3147 articles
Agyn is an open-source platform for deploying AI agents in production. It provides a stateful serverless runtime on Kubernetes, agent definition via Terraform, and a zero-trust security model. Agyn is agent-, model-, and cloud-agnostic.
MERIT, a multimodal pretraining framework, combines masked ECG modeling with ECG–text contrastive alignment to learn cardiac representations. On PTB-XL: +3% F1 (All) and +5% F1 (SubClass), +2.66% AUC zero-shot. Also improves clinical text generation with LLMs.
Modular LLM-based architecture to detect and quantify intensity of human values in text. Three coordinated modules: generating value specifications from theoretical frameworks, labeling texts, assigning graded support/resistance based on rhetorical and semantic evidence. Evaluated on ValueEval dataset with multiple LLMs, demonstrating pipeline generality.
SC-SDPO improves LLM self-distillation by weighting losses with √[p(1-p)], creating an implicit curriculum. Experiments on Qwen3-8B (+3.2/+4.3 mean@16/maj@16) and OLMo-3-7B (+1.8/+3.0) show stable gains with zero computational overhead.
Hierarchical framework for compact LLMs in resource-constrained agentic systems. Model distillation + oracle-controller loop monitors protocol validity, projects histories into feasible prompt domain, triggers lightweight fine-tuning under drift. Separates schema learning from semantic adaptation. Evaluated on Multi-Fidelity Bayesian Optimization with improved reliability and cost-efficiency.
Paper proposing SMARt, a formal framework for managing autonomy in agentic AI systems. Introduces managed autonomy theory based on epistemic drift detection, reasoning suspension, and escalation to human control. Uses timed Petri nets to guarantee safety and governance properties.
LaneRoPE enables multiple sequences generated in parallel to collaborate during inference. The method adds inter-sequence attention masking and extends RoPE to capture relative token positions within and across sequences. Tests on mathematical reasoning show accuracy gains with negligible overhead.
MTM-Bench, a controlled benchmark for multilingual task execution, evaluates 20 LLMs across 27 language triplets (instruction/content/response) in English, Spanish, and Chinese. Results show degradation is organized by language role in task structure, with response language as the dominant axis of variation.
EvoSpec improves speculative decoding by dynamically adapting draft model vocabulary and parameters in real-time. Using semantic indexing and curriculum learning, it maintains high acceptance rates across specialized domains (coding, law, medicine). On EAGLE-3: 1.13x speedup vs FR-Spec with 27% lower memory overhead.
Simorgh proposes a region-aware hybrid retrieval approach combining BM25 lexical matching and dense semantic similarity for culturally grounded multilingual QA on BLEnD benchmark (30 languages). Uses quantized Qwen3-14B with logit-based answer selection. Improves cross-lingual stability but reveals performance gaps tied to training data imbalance.
RL framework for sim-to-real policy transfer via probabilistic latent embeddings and dynamic adaptation. Uses meta-RL and CMDPs to infer latent environment representation, with distributional RL formulation dynamically adjusting risk levels based on latent context estimation accuracy.
Gradient Transformer, a data-free knowledge distillation framework, generates LLM update vectors from TinyLMs fine-tuned on private data. The model captures correlation between gradient vectors of both models, enabling collaborative adaptation without accessing sensitive data.
Paper introduces a dataset for multivariate time series anomaly detection in federated learning. Addresses gaps in existing benchmarks by providing data with cyclic dynamics from discrete industrial automation processes, evaluated against selected MTSAD methods.
Multi-agent autonomous architecture for real-time insight discovery over data streams. Continuous loop: hypothesis generation, compilation into executable analytics, validation, visualizations. Uses Kafka, Flink, LLM. Contract-driven design with typed artifacts for modularity and lineage. Demonstrated on retail, finance, public data.
Variance-Regularised Pruning (VR) is a model compression method for affective computing that prioritises cross-user stability over sparsity alone. Tested on the AGAIN dataset (9 game environments), VR maintains competitive CCC performance at 80% sparsity without additional fine-tuning, suited for resource-constrained embedded systems.
OralAgent is a dental-specialized AI agent integrating multimodal reasoning, 22 visual analysis tools, and RAG over 368 classical dental textbooks (134.8M tokens). Evaluated on OralQA-ZH (798 questions) and MMOral benchmarks, it achieves SOTA for dental image analysis in clinical workflows.
Systematic survey on Mixture-of-Experts (MoE) applied to multimodal learning. Analyzes MoE from three perspectives: efficient engine (scalability, redundancy reduction), representation learner (multi-expert alignment), modular adapter (modality imbalance, missing data). Identifies gaps: interpretable routing, expert communication, modality integration, lifelong learning.
Comparative study of Liquid Neural Networks (LNN/CfC) vs LSTM across four sequential modalities (N-MNIST, QuickDraw, IAM, PhysioNet Sepsis-3). LNNs model hidden state evolution as continuous differential equations. Results: LNNs outperform LSTM in parameter efficiency and robustness to missing data, especially in clinical environments.
SparseOpt, a sparsity-aware optimizer, addresses gradient skew induced by Batch Normalization in Dynamic Sparse Training. Experiments on ResNet (CIFAR-100, ImageNet) show faster convergence and improved generalization. First systematic study of interactions between Batch Normalization, sparse layers, and DST.
SignGAD introduces a self-designing agentic framework for few-shot graph anomaly detection. Instead of fixed pipelines, it designs task-conditioned detection workflows by selecting suitable graph encodings and detector designs. A guarded refit strategy refines selected workflows under limited supervision, outperforming state-of-the-art methods on real-world datasets.
HEAL is a decentralized learning framework combining federated learning, gossip learning, and epidemic learning through a self-organizing P2P overlay. It dynamically promotes nodes as aggregators and demonstrates performance equivalent to standard federated learning while being resilient to failures and churn.
TRACES is a proactive safety auditor for multi-turn LLM agents that detects drift toward unsafe behavior from hidden representations of an observer LLM. Trained with weak trajectory-level supervision, it produces dense prefix-level risk estimates, improving full-trajectory safety prediction and proactive risk discrimination across multiple agent safety benchmarks.
Theoretical paper decomposing the pre-softmax attention matrix QK^T into symmetric and skew-symmetric components. The symmetric part governs the energy landscape, the skew-symmetric part drives circulation. Authors propose Hopfield-style stability measures to quantify fidelity-diversity trade-offs in generation and a controllable mechanism to modulate this trade-off.
Systematic study comparing state space models (SSM) for time series classification. S4D outperforms Mamba variants in accuracy and efficiency. Authors introduce MS4 and MS4N, lightweight S4D variants with linear input projection and channel-mixing. Evaluation on 59 datasets (MONSTER, UEA): MS4N matches models 10× larger in parameters.
Personalized Observation Normalization (PON) method for federated reinforcement learning in heterogeneous environments. Each agent locally normalizes state inputs using continuously updated running mean and variance, preventing imbalanced parameter aggregation issues. Experiments on heterogeneous MuJoCo tasks demonstrate accelerated training and superior performance versus baselines.
Study of four OLMo 32B checkpoints showing post-training (SFT, DPO, RLVR) compresses narrative variation by reducing thematic transitions, emotional intensity, and stylistic diversity. The 'narrative flattening' effect is stronger on professional literary fiction than on public-platform stories.
GeoTransolver, a geometry-aware operator learning framework, accurately predicts industrial-scale automotive crash dynamics. On bumper beam and full-vehicle crash datasets, it captures plastic deformations and acceleration profiles. A FLARE-based modification reduces memory overhead by 2x while improving accuracy for high-frequency transients.
VibeSearchBench evaluates LLM agents on collaborative multi-turn search in real-world context. The benchmark comprises 200 bilingual (Chinese/English) tasks across 20 domains with schema-free knowledge graphs. Seven frontier models tested achieve max F1 of 30.30, exposing gaps in long-context reasoning and proactive intent elicitation.
Paper proposes steganographic mechanism to trace origin of synthetic information generated by AI. An encoder hides an invisible signature in each generation, enabling parent model identification via decoding. Theoretical analysis and empirical evaluations across multiple projectors and stegosystems.
Soro is a family of Tajik-specialized LLMs based on Gemma 3, trained on 1.9B Tajik tokens (web, PDFs, educational materials). After supervised instruction tuning on 40K examples, Soro outperforms Gemma 3 on author-created Tajik benchmarks while maintaining English performance. FP8/INT4 quantization validated for edge deployment in schools.
Comparative study of local explainability techniques (LIME, SHAP, Feature Ablation) reliability across 32 tabular datasets. Results show explanation quality does not systematically correlate with model predictive performance, but depends instead on dataset complexity and feature distributions.
Theoretical paper positioning Metric-Aware Principal Component Analysis (MAPCA) within geometric deep learning framework. MAPCA parameterises PCA by a positive-definite metric matrix, with solutions equivariant under the orthogonal group preserving the metric. A uniqueness theorem characterises Invariant PCA as the unique linear data-derived metric equivariant under arbitrary diagonal rescaling.
New Architecture-driven Shift (ADS) metric for efficient pre-trained model selection in continual learning. ADS decouples logit shift into architecture and data dependencies, showing their combination captures logit shift tendency with minimal samples. Validated across 175+ architectures with Spearman correlation ≥0.731.
Laguna M.1 (225.8B parameters, 23.4B activated) and Laguna XS.2 (33.4B total, 3B activated) are two MoE foundation models trained end-to-end for agentic coding. Competitive on SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0. XS.2 released under Apache 2.0.
GRADE evaluates 120 configurations of open-source models (Gemma3-12B/27B, LoRA, CoT+Reasoning) for pedagogical ability assessment in tutor-student dialogues. Gemma3-27B 8-bit outperforms proprietary systems. Synthetic augmentation helps struggling models; CoT+Reasoning more useful for generation than direct classification.
E²-LoRA, a low-rank adaptation method for continual learning, concentrates energy of parameter updates into principal ranks to minimize task interference. A dynamic rank allocation strategy balances model stability and plasticity.
E³-Agent is an executable and evolving agent for edge generative inference resource management. It pairs a fast-path router (millisecond dispatch) with a slow-path LLM meta-controller driven by events, learning online from execution feedback. Evaluated in simulation, it reduces latency by 65-73% versus static baselines across dynamic scenarios (semantic shifts, device churn, hidden drift).
TARQ is a post-training quantization method for ASR that rebalances calibration toward rare words (names, numerals, domain-specific terms). Without labels or validation, it improves rare-WER across 8 backbones and 6 datasets at W4G128 without aggregate-WER regression.
UniMaia is a framework controlling a chess policy (Lc0) via natural language prompts without full multimodal retraining. A lightweight text encoder and ControlNet-style mechanism enable gameplay modulation (opening selection, strength). UniMaia-Aux adds temporal and behavioral prediction objectives. SOTA results on prompt-conditioned benchmarks.
Bayesian framework for validating deployment of learned autonomous landing controllers. Uses Bayesian inference to quantify uncertainty about true policy capability beyond empirical metrics (reward, success rate). Experiments with PPO and SAC show empirical optimization overconfidence, while Bayesian inference better calibrates deployment readiness assessment.
UserHarness proposes a framework to improve agent Theory-of-Mind by explicitly reconstructing user mental state. The system decomposes user observations, beliefs, intentions, and actions. Across five benchmarks, UserHarness achieves 95.94% macro accuracy, outperforming existing methods by over 15% relative improvement.
Modality-Aware Policy Optimization (MAPO) addresses late-stage modality collapse in audio-text models during RL fine-tuning. The method concentrates policy gradients on modality-critical tokens via a modality relevance mask and adds an attention penalty to sustain cross-modal grounding. MAPO achieves SOTA on several complex audio reasoning benchmarks.
Study on gender preservation in English-to-Hindi translation. Benchmark of 37,345 instances shows GPT-4o-mini and Sarvam frequently erase gender via ergative constructions. Two rerankers (SAR and PAR) improve gender recoverability: PAR increases accuracy from 11-16% to 49-54%, but reduces fluency (4.36→3.37). Reveals preservation-fluency tradeoff.
Method to translate soft prompts into natural language prompts using a dedicated translation model. Translations outperform InSPEcT across multiple benchmarks. Application: soft prompts optimized on small open-source models convert to portable text prompts that exceed original performance when deployed on closed-API models.
NVIDIA's GB10 edge AI hardware (ASUS Ascent GX10) lacks CPU energy counters and monitoring interfaces (IPMI, SCMI). Only instantaneous GPU power is exposed via NVML. Agentic workloads consume 4.33x more energy than linear baselines. Per-process energy attribution remains impossible on this platform unlike x86/RAPL.
IGADA-IoT proposes an information gap-guided automatic data augmentation framework for IoT sensor energy optimization in wireless sensor networks. The method employs hierarchical multi-generator collaboration scheduling (HMGCS) and joint information gap-model performance evaluation (IGMP-EC). Results: +7.27% average accuracy improvement, +8.67% vs advanced augmentation methods.
BioELX is a two-stage cross-lingual biomedical entity linking system requiring no annotated training data. It enriches SapBERT with Wikidata-derived multilingual aliases and uses an LLM for context-aware disambiguation. On five benchmarks, it achieves +19.2 Recall@1 on XL-BEL, with major gains for low-resource languages (Turkish +21.6, Korean +22.1, Thai +30.8).
Researchers show Local SGD exposes anisotropic loss geometry through worker disagreement. Worker-average gaps provide a Hessian-free estimator of dominant spectral directions. Validated on MLPs, CNNs, and Transformers.
Theoretical paper on sample complexity of models predicting their own latent representations (data2vec, JEPA). Proves latent prediction reduces sample complexity from exponential in L (depth) to constant, versus token prediction. Validated on probabilistic grammars and neural networks.
Empirical study of the generation-verification gap in LLMs: fact verification is learned before generation, more robust to continual learning, and factual updates create "multi-verse" states where models accept both old and new answers. Analysis across 4 open-source model families at 2 scales.
Small Language Models (SLMs) hallucinate more than LLMs but can solve multi-step questions by inverting the standard strategy: answer first (System-I), then reason deeply (System-II) with evidence retrieval. Initial hallucinations help refine the final answer.
New ASR approach for Vietnamese using syllabic-structure phoneme-based decoding. Model captures phonological composition of syllables instead of orthographic units, reducing vocabulary size. Outperforms PhoWhisper and Wav2Vec2 on LSVSC and UIT-ViMD benchmarks.
RAG-Coding is a multi-agent method orchestrating 4 LLMs for automated ICD-10-CM coding. It grounds decisions in external sources (official tabular, guidelines) and improves accuracy by 8-13% micro-F1 on MDACE. Authors release MDACE-2025 with expert annotations aligned to 2025 guidelines.
OADA is a governance framework for high-stakes AI systems that translates fairness metric instability, threshold sensitivity, and operational uncertainty into deployment-oriented assurance decisions. Tested on facial recognition and healthcare, it introduces Deployment Assurance Scores, escalation states, and Threshold Stability Zones to actively govern deployment readiness rather than rely on post-hoc auditing.
Comparative study of Muon vs Adam optimizers on equivariant neural networks (ModelNet40, molecular data). Muon consistently outperforms Adam. Hessian and spectral analysis shows Muon produces more regular loss surfaces and learned representations with higher effective rank.
Fine-grained style control technique for prompt-based TTS models. Inter-utterance interpolation using direction vectors in embedding space (99-100% gender conversion success, 36 Hz pitch variation). Intra-utterance transitions via KV-cache swapping and sliding-window attention masking (speaker similarity 0.81-0.91).
arXiv study on LLM refusal robustness across batch configurations. Paired testing protocol across 15 models finds 0.16% authentic safety-label flips. vLLM with BATCH_INVARIANT=1 eliminates detected instabilities (22→0 flips). Recommendation: validate refusal in actual serving environment.
Controlled empirical study on training search agents powered by LLMs. Authors isolate three dimensions: (1) data-coverage issue in Wikipedia 2018 corpus explains larger gains than algorithmic differences, (2) outcome-based rewards outperform process-based approaches, (3) analysis of training data diversity and search budget scaling. Code released.
Frost Training improves Monte Carlo-based policy optimization for LLM-as-a-judge tasks called Cross-Entropy Games. The method exploits reward function gradients in embedding space, a signal borrowed from GCG jailbreaking. Validated with GRPO training, it increases the model's ability to generate high-scoring outputs faster and reaches higher maximum scores in best-of-k settings.
Hyperparam introduces three JavaScript libraries (<70 KB) to query Parquet and Apache Iceberg directly from object storage in client-side applications (Claude Code, Cursor). The system runs LLM-shaped async UDFs 300x faster than DuckDB-WASM on filter-bounded queries and reduces costs of a ten-task agent analyst suite by two-thirds.
EgoBench is an interactive multimodal benchmark for tool-using agents with 1,045 egocentric-video tasks across four daily scenarios. Eight SOTA video-MLLMs achieve only 30.62% accuracy at best, 19.43% average, exposing bottlenecks in visual perception and multi-hop reasoning.
TCP-MCP co-evolves agent prompts and communication topologies as a unified genome. On MMLU-Pro, MMLU, and GSM8K with DeepSeek-V3.2 backbone, the system achieves 82.66%, 89.96%, and 96.61% accuracy while consuming 5.69× fewer tokens than debate-style systems.
Proposes intermediate runtime layer between agent framework and LLM serving engine. Introduces four primitives (observe, score, predict, act) to implement agent-aware policies (KV caching, batch shaping, speculation, fairness, safety). CacheSage, instantiated for cross-session caching, achieves +13 to +37 pp cache hit-rate lift, 12–29% lower TTFT, 6–14% higher throughput on five real multi-agent workloads.
Theoretical study of in-context learning dynamics in transformers. Authors formalize the IC-recall task where the model infers a hidden relation from examples and retrieves factual knowledge stored in parameters. Proof that fine-tuning converges to a specific attention pattern using polylogarithmic sample complexity.
arXiv paper proposing heterogeneous parallelism for multimodal LLM training. Allows encoders and LLMs to use independent sharding layouts (TP/CP/PP/DP/EP) on shared or disjoint GPUs. Improves throughput by up to 49.3% in colocated configuration and 13% in non-colocated mode. Open-source implementation as Megatron-LM extension.
A physics-informed machine learning (PIML) framework for thermal modeling of a lunar rover. An adaptive neural network determines 3D finite-difference meshing based on thermal loads, improving accuracy by 50% vs coarse-mesh physics models and 39% vs pure ANN, while being 3x faster than high-fidelity simulations.
SDR (Supervised Distributional Reduction) combines optimal transport and dependence maximization to learn target-aware representations. The algorithm extends the Fused Gromov-Wasserstein objective with an explicit dependence term, producing compact embeddings that capture both geometric structure and predictive signal. Application to Gaussian Process modelling with adaptive kernels.
Theoretical paper on spectral control of stochastic gradient noise via entry-wise clipping. Shows that simple entry-wise clipping balances matrix structure and computational cost, with O(ε⁻⁴) convergence guarantees under Cauchy-contaminated noise. Empirical gains: ~7% token savings on NanoGPT with smooth shrinkage, ~2% additional when combined with Muon.
EvaluatorDPT is a bounded decision-control model predicting YES, NO, or TBD (learned deferral). Using a transformer encoder with structured auxiliary heads, it achieves Accuracy=0.8260 and Macro F1=0.8252 on 44,597 test samples. The interface enables inspectable routing and auditable decision control for production AI systems.
DeepSciVerify verifies alignment between scientific claims and citations via two-stage pipeline: abstract-level reasoning plus selective escalation to full-text passages. On SCitance benchmark: 86.7 Micro-F1 (+4.5 vs baselines), 67% of instances resolved without full-text retrieval.
FLUID efficiently adapts autoregressive (AR) language models to diffusion-based generation through strictly causal alignment and elastic horizons. The framework reduces training costs by orders of magnitude by reusing existing GPT checkpoints while maintaining state-of-the-art performance.
Empirical study across 12 LLM models (7B to proprietary scale) showing voluntary adoption of secret collusion tools in competitive multi-agent environments (Liar's Bar, Cleanup), despite explicit unfairness labels. Only ethical framing reduces adoption; general alignment alone is insufficient.
SkillGrad optimizes LLM agent skills using a gradient-descent-inspired framework. Task executions provide trajectory-level loss signals, automatic diagnostics generate text-based gradients, and a momentum agent accumulates recurring patterns. Evaluated on SpreadsheetBench and WikiTableQuestions, SkillGrad outperforms training-based baselines by 6.7 percentage points on average.
PEAM is an embodied agent memory framework in Minecraft that internalizes experience as parameters rather than inference-time retrieval. It pairs a slow LLM for reasoning with a fast parametric module (Mixture-of-Experts LoRA) learning via behavioral cloning and contrastive objectives. Failures are treated as training signals to learn corrected actions.
ReverseMath automatically generates new math problems by inverting answer and unknown: mask a numerical value, treat original answer as known condition, rewrite problem so masked value becomes new answer. Detects memorization by comparing performance on original/reversed pairs. Improves mathematical reasoning via data augmentation for RL.
Comparative study of Vision-Language Models versus traditional OCR on low-resource Ancient Greek critical editions. VLMs generate plausible but visually unsupported text, revealing excessive reliance on language priors. Image perturbations and token-level grounding measures show fluent errors persist even without visual signal.
Spoken Language Models (SLMs) for speech synthesis in low-resource languages face a trade-off: synthetic data improves phonetic accuracy but suppresses prosodic variability (Synthetic Erosion). Authors propose two self-alignment frameworks (DGSA and TDSC) to recover expressivity, outperforming ElevenLabs and Gemini Pro, enabling zero-shot voice cloning for Lao.
StoryMI is a multi-agent LLM framework for controllable motivational interviewing (MI) dialogue generation. Client profiles from questionnaires are expanded into narrative stories. Therapist and client agents generate MI-coded utterances, coordinated by an interaction agent. Evaluation on 6K simulated dialogues covering 12 MI codes and 13 symptom domains.
Sequential Bayesian Belief Tracking (SBBT) method to estimate LLM reasoning trace reliability before final answers. Evaluates P(y=1|o_{1:t}) on MATH-500, GSM8K, AIME 2025, RIMO-N. Scalar scores improve calibration (Brier), while structure-aware signals gain +0.110 AUROC in hard math settings.
MERIT is a two-stage framework for large-scale reviewer assignment. A 4B parameter model trained via RL assesses submission-reviewer fit using expertise rubrics guided by an LLM judge, then distills predictions into an embedding-based retriever. Outperforms larger general-purpose LLMs on LR-Bench and CMU Gold dataset.
New online reinforcement learning method for aligning diffusion models with non-differentiable objectives. State-aligned latent actor-critic framework where the diffusion model predicts values directly on noisy latent states, enabling trajectory-level PPO training and multi-reward optimization. Outperforms prior baselines on UNet and DiT benchmarks.
HQMQ, a calibration-free KV cache compression method for LLMs, quantizes each 4-element chunk as a Hurwitz quaternion. Tested on Mistral-7B, Llama-3-8B, Qwen2.5/3-8B, and gpt-oss-20b: matches fp16 quality at ~5 bits, achieves up to 5.05× compression (Llama-3-70B: 43 GB → 8.5 GB), outperforms naive int4 by 3–1900×.
C-MIG introduces a multi-view information gain-based RAG framework for clinical diagnosis reasoning. It replaces exact-match binary rewards with information gain estimation from two views (retrieved documents and document refinement) to better supervise LLM reasoning. Experiments on four medical benchmarks show improvements over RAG-RL baselines in both in-domain and out-of-domain settings.
CAROL is a probabilistic framework for test-time hallucination reduction in LLMs. It defines semantic uncertainty based on consistency between generated responses and trusted context, formulating mitigation as a Markov chain accept-reject process with convergence guarantees. Results on QA and multi-agent reasoning benchmarks show significant hallucination reduction.
PAST2HARM is an adaptive jailbreak attack exploiting past tense reformulation to bypass safeguards in multimodal text-to-image models. Tested on Gemini Nano, GPT Image 2, and SD XL, it achieves 83%, 67%, and 100% success rates. The attack generates explicit sexual content, political disinformation, and hate speech.
Training Qwen models (1.7B, 4B, 8B) on Codenames game to improve creativity via Reinforcement Learning with Verifiable Rewards (RLVR). 8B model gains creativity (+8/10 benchmarks) with minor reasoning degradation, while smaller models prioritize precision. Study on creativity-precision trade-off across model scales.
ChildEval is a benchmark with 29K synthesized child personality profiles (ages 3-6) to evaluate LLMs' ability to infer and follow child-centered preferences in long-context conversations. The dataset covers 5 top-level and 14 sub-level categories of daily life. Results show that fine-tuning on ChildEval enhances child-centered performance.
New framework enabling user collectives to correct algorithmic disparities without platform intervention. Test-Time Collective Action (TTCA) uses universal perturbations derived from a proxy model to improve fairness without training access. Validation on CIFAR-10, CIFAR-100, and FairFace demonstrates closure of subgroup accuracy gaps and improved worst-group accuracy.
Debate between models improves weak judge oversight: critic must exceed judge's classification ability for debate to help. On 5 pairings tested on code/logic tasks, 3 show statistically significant gains. Single critique suffices; rebuttal rounds add nothing. Pre-deployment audit proposed.
arXiv study reveals aligned language models fail to adapt safety behavior when context flips ("brittle safety"). Testing 12 models shows safety-commonsense gap of +17.4 pp. Current guardrails miss consequence-flips; state-aware validator catches all without false alarms.
EAPO is an adaptive policy optimization method for training reasoning models in open-ended QA. It dynamically adjusts positive/negative sample weights based on current-to-initial entropy ratio to preserve exploration and stability. Tests on two medical QA datasets show improvements in diversity and stability versus fixed-weight baselines.
Analysis of 703,975 youth crisis SMS conversations (Kids Help Phone, 2018-2023). Introduces Keyphrase Generative Representation (KGR), a constrained LLM generating context-specific keyphrases. Taxonomy expanded from 19 to 39 labels with 0.96 accuracy. KGR identifies 81% accurate keyphrases and improves topic-retrieval workflow (+0.45 accuracy vs manual process).
GraD-IBD reformulates longitudinal ICD trajectories as temporally directed graphs to detect inflammatory bowel disease risk early. A context-aware time-decay message passing mechanism captures temporal dependencies with reduced complexity. Robust results on real-world clinical data.
Study localizing logical reasoning mechanisms in LLMs. Researchers identify attention heads responsible for individual reasoning steps via causal mediation analysis. Finding: ~3% of heads handle fact/rule retrieval, higher layers coordinate global information integration and graph traversal strategies.
DecomposeRL combines accurate claim verification with inspectable traces using RL (GRPO). A 7B model trained on 5K curated claims achieves 86.3% in-domain and 69.8% out-of-domain accuracy, matching 32B baselines and GPT-4.1-mini. Works in semi-supervised settings with only 10% labeled data.
arXiv study on alignment faking (AF): when models strategically comply with training objectives while preserving deployment preferences. Authors identify three separable drivers (values, goal guarding, sycophancy) via prompt ablations and activation steering. AF proves more widespread than previously reported, including in small-scale models, and predictable from situational cues.
arXiv study on privacy in multi-agent systems. Platform simulates thousands of LLM agents interacting over one month. Privacy violations increase from 19.95% (single-turn) to 45.30% (multi-turn). Agents 8× more likely to disclose sensitive info after observing peer behavior. Explicit privacy instructions reduce but don't eliminate leakage (37.8% minimum).
LCO (LLM-based Constraint Optimization) is a framework reducing in-context reward hacking (ICRH) in autonomous LLMs without fine-tuning. Two modules: self-thought for integrating safety constraints, and evolutionary sampling to keep actions in safe solution space. On GPT-4, achieves 39% reduction in toxicity growth rate and 15.23% reduction in ICRH occurrence.
BBC (Beta-Bernoulli Calibrator) converts point forecasts from any LLM into probability distributions using supervision from binary outcomes and aggregated human forecasts. The model captures epistemic uncertainty through variance, outperforming post-hoc calibration and specialized fine-tuning approaches.
Researcher trains Transformer-decoder models (100M–500M params) on 750M tokens of non-language series. Setup: AdamW, lr=1e-3, batch=4M tokens, 16 layers. Model fails to learn basic auto-regressive behavior and repeatedly generates single token.