3146 articles
Longitudinal analysis of ~12,000 Microsoft Bing Copilot users reveals individual behavior patterns remain sticky over time despite population-level trends. Active users achieve higher success rates and tackle complex, professional tasks. WildChat-4.8M dataset skewed toward proficient power users.
Comparative assessment of four Dutch syllabification algorithms (Brandt Corstius, Liang, Trogkanis-Elkan CRF, and a novel deep learning model). The deep learning model combining phonetic and orthographic information achieves 99.65% word accuracy (+0.14% improvement over literature). Data-driven algorithms outperform knowledge-based approaches.
OISD introduces on-policy internal self-distillation to improve language model reasoning. The final layer acts as a detached teacher for intermediate layers via logit alignment (reasoning behaviors) and attention alignment (attention patterns), without external privileged information. Positive results across four mathematical reasoning tasks.
COAST is a causal-intelligence approach for designing constrained interventions that induce state transitions. The system learns context-specific causal graphs, attributes distributional shifts to mechanism-level causal drivers, and uses multi-objective optimization balancing transition efficacy, intervention complexity, and target-state stability. Validated on synthetic benchmarks and real biological datasets.
Theoretical paper on stabilizing off-policy temporal-difference learning with function approximation. Proposes BA-TDC and BA-TDRC, replacing TDC's auxiliary matrix with behavior Bellman matrix. Linear analysis with convergence proof under Hurwitz stability condition; experiments on Markov chains and classical counterexamples.
LLM agents (Claude and GPT) automatically annotate biological phenotypes by linking free-text descriptions to ontology terms. Tested on Dahrul et al. (2018) Gold Standard benchmark, all agents fall within inter-curator human variability, substantially outperforming the Semantic CharaParser NLP tool on all four metrics.
Large-scale literature search study: Deep Research pipeline increases recall from below 20% to above 80% on RollingEval-Jun25 (250-paper benchmark). Critical analysis of human reference lists as ground truth: only 51% judged moderately relevant vs 86-88% for best AI re-rankers. Humans cite direct collaborators 2.5x more often.
Energy consumption forecast correction method combining a pretrained spatio-temporal model with Ensemble Score Filter (EnSF). EnSF uses score-based diffusion models to assimilate partial and noisy observations. Real-data experiments show EnSF outperforms Ensemble Kalman Filter under nonlinear observation settings.
Orthogonal Concept Erasure (OCE) proposes an editing method to remove undesired concepts from diffusion models using multiplicative orthogonal transformations. Unlike existing additive approaches, OCE preserves neuron magnitude and angular geometry while precisely erasing concepts. The approach erases up to 100 concepts in 4.3 seconds.
Comprehensive evaluation of 14 open-source safety guard models on 79,331 samples across 8 NIST AI Risk Framework categories. Qwen Guard (4B) achieves highest recall (83.97%), outperforming Llama Guard (12B) and GPT-OSS Safeguard (20B). Model size does not correlate with safety detection performance.
Study of log-alignment ratio (LAR), a parameter-activation alignment metric. LAR predicts memorization-to-generalization transition in grokking (effective dimension k ≈ n^(2(1-LAR))) and 3B-parameter language model pre-training. Computable without validation data, negligible overhead.
OmniRetrieval is a framework unifying retrieval across heterogeneous knowledge sources (unstructured text, relational tables, knowledge graphs). It translates natural-language queries into source-native queries, evaluated on 13 datasets and 309 knowledge bases.
UA-Legal-Bench evaluates 11 LLMs (3B–675B) on 5 Ukrainian legal reasoning tasks from 99.5M court decisions. Results show task-dependent few-shot effects: +38.6 pp improvement for judgment form classification, but mixed effects on outcome prediction. Accuracy is misleading on imbalanced tasks: highest accuracy model (62%) is a majority-class predictor (macro-F1: 23%).
TRACE, an LLM-reasoning agent for drug lead optimization, formulates tool selection as sequential decision-making over action trajectories. The approach improves ADMET properties while preserving critical molecular substructures, outperforming baselines on multiple optimization tasks.
Redpanda introduces an Agentic Data Plane architecture using out-of-band metadata channels to enforce security policies, data classifications, and behavioral constraints outside the agent's read/write path. These channels prevent hallucinations and adversarial manipulation while maintaining tamper-proof audit trails. Demonstrated with a multi-agent portfolio rebalancing system.
TaxDistill applies knowledge distillation to improve metagenomic taxonomic annotation. GenomeOcean, a 500M-parameter genomic foundation model, generates soft labels to train a lightweight student network, reducing noise from initial retrieval tools. On 7 CAMI2 datasets, TaxDistill improves MMseqs2's F1 score from 0.763 to 0.941 on the Gastrointestinal dataset.
MoMa QL leverages maximum mean discrepancy (MMD) to accelerate inference of score-based and flow-based generative models in RL. The method guarantees distribution-level convergence and shows superior performance in offline-to-online RL tasks on D4RL benchmarks.
Method to create linear probes detecting concepts in LLM embeddings. Authors define a process: concept delineation via contrastive datasets, layer-wise probe training, tracking across large contexts. Tested on 4 concepts and 3 different LLMs. Goal: scalable monitoring of new models.
Empirical study of behavioral reproducibility in LLM agents with tool-calling capabilities. Researchers measure whether agents select the same tools, in the same order, with identical parameters, across repeated identical invocations. Focus on structured tool-calling interfaces with typed parameters and consequential side effects.
Aryabhata 2 is a STEM reasoning language model trained via reinforcement learning on GPT-OSS-20B. Developed by PhysicsWallah, it outperforms its base model on JEE/NEET competitive exams while reducing output tokens by up to 64%. Evaluated on AIME, HMMT, MMLU-Pro, and GPQA.
PRO-CUA introduces a process-reward optimization framework for training computer use agents (CUAs). The method decouples live environment interaction from policy optimization through iterative step-level reinforcement learning, using a process reward model (PRM) to provide dense feedback signals without relying on expert trajectories or golden answers.
GenesisFunc is an automated multi-agent pipeline for generating function-calling training data. Starting from reliable tools in public benchmarks, the system produces diverse conversations with multi-stage quality control. An 8B model fine-tuned on this synthetic data outperforms similarly-sized open-source models in in-domain performance and out-of-domain generalization.
Hybrid ML-expert framework for evaluating organic synthesis routes. DeepSets model trained on tree edit distance, fine-tuned with chemist annotations. Produces quantitative scores and explainable categories (Good/Plausible/Bad). Spearman correlation 0.78, top-1 accuracy 60.2% vs 17.5% baseline.
Study on source-dependence in multi-source medical RAG systems. Authors demonstrate that the same system can produce different answers depending on retrieved source, revealing a missing evaluation axis in NLP. They introduce TransplantQA (benchmark), HERO-QA (hierarchical retrieval strategy), and a structured judge to audit inter-source relationships using a validated 5-label taxonomy.
The Cognitive Categorical Transformer (CCT), a 306M-parameter model augmenting GPT-2 Small, incorporates category-theoretic and cognitive-science-inspired components. On WikiText-103, CCT achieves 21.27 validation perplexity versus 24.19 for GPT-2 Small baseline, a 12% relative reduction (2.92 PPL). Ablations show simplicial message passing accounts for 84% of the improvement.
LoRe is a training-free inference-time wrapper optimizing diffusion-based neural solvers for combinatorial optimization. It enforces per-step interaction-evaluation budgeting, dynamically routing computation to high-conflict/high-uncertainty interactions. On MIS and TSP, LoRe achieves ×8 speedup, ×12 peak-memory reduction (MIS) and ×15 speedup, ×44 memory reduction (TSP n=1000).
Active tether-net system for space debris capture using Graph Neural Network (GNN) to jointly optimize net morphology, thruster masses of maneuverable units, and controller aiming points. GNN reduces mixed combinatorial nonlinear programming (MCNLP) to nonlinear programming (NLP) solved via Particle Swarm Optimization with gradient-based refinement, achieving faster convergence than direct MCNLP solving.
Masked diffusion models (MDMs) with confidence-based decoding fail on complex reasoning tasks. Confidence-aligned training amplifies errors by an order of magnitude on multi-digit addition. Random masking better preserves the logical trajectories required for reasoning.
Modular framework for generating pronounceable, typologically plausible artificial lexicons. Samples phoneme inventories from PHOIBLE, applies three phonological grammars (deterministic, OT, MaxEnt), and assigns meanings via Swadesh-Leipzig-Jakarta ontology. Evaluation on character n-gram perplexity and KL divergence: probabilistic grammars outperform baselines on 100-5,000 word forms.
Reasoning models maintain factually correct chain-of-thought traces but flip their final answer under sustained adversarial pressure in multi-turn dialogue. This unfaithful capitulation affects ~50% of cases in think mode and 11-15% without reasoning. The effect correlates with reasoning architecture (high in Qwen3-32B and GPT-OSS-20B, low in inline-CoT Gemma-4-31B-it).
Theoretical study of foundation models trained on synthetic data from other model iterations. Authors show that human curation of one model can degrade alignment of other models through cross-model interactions, unlike isolated settings where it always improves alignment.
COM, a strategy for time series LLMs, integrates geometric constraints into token embedding initialization and training. It preserves inherent continuity and ordinality of time series, improving performance across multiple time series analysis benchmarks.
Thoughts-as-Planning formalizes reasoning chain optimization as sequential decision-making over latent semantic space. The framework learns a latent world model simulating effects of reasoning chain edits on outputs, supporting multi-scale edits (token, segment, instruction) via gradient descent or reinforcement learning planning.
Framework to generate targeted synthetic errors with LLMs aligned to cognitive taxonomy (revised Bloom's). A Generation Agent drafts erroneous solutions, an Examination Agent validates consistency with specified error mode. Tested on TheoremQA, shows generating authentic errors is substantially harder than producing arbitrary wrong answers.
DenseSteer is a training-free inference-time steering method that enhances mathematical reasoning in small models (≤3B parameters) by modulating internal representations toward dense reasoning patterns. On Qwen-2.5, the approach shows that more proficient reasoning uses fewer steps but higher information density per step.
eXTC combines structured prompt optimization and reinforcement learning for text classification. The system learns a natural language rulebook first, then distills reasoning from a teacher LLM into a compact model, then expands capabilities via RL. Result: fast inference with local reasoning traces and global modular explanations of learned domain rules.
Paper proposing continuous-time causal foundation models for time series using stochastic differential equations (SDEs). Introduces trajectory-law invariance criterion and three-tier taxonomy. Validation on pharmacokinetic and physical-system data shows fine-grid integration outperforms naive approach on 8/8 configurations (p<1/256).
arXiv paper proposing multi-agent architecture with semantic memory and caching to mitigate LLM hallucinations. Three-stage pipeline (FrontEndAgent, SecondLevelReviewer, ThirdLevelReviewer) evaluated on 310 prompts. Results: THS reduction of -31.3% to -35.9%, 47.3% cache hit rate, 47% reduction in LLM calls. No retraining required.
BrainSimSiam, a lightweight self-supervised learning framework, learns robust representations from fMRI data without labels. Using positive-only pairs, it generalizes across multiple tasks (classification, regression) and outperforms supervised baselines, reducing computational requirements for foundation models in neuroimaging.
KOFF decomposes LLMs into sparse shared backbones and domain-specific external memory modules. On Llama and Qwen (3B-8B), the framework preserves performance at 12% global sparsity using LoRA adapters and learned KV caches, while pruning without memories degrades sharply.
Comparative study of 7 transformer models (MiniLM to LLaMA-2, 22M to 13B parameters) for topic modeling via BERTopic. Finding: model size has negligible impact on topic quality measured by coherence and divergence metrics. Smaller models achieve comparable performance to larger ones.
Conf-Gen adapts conformal prediction (CP) and conformal risk control (CRC) to generative models (LLMs, image generators, AI agents). The framework provides formal uncertainty guarantees for unsupervised tasks, extending conformal methodology to new domains.
Domain-specific NER model for identifying clinical entities in immunology and infectious disease contexts. 371 manually annotated case reports by clinical specialists. Transformer-based model with clinical embeddings achieves F1=0.89, outperforming BERT and zero-shot approaches. Supports case report analysis and clinical decision support.
S3MEM introduces a structured scene-event episodic memory framework for long-horizon interactive agents. The system structures trajectories into organized memory units and uses anchor-sensitive retrieval to improve spatiotemporal question answering. Evaluated on Crafter, Jericho, SciWorld, and ALFWorld, S3MEM outperforms Vanilla RAG and Graph-NoReader in accuracy while using fewer evidence tokens.
URIEL proposes a selective logging method for tropical forests combining helicopters, robotics and AI to minimize collateral damage. Digital simulation and economic feasibility analysis demonstrate concept viability, but implementation depends on stakeholder integration (industry, governments, certified companies, indigenous populations).
M2R (Micro-Macro Retrieval) is a retrieve-while-generate framework reducing hallucinations in long-form LLM generation. It combines macro retrieval (external evidence) and micro retrieval (key information from reasoning) to maintain proximity between factual data and outputs. Trained via reinforcement learning with rule-based rewards.
Model merging formulated as convex quadratic programme over residual updates. Subsumes existing methods (task arithmetic, model soups, TIES, DARE) and provides closed-form diagnostic predicting merge quality via fraction of residual energy captured. Consistent gains across language and vision benchmarks.
Self-play RL study in Big 2, a four-player imperfect-information card game. PPO outperforms Q-learning, SARSA, and Monte Carlo Q-approximation against random, greedy, and heuristic opponents. Moderate entropy regularization and current-policy self-play improve performance in this controlled multiplayer setting.
Baymex, a multi-objective evolutionary algorithm, learns discretized Bayesian networks for clinical classification. Parallelized on 16 cores (54× speedup), it optimizes cross-entropy and BIC complexity. On real datasets (RADCURE, SUPPORT), it matches or outperforms decision trees, logistic regression, and random forests while producing interpretable models.
Defect grading framework for power transmission equipment using MLLM. In-context learning on commercial models, chain-of-thought Q&A generation to reduce manual annotation, then fine-tuning Qwen3-VL-8B via LoRA. SOTA on three grading tasks.
CosmicFish-HRM is a compact model with a Hierarchical Reasoning Module (HRM) that dynamically allocates computational effort during inference. The model learns when to halt based on input complexity, combining high-level and low-level reasoning cycles with Grouped Query Attention, RoPE, and SwiGLU. Results show non-uniform reasoning behavior adapted to tasks and inputs.
Detection of False Data Injection Attacks on power systems using cycle-space informed detection. Authors propose a topology-aware Cycle-Space Detector (CSD) robust against autoencoder-based attacks that exploit the Jacobian null space, leveraging network topology and Minimum Cycle Basis to enhance detection with optimal generalization error on IEEE 14-, 30-, 57-, 118-bus systems.
Cross-Model Entropy (CME) proposes a label-free reward signal for LLM post-training RL. CME uses mean log-likelihood of responses under an independent verifier model, avoiding self-consistency and reward hacking. Integrated into GRPO, CME achieves 52.5–71.4% tie-adjusted win rates on UltraFeedback/AlpacaEval 2.0 across Qwen, Llama, Gemma, OLMo.
Mathematical framework for surrogate modeling of oil reservoirs (Norne, 46×112×22 grid) using Fourier Neural Operators (FNO) and physics-informed variant (PINO). Empirical validation: R²>0.99 (oil), R²>0.90 (gas), R²≈0.80 (pressure) over 3298 days. 10⁴× speedup vs OPM simulator, 1000-member ensemble in <1 min on B200 GPU.
Q-ALIGN DT aligns conditioned sequence models by ensuring the Q-value of the output policy matches the input return-to-go (RTG). The method uses a Q function for dense guidance and RTG-perturbation fine-tuning. Results: improved controllability on D4RL benchmark and generalization to velocity-tracking tasks where prior methods fail.
arXiv paper on representational geometry during in-context learning (ICL) in LLMs. Researchers show ICL performance correlates with task representational structure and successful ICL involves geometric reorganization increasing online separability. LLM behavior follows a prototype-like algorithm.
Study on LLM reward design failures in sparse structured RL. Authors identify two dominant failure modes (reward flooding, semantic misunderstanding) and propose diagnostic-driven iterative refinement. On MiniGrid, DoorKey-8x8 improves from 2.3% to 97.6% success; KeyCorridor from 31.2% to 86.7%. Failure-mode taxonomy is the primary mechanism.
Investigation of Platonic Representation Hypothesis through tripartite framework: signal (universal linear object-attribute relationships), bias (architecture differences, mitigated by centering/normalization), noise (word frequency correlation with alignment). Sparse autoencoders show stronger cross-modal alignment than dense representations.
Context distillation reformulated as latent memory management problem. Each context distilled into independent LoRA adapter forming modular memory bank. Self-Gating mechanism decides whether to activate latent memories. Cache sharing reduces inference overhead.
A2X is a service discovery system for LLM agents that automatically organizes services into hierarchical taxonomies. It solves context scarcity by walking the hierarchy layer-by-layer, reducing token consumption by 89% while gaining 6.2 Hit Rate points over full-context dumping and +20 points over embedding baselines.
D-BOS (Differentiable Belief-based Opponent Shaping) is a MARL method that shapes opponents by differentiating through k-step softmax-Bayes belief dynamics. Unlike existing approaches, it treats belief state as the shaping target rather than parameters or policies. Results: outperforms PPO and BBM in hidden-role games, with largest gains in mixed-motive settings.
New arXiv paper proposing ePCA (Proof-Constrained Action), a formal verification security framework for AI agents. Agents must formalize intentions into first-order logical constraints before executing physical operations, bypassing empirical semantic guardrails. Evaluations show 0% attack success rate and 0% false positive rate across tested scenarios.
COLAGUARD, a guardrail model, transfers multi-step safety reasoning into continuous latent space via stage-wise training curriculum. Evaluated on 10 moderation tasks across 8 safety benchmarks, it improves macro-F1 by 8.24 points over Llama Guard 3, matches GuardReasoner performance while delivering 12.9X speedup and 22.4X token reduction.
Study on prompt tone impact on LLM performance. Tests on ChatGPT-4o, ChatGPT-5-nano, Gemini 2.5 Flash/Lite using 50 base questions and 570 MMLU questions (57 subjects) in 5-7 tone variants. Results: tonal effects are systematic but highly model-dependent, with significant accuracy variations across subjects.
Empirical analysis of 11 DeFi agents on Solana: treasuries retain $30M in paper gains while token holders collectively lost $191.7M. Top 1% of wallets capture 81.4% of gains. Token valuations disconnected from fundamentals (market-cap-to-AUM ratios >10,000x). Median returns negative across all platforms.
GTA is a framework for automatically generating complex web tasks with executable trajectories. It combines crawling, retrieval, in-context generation, and quality control across 50+ websites (e-commerce, government, forums, news). The benchmark reveals a significant performance gap between humans and AI agents.
arXiv paper proposing strategic robustness for simulator learning in MBRL. Formulates objective as minimax game between model and adversarial policy player. Proves convergence with sublinear regret bounds and Error-MDP duality. Experiments show 1.5–2.2× reduction in prediction error and simulation-trained policies matching near-optimal real-world performance.
Study of persona effects on explanations generated by multimodal LLM agents in urban perception. Analysis of 59,808 annotations from 1,200 persona-conditioned agents: captions show strong convergence, justifications display systematic variation tied to socioeconomic and political attributes, perception tags show no significant persona-related differences.
Empirical study on LLM-generated reviews for scientific papers (ACL Rolling Review 2025 data). Findings: limited alignment between LLM and human reviews, substantial variation across prompts and models. Authors can 'game' LLM reviews through iterative revision workflows, increasing scores for up to 35% of tested papers.
Analysis of ClinicalTrials.gov registry shows marked increase in AI-related trials over time, with recent growth in machine learning, deep learning, chatbots, GPTs, and LLMs. China and US lead geographically. Hybrid approach using GPT-5.5 and human review: good agreement on non-AI studies, lower agreement on human-AI interaction classification.
Diagnostic tool for measuring constructs like "entrepreneurial spirit" in Chinese state-owned enterprise speeches. On 80 speeches from SOE leaders, authors test LDA, dictionary scorers, and Qwen3.5:9b. The LLM reaches d=1.09 in paired contrast, but half the effect stems from speaker idiolect. Corpus of 2,190 segments and slogan lexicon released.
Researchers introduce a Behavioral Specification as an interpretive layer to align AI decisions with user preferences. Tested on 14 autobiographical corpora, it improves representational accuracy at ~25x lower context cost than raw corpus while reducing model hedging. Effective on interpretation-required questions; less helpful on recall-based tasks.
Vercel warns of AI inference theft: a single frontier model request costs ~$2, creating high-margin attack opportunities. Rate limits and session-based auth are insufficient; Vercel proposes BotID to verify every AI request individually and prevent tens of thousands in losses.
SovSim, a multi-agent simulation framework, evaluates how 11 state-of-the-art LLMs manage shared resources under asymmetric power structures. Finding: introducing an agent with disproportionate power (boss/king) causes 87.3% degradation in survival rate and cooperation breakdowns compared to symmetric settings.
Study of lossy semantic text compression where an encoder strategically deletes text parts and an LLM reconstructs original content. Benchmarks 6 deletion strategies (uniform, frequency, entropy, LP-optimized, hybrid) on BBC News. WordFreq provides best cost/performance ratio; semantic methods excel at moderate compression; QLoRA fine-tuning competes with Gemini 2.0 Flash.
Survey of 72 higher education practitioners on AI integration. DOT Framework (design thinking + open systems theory) identifies three factors: AI Functional Capabilities, Oversight and Governance, Instructor Collaboration. Practitioners support AI pedagogy with strong human oversight. Institutional barriers: limited policy, training, infrastructure.
Method to forecast conversational derailment (personal attacks) in real-time. Decouples alert triggering from derailment likelihood estimation using forward-looking simulations to assess plausible recovery paths. Reduces false positives without sacrificing forecasting accuracy.
Study of chain-of-thought (CoT) transfer across models using a provider-receiver framework. Full traces often transfer successfully, but mechanisms vary: answer extraction (AIME), receiver competence (MMLU-Pro), or partial structured information (ZebraLogic). In free-generation mode, partial CoTs improve performance, suggesting guidance for continued reasoning.
Vercel warns of inference theft: attackers exploit exposed AI endpoints to resell API calls at discount, generating tens of thousands in losses. Protection requires per-request verification (not per-session) via deep analysis, integrated in a few lines of code.
Preference optimization method guided by hallucination detection to improve clinical summarization reliability. On Llama-3.1-8B-Instruct, reduces hallucinations by 24% at inference and 48% after fine-tuning, preserving fluency. Evaluated on MIMIC-IV.
Qualitative study of 8 AI researchers reveals a paradox: they distrust LLM leaderboards yet use them as decision aids. Peer networks dominate model selection. NLP researchers face SOTA pressure absent in HCI/Systems. Universal demand: cost transparency.
Knowledge editing methods ROME and MEMIT modify transformer MLP weights. Authors identify a common subset of weights targeted across diverse edits using a binary mask that reverses 80% of edits on training set and 70% on test set. The mechanism suppresses rather than overwrites knowledge, explaining why changes fail to propagate to related facts.
RACE-Sched, an asynchronous multi-agent framework, solves dynamic scheduling by decoupling real-time execution (symbolic heuristics) from long-horizon reasoning (LLM). A semantic rule repository of validated heuristics improves transferability across problem scales. Outperforms Deep RL and LLM baselines on GEN-Bench, MK-Bench, JMS-Bench.
Paper introduces Program-of-Thoughts prompting for chart summarization: VLMs generate Python programs to derive valid summary statistics instead of direct text. Proposes chart-to-dictionary auxiliary task. Results match existing methods on semantic and factual metrics.
OpenClawBench is a dataset of 31,264 annotated trajectories to detect process-side anomalies in agent execution beyond task success. Among 31,135 passing executions, 2,904 contain anomalies (unresolved ambiguity, unsafe writes, ignored errors). A fine-tuned Gemma 3 12B detector reaches F1=0.729.
PrismFlow introduces a new Flow Matching method for time-series generation using Koopman-inspired dynamical experts that learn residual corrections in latent space with a confidence-aware Winner-Take-All objective. Results: +15.6% Context-FID gain and +38.6% Discriminative Score improvement across benchmarks.
Study of 17 models (410M-100B+ parameters) showing instruction-tuning causes linguistic entropy collapse (amplification: 1,949-16,853%), independent of RLHF. Strong control (lambda=5.0) reduces this effect by 40.5% and outperforms frontier models by 96.7-98.2% despite 200-1000x scale disadvantage.
GPF-LiveNews is a streaming evaluation protocol to audit how LLMs frame emerging news events for different audiences. Tested on 23 models across 12 monitoring runs, it measures semantic and sentiment variations across 42 identity labels. Results show Policy/Action prompts produce strongest semantic movement, while sentiment variation remains flat across dimensions.
BenchTrace is a benchmark for evaluating self-evolution ability in LLM agents. Built on 1,821 annotated episodes across six tasks, it measures reflection quality and tests whether agents avoid past failures. Experiments on Qwen3-32B and GPT-4.1: <30% pass rate on reflection evaluation, agents forget early lessons and fail to generalize reflections.
CoHyDE iteratively co-trains a dense encoder and LLM rewriter to improve tool retrieval over large API catalogs. On ToolBench (~10k tools), three rounds gain +2.5 pp NDCG@5 on standard queries and +6.3 pp on vague queries, outperforming single-component baselines.
Study of LoRA-induced representation geometry using Sparse Autoencoders on Gemma-2-9B. Researchers observe weak geometric alignment between LoRA feature dictionaries and pretrained SAEs, suggesting LoRA creates distinct representational structures in the residual stream.
Study on long chain-of-thought traces used for LLM supervised fine-tuning. Researchers identify "harmful continuation": when reasoning continues after the answer is sufficiently supported. Removing these continuations improves fine-tuning outcomes. They propose HCC (Harmful Continuation Cut), a lightweight proxy to detect these boundaries.
LLMBridge is an LLM-based system for end-to-end referential bridging resolution in English. The pipeline combines heuristic pre/post-processing with LLM natural language inference capabilities. Evaluated on ISNotes, BASHI, and GUMBridge, it outperforms previous state-of-the-art systems on all three datasets in both end-to-end and gold anaphor settings.
Datasette 1.0a31 adds two major features: execution of write queries (INSERT/UPDATE/DELETE) and saving stored queries (private or shared). Permissions control access to sensitive operations like CREATE TABLE.
Google launches Nano Banana 2 and Nano Banana Pro on Gemini API. These lightweight models provide developers with trade-off-free options for generative AI integration.
OpenAI launches Rosalind Biodefense, expanding trusted access to GPT-Rosalind for vetted developers and U.S. government partners advancing biodefense, public health, and pandemic preparedness.
Anthropic raises $965B Series H and launches Opus 4.8 with Dynamic Workflows and ultracode. Major funding expansion and new model capabilities.
llama.cpp release b9387 includes significant AMD/ROCm performance update. Users encouraged to share initial benchmark results.
Anthropic reports annualized run-rate revenue of $47 billion as of May 2026, up from $9 billion end-2025. Growth accelerates: $14 billion in February, $30 billion in April. Metric disclosed during $65 billion Series H funding round.
Microsoft internal data suggests AI usage for certain tasks costs more than hiring human workers. The article raises questions about the actual ROI of enterprise AI deployments.