AI/ML Reading List

Google DeepMind and Isomorphic Labs launched a biosecurity programme called 'bioresilience' to prevent AI model misuse for biological threats while enabling legitimate biosecurity research. This represents a significant governance precedent for large capability labs addressing dual-use risks and aligns with emerging industry standards around model safety and responsible deployment in high-stakes domains.

PnP-CoSMo introduces a plug-and-play multi-contrast MRI reconstruction framework using content/style modeling that eliminates k-space training data bottlenecks while generalizing across contrasts—published in Medical Image Analysis with code released. The work demonstrates how explicit latent structure modeling can improve both performance and interpretability in inverse problems, relevant to practitioners building domain-specific reconstruction systems and researchers exploring generalizable priors for ill-posed problems.

ExTernD proposes expanded-rank ternary decomposition for post-training quantization of LLMs, decomposing weight matrices into two ternary matrices with a learnable diagonal scaling matrix to achieve arbitrary accuracy with minimal VRAM overhead. This addresses a core bottleneck in efficient LLM inference by making ternary quantization (extreme compression) practical without severe accuracy loss, directly relevant to practitioners optimizing model deployment.

The RTCA Workshop at NeurIPS 2026 (Sydney, December) solicits papers and demos on real-time conversational agents—streaming speech, video, naturalness, and live-system evaluation. This signals growing field interest in latency-critical, interactive AI systems beyond offline chat, positioning turn-taking, backchanneling, and perceptual naturalness as first-class research problems.

Independent researcher released DABSN, a recurrent language model architecture with preprint, PyTorch/C++/Triton implementations, and 24M-parameter LM results, now seeking collaborators for scaling and independent evaluation. Novel architecture approach to reasoning and long-context modeling with full reproducibility is relevant to practitioners exploring alternatives to transformer-dominant approaches.

A new inference harness called 'Schema' reports 99% accuracy on ARC-AGI-3 public benchmark by wrapping Claude Opus 4.8 and Fable 5 with process-level improvements (observation modeling, prediction testing, plan revision) rather than weight changes. This matters because ARC-AGI is a high-signal reasoning benchmark; if reproducible, the result demonstrates that post-hoc algorithmic scaffolding can substantially improve structured problem-solving without retraining.

A developer released EU AI Act OpenRAG, a 933-chunk SQLite dataset with BGE-M3 embeddings specifically structured around legal document hierarchy (articles, recitals, definitions) rather than sliding windows, demonstrating 9.2pp recall improvement on article retrieval vs. baseline chunking. This matters to practitioners building legal-compliance AI systems and researchers optimizing RAG performance on structured regulatory text, establishing a replicable benchmark for domain-specific document chunking strategies.

Researchers demonstrate that answer-conditioned chain-of-thought generation (showing models the correct answer and asking them to justify it) produces degraded training data that correctness filtering cannot detect, causing up to 27-point accuracy drops on hard reasoning tasks. This finding directly challenges a standard distillation recipe and provides actionable guidance: practitioners should generate reasoning chains answer-blind to avoid backward rationalization that masquerades as forward reasoning.

DS@GT ARC evaluated corrective RAG pipelines (CRAG + CiteFix) at CLEF 2026 LongEval and found frontier models optimize answer fluency over citation faithfulness, proposing evaluation metrics that enforce strict answer grounding to cited sources. This addresses a material failure mode in production RAG systems: high BLEU/ROUGE scores masking hallucination risk, directly relevant to practitioners deploying trustworthy QA infrastructure.

HABIB_TAZ achieved top ranks on SemEval-2026 Task 11 by using mDeBERTa fine-tuned on synthetic syllogistic data combined with multi-objective loss functions (DRO, bias penalty, KL regularization) to isolate formal reasoning from real-world plausibility biases across 12 languages. This directly addresses a known failure mode in LLMs—conflating logical validity with semantic plausibility—and demonstrates that structured training on synthetic logic can achieve near-perfect performance on formal reasoning tasks while maintaining low bias.

ReportMedSAM proposes a learnable concept bank approach to align free-form radiology reports with medical image segmentation, replacing brittle rule-based extraction with contrastive learning over BiomedCLIP embeddings and dynamic MoE routing. Directly relevant to practitioners working on clinical AI: demonstrates a scalable, extensible mechanism for handling linguistic variability in medical text and addresses the semantic robustness problem that limits production deployment of report-guided segmentation systems.

FindMyText is an open-source Python package that detects whether text appears in large web-crawled corpora using fingerprinting chains, outperforming existing methods on ArXiv, Wikipedia, and web datasets. Relevant to practitioners building datasets, validating training corpora for copyright compliance, and addressing reproducibility concerns in large-scale ML training.

Researchers optimized a 0.6B parameter model for on-device English-to-Traditional-Chinese subtitle translation by replacing the vocabulary projection layer with a domain-specific 64k-token tokenizer, achieving 1.63x speedup on Apple M2 and 59.2% win rate vs. Google Translate. The work is technically sound and open-sourced, but addresses a specific use case (subtitle translation) rather than advancing core ML architectures or reasoning capabilities.

Researchers conducted a 10-session longitudinal study with 24 participants using a memory-augmented conversational agent, finding that perceived memory (not just capability) drives self-disclosure and sustained enjoyment, while relationships exhibit discrete turning points traceable in behavioral signals. This informs how practitioners should design for repeated interaction, relationship persistence, and early intervention in user engagement patterns.

Researchers introduce SD-MAR, a framework combining synthetic data generation and reinforcement learning to improve Vision Language Models' analytical reasoning across multiple images. The approach achieves 36.95% in-domain improvement on custom benchmarks while maintaining or improving generalization on standard evals—a methodological advance relevant to practitioners building multimodal reasoning systems.

Team CoTu developed a Program-of-Thought pipeline combining neuro-symbolic reasoning (Z3 SMT solver for logic, Python for physics) with a 4B LLM backbone to achieve perfect scores on physics reasoning and 3rd place overall in EXACT 2026, a competition requiring explainable QA from small self-hosted models. This demonstrates that grounding LLM outputs in symbolic solvers can unlock reasoning capability at scale that would otherwise require much larger proprietary systems, with direct implications for practitioners building interpretable, deployable AI systems.

Researchers evaluated LLM-generated writing feedback across 20,000+ EFL student essays, comparing expert teacher ratings against actual student engagement and helpfulness. The key finding—low alignment between expert evaluation and learner utility—challenges how we assess AI systems for education and suggests learner-centered evaluation frameworks are essential for real-world pedagogical deployment.

Researchers identify the 'Severance Problem'—LLMs' inability to explicitly represent what they don't know about users—and propose the Severance Schema, a structured approach to encode ignorance that reduces sycophancy, hallucination, and harmful advice across major model families. This addresses a fundamental limitation in personal AI assistants and offers a practical intervention for improving alignment and reliability in user-facing systems.

First systematic cross-dataset generalization study for Urdu fake news detection reveals critical asymmetry in transfer performance driven by article-length confounds, demonstrating how shortcut learning distorts model evaluation. The reusable diagnostic methodology—combining bidirectional transfer analysis with confound detection—has methodological value for multilingual NLP robustness evaluation, though impact is constrained to low-resource language communities.

Researchers quantify 'semantic register compression'—a failure mode where intermediate agents in LLM cascades systematically compress semantic distinctions during text transformation, reducing label separability by up to 41.7% and shifting outputs toward particular distributional biases. This work directly impacts safety evaluation and architectural design of multi-agent systems in high-stakes domains like medical triage and fact-checking, offering practitioners concrete diagnostic methods and testable intervention points.

CoEvoT proposes a co-evolving chain-of-thought framework that dynamically updates graph token representations during LLM reasoning steps, rather than reasoning over fixed graph snapshots. This advances the practical capability of graph-LLM systems for distribution shift scenarios—a frontier problem combining structured reasoning with language model multi-step inference.

New arXiv paper formalizes information-theoretic ceilings on LLM reliability independent of scale, decomposing performance limits into resolvable context and inherent task ambiguity. Framework unifies scaling laws (recovers Chinchilla as special case), explains RAG benefits and catastrophic forgetting—providing practitioners and researchers a principled theory for predicting when scaling improves performance versus hitting fundamental limits.

Researchers introduce Token Time Continuous Diffusion (TTCD), a diffusion language model that operates in continuous space with per-token timing to improve generation speed and conditional modeling accuracy. The approach outperforms discrete diffusion baselines at high speedups and demonstrates gains on both language generation and structured tasks (Sudoku), representing a meaningful architectural contribution to efficient decoding in language modeling.

Researchers propose function-aware fill-in-the-middle (FIM) mid-training to improve coding agents by leveraging structural isomorphism between agent loops and function call sites, trained on 2.6B tokens from GitHub. The approach yields +2.8–+5.4 points on SWE-Bench across multiple models and post-training pipelines, with transfer benefits to non-agent coding and tool-use tasks, directly addressing a known capability gap in agent reasoning.

Researchers propose a lightweight machine unlearning framework that modifies RAG external knowledge bases to simulate forgetting without direct model access, particularly effective for closed-source LLMs like ChatGPT and Gemini. This addresses critical gaps in existing unlearning methods (high compute, catastrophic forgetting) and has direct implications for practitioners deploying LLMs with sensitive data retention risks and compliance requirements.

Researchers introduce Idea2Plan, a benchmark task and evaluation framework measuring LLM ability to translate research ideas into structured plans, grounded in ICML 2025 and Nature Mental Health papers post-cutoff. This addresses a systematic gap in understanding LLM reasoning for scientific workflows and directly informs development of autonomous research agents, a key long-horizon AI capability.

MedFailBench introduces a clinician-authored open-source benchmark that categorizes medical AI failures by safety gate type and severity rather than correctness alone, with 44 annotated synthetic cases and public leaderboard. This addresses a critical blind spot in medical AI deployment: current benchmarks don't surface failure modes (missed escalations, evidence fabrication, unsafe discharge) that matter most in clinical safety, making it immediately relevant to practitioners building and validating medical models.

Researchers propose Contrastive Policy Optimization (CPO), a method that replaces entropy-based advantage shaping in RL with token-level contrastive disagreement to distinguish genuine correctness from uncertainty. This addresses a fundamental limitation in reinforcement learning with verifiable rewards (RLVR)—particularly relevant for code generation, reasoning, and other tasks where correctness matters—and shows substantial empirical gains while maintaining generalization.

Researchers developed and validated GenAI-RTS, a 20-item psychometric scale measuring four theoretically grounded types of student reliance on generative AI in academic writing (Strategic, Instrumental, Dependent, Dialogic), with confirmatory factor analysis and measurement invariance across demographics. The work bridges AI literacy, educational integrity, and equity concerns—establishing a replicable assessment tool for researchers and educators to profile how students interact with GenAI rather than just adoption rates.

Researchers developed a multi-agent LLM framework using SFT, DPO, and RAG to simulate political coalition negotiations with partisan fidelity, introducing interpretability tools (Multi-Layered Information Lineage Topology, Coalition Influence Score) to trace policy origins and validate outputs against real agreements. This advances both LLM agent design for ideological consistency and computational political science, demonstrating how to audit emergent multi-agent negotiations for factual grounding and bias mitigation.

Researchers empirically demonstrate that OpenAI's GPT-2 detection model disproportionately flags autistic writing as AI-generated, despite neither corpus having high false-positive rates overall. The finding signals a critical bias in widely-deployed detection systems and raises governance concerns about automated flagging in academic integrity workflows affecting neurodivergent populations.

Researchers evaluated a LoRA-adapted Gemma model on 12,100 TOEFL essays across 11 L1 backgrounds and 8 unseen prompts, achieving 77.79% band agreement but discovering systematic scoring bias: essays from European-language speakers scored higher than East-Asian speakers within identical proficiency bands. This is the first large-scale L1 fairness audit of fine-tuned open-weight LLMs for essay scoring, surfacing critical deployment risks in educational AI systems.

Researchers show that LLM evaluators (reward models and judge prompts) exhibit significant language-dependent scoring bias, with lower-resource languages scored more generously despite 90%+ pairwise accuracy—creating a safety vulnerability where harmful content in low-resource languages is more likely to pass filters. The bias persists across eight open-weight evaluators and frontier models, is linked to model uncertainty, and cannot be mitigated by simple thresholding without language identification.

Researchers propose Budgeted Subset Refinement, a strategy for allocating limited computational refinement effort to improve LLM-generated research idea portfolios for diversity and executability. The work introduces a controlled proxy benchmark showing that selective refinement (especially diversity-aware MMR-k) outperforms uniform refinement on portfolio-level metrics—relevant to practitioners building AI-assisted research systems and teams evaluating how to deploy LLMs for ideation at scale.

ArogyaSutra introduces a multi-agent reasoning framework and large-scale multilingual medical QA dataset covering 21 clinical domains across English and seven Indic languages, addressing a critical equity gap in healthcare AI for rural India. The actor-critic architecture with dual-memory mechanisms and distillation improves medical reasoning accuracy across low-resource language settings—a technically rigorous contribution to underexplored terrain in specialized multimodal reasoning.

Researchers propose WavePhaseNet, a DFT-based method that reformulates Transformer attention through measure theory, proving hallucination is a structural inevitability and proposing cohomological regularization to suppress it via semantic hierarchy construction. The work connects embedding geometry, Zipf's law, and Hodge theory to derive a 3,000-dimension lower bound for LLM semantic faithfulness—directly addressing a field-critical problem (hallucination mitigation) with theoretical rigor and potential deployment implications.

Researchers demonstrate instruction tuning + model merging can improve reasoning language model performance on both verifiable domains (math/coding) and hard-to-verify domains (summarization) using existing supervised fine-tuning data, at sub-$3 cost. Relevant to practitioners scaling reasoning models beyond RL-tractable problem spaces and signals progress in making reasoning capabilities broadly adaptable across diverse domains.