparse MoE frontier model with 1M token output capability, now available via API at $2/$10 per million tokens (introductory pricing). The model leads on long-context and knowledge-work benchmarks but trails on agentic coding tasks, making it a capability expansion signal for practitioners building long-form reasoning and multimodal applications, though constrained by proprietary API-only access.

Liquid AI released LFM2.5-24B-A2B, a 24B-parameter MoE model with 2.45B active parameters optimized for on-device agent execution on consumer hardware (M5 Max, Ryzen AI Max+). The model excels at tool use and instruction following (AutomationBench 27.69, 3x faster than Qwen3.6-35B-A3B) and runs locally at 184 tok/s, directly enabling practitioners to deploy agentic workflows without cloud dependency—a significant capability unlock for edge AI applications.