BiasGRPO introduces Group Relative Policy Optimization (GRPO) to mitigate social bias in LLMs. By normalizing rewards across sampled completions, the method stabilizes training compared to DPO and PPO. Authors release a compute-efficient bias reward model and extended dataset for multi-objective RLHF.
Diagnostic study on offline model selectors for edX dropout prediction. Authors identify three failure causes (mismatched learner, non-predictive state, label shift) via three stages: oracle ceiling via k-NN, BC/DQN/CQL evaluation, state ablation. Across 5 models, oracle gains 9.7 accuracy points, but learners plateau due to local representational ambiguity.
Metric-aware hybrid approach for CTF4Science Lorenz challenge mixing short-horizon forecasting, long-time distribution matching, and trajectory reconstruction. Multi-model system: synthetic-pretrained denoisers, Lorenz ODE fitting for first 20 steps, histogram-tail substitution with synthetic libraries. Public leaderboard score 83.84.
AXON is a training-free module improving parallel decoding for discrete diffusion language models. It selects anchor tokens (confident masked tokens) that uncertain positions attend to via attention signals, reducing denoising steps while maintaining quality on reasoning and code-generation benchmarks.
ADAPTOOD is a framework for adapting ECG time series models to distribution shifts. It uses data uncertainty to quantify shift severity and guides fine-tuning with low-rank updates and adaptive hyperparameter optimization. Results: +7% accuracy and +12.9% precision on OOD tasks.
Boolean Task Algebra (BTA) enables zero-shot task composition in RL. Authors prove that in deterministic MDPs, optimal Q-value functions collapse to universal and empty tasks, making logarithmic base tasks redundant. They propose a goal-set-based composition method reducing learning and composition costs while maintaining policy performance across tabular, visual, and continuous-control domains.
SANE introduces a schema-aware text-to-SQL evaluation paradigm for domain-specific biological data, using automatically generated benchmarks tied to real experimental structures. Few-shot LLMs achieve accurate query generation without fine-tuning when combined with schema-aware prompting; failures stem mainly from ambiguous inputs rather than SQL generation errors.
Authors propose a framework for multi-agent systems where disagreement between agents is treated as an informative signal rather than error to eliminate. They define four disagreement states based on reasoning-trace similarity and conclusion agreement, applied to content moderation with defeasible strategic routing.
New FFR approach retrieves coherent multi-utterance, multi-image fragments from long-form multimodal dialogues. Two models: F2RVLM (generation + RL with multi-objective rewards) for single-dialogue, FFRS (two-stage indexing + retrieval) for corpus-scale. MLDR dataset introduced, superior performance on benchmarks.
XGBoost model for three-class Alzheimer's detection (normal cognition, mild cognitive impairment, Alzheimer's) on 1,641 ADNI subjects. Macro AUC-ROC 0.983 in cross-validation, 0.982 on test set. SHAP analysis identifies CDR Global as dominant predictor for NC/MCI, CDR-SB and MMSE for AD.
ACAT is a web-based collaborative platform for Aspect-Based Sentiment Analysis (ABSA) dataset annotation. It supports 4 ABSA workflows, automates ETL and computes Inter-Annotator Agreement metrics. On 1,002 restaurant reviews, median annotation time 31.58s, IAA 0.78–0.86.
Geometric study of counterfactual explanations in modern ML systems. Researchers show that models with similar predictive performance can differ substantially in the feasibility and distance of counterfactual changes, determined by decision-boundary proximity and local data support.
CausalBERT-based methodology to isolate the impact of each aspect on overall online review ratings. Enhancements: temperature scaling, hyperparameter optimization, interpretability methods. Validation on 600K+ reviews of U.S. schools: school administration and benchmark performance are key drivers.
Foundation models based on physical principles (Fourier decomposition, energy conservation, symmetry) rather than statistical correlations. Trained on RF data, a 1.99M parameter frozen encoder achieves 77.7% average accuracy across 15 diverse tasks via linear probing without fine-tuning. Superior performance on physically-grounded tasks (84.5%) versus semantic tasks (70.0%).
SGR integrates LLMs with external knowledge graphs via schema-guided subgraph generation. Framework extracts entities and relations, retrieves compact subgraphs, then combines direct Cypher reasoning and multi-path validation. Experiments on CWQ, WebQSP, GrailQA, KQA Pro show improved accuracy and Hits@1 over baselines.
Theoretical paper defining Bayes-sufficient representations in supervised learning. A representation is Bayes-sufficient if a prediction head can use it to implement a Bayes-optimal action rule. The framework connects sufficiency to property elicitation and shows that loss determines the relevant information required.
Longitudinal study showing that daily 5-minute conversations with AI over 28 days decreased preference for human support by 10.3% and increased AI preference by 11.6%. Emotional support emerges incidentally within routine interactions, not just dedicated chatbots, progressively redirecting relational choices.
MechSim is a neuro-symbolic reasoning framework enabling LLM agents to reason about mechanisms, assumptions, and execution behavior of scientific simulators. It represents simulators through structured schemas capturing assumptions, variables, and mechanism dependencies, generating evidence-grounded explanations linking simulator outcomes to underlying mechanisms rather than treating them as black boxes.
The Digital Apprentice is a framework for agentic AI that calibrates autonomy progressively based on empirical evidence. The agent internalizes a supervising human's tacit methodology, escalates autonomy tiers per-skill, and corrects alignment drift at runtime. Three pillars: methodology capture, authorization with explicit escalation, continuous alignment.
MONIR is a Modalized-Output Normative Intermediate Representation for ASP-based compliance reasoning. The framework features staged operational semantics, executable compilation (MONIR-ASP), and extensions for external functions and temporal rules. Validated on Chinese ADAS regulations using an LLM-assisted pipeline.
Study on affinity-based reinforcement learning to instill virtuous behavior in AI agents. Researchers test this technique in Fog of Love, a complex multi-agent environment where two agents must balance individual competition and relational cooperation. Localized affinities improve performance and make agent behavior interpretable.
Tree-based formalization of complementarity in multi-agent human-AI interactions. Framework models HAI protocols as rooted planar binary trees with prediction vectors at leaves. Key results: complementarity unattainable via selector-based HAIs, achievable in regression (closed form for N=2), obstructed in classification under endpoint-monotone losses.
Mixed-methods study on AI integration into mathematical proof formalization workflows. Qualitative survey plus controlled user study show mathematicians prefer high-level control but achieve better formalization accuracy with AI tool access, flexibly combining multiple models.
Survey paper positioning LLMs within the longer history of computational approaches to concept analysis in history, philosophy, and sociology of science. Examines what LLMs add to prior methods (early digital methods, distributional approaches, lexical semantic change detection) and persistent methodological challenges: corpus construction, operationalization, evaluation and interpretation.
Tabular reinforcement learning for metro network expansion (MNEP). Reformulated as Non-Markovian Rewards Decision Process (NMRDP): matches Deep RL performance with 18× fewer training episodes and 12× lower carbon emissions. Incorporates social equity criteria. Validated on Xi'an and Amsterdam.
DMAIC-IAD is a multi-agent system inspired by the DMAIC quality-management framework for industrial anomaly detection. It plans strategies first via standardized operating procedures, then uses a pre-trained judge model to evaluate candidates without costly execution. 37.76% improvement over existing agentic baselines.
KODA is a kernel-based framework for comparing and aligning vision-language model representations (CLIP, SigLIP). The method identifies sample subsets weakly clustered under one representation but strongly clustered under another through constrained optimization and low-rank approximations. Code released.
Theoretical framework for deep reinforcement learning in continuous environments modeled as continuous-time stochastic processes. For single-hidden-layer networks, authors characterize state distribution evolution via stochastic differential equations in the infinite width limit.
User converted an Android phone (Z Fold 6) into a local LLM inference node using Vulkan, GGUF, and llama.cpp. The device exposes an OpenAI-compatible endpoint integrated into a Tailscale mesh with LiteLLM routing and fallback to larger nodes (Mac Studio, RTX box).
Local test of Gemma 4 12B vs 26B-A4B on RTX 4090: HTML5 canvas animation with physics. 26B uses 15 GB VRAM (6.9k tokens, 138 tok/s), 12B uses 9 GB (8.9k tokens, 80 tok/s). 26B-A4B outperforms 12B (~1.7x faster with 4B active params), but 12B remains competitive on 16 GB.
AMD RX7900XTX tested with ROCm7 on llama.cpp (Qwen 3.6-35B). Quiet mode (171W) vs standard mode (271W): 99W reduction, 9.4% generation drop, 17.1% long-context prefill drop, junction temp -11°C.
Gemma 4 12B tested as coding agent on RTX 4080 Super via VSCodium + Pi Agent extension. Model generated complete Python script (log parsing, error extraction, JSON export), created test data, executed code in terminal, and validated output with zero bugs. 32K context, Q4_K_XL quantization, llama.cpp + CUDA.
llama.cpp optimizes Qwen 3.6/3.5-MTP support following multiple PRs and fixes. Community shares tokens/s benchmarks with full configurations (quantization, context, KVCache, MTP). Example: 207.90 t/s prompt eval, 24.07 t/s generation with 52.6% draft acceptance rate.
Developer built a compiler converting Python code into a model-optimized representation. Tested on ~13k files, achieving 14% token reduction with 99.8% AST-equivalent roundtrip success. Open-source project called Vulpine.
Portable C++ implementation of Meta's EnCodec using Eigen. Lightweight audio codec with no runtime dependencies, easy CMake integration, performance comparable to or exceeding onnxruntime. Weights compiled into binary.
Uber caps monthly token spending at $1,500 per AI coding tool (Claude Code, Cursor) to control costs. The limit represents ~11% of median Uber engineer salary ($330k). Policy follows exhausting 2026 AI budget in four months.
Spatial reasoning benchmark on LLMs using Sokoban under zero-shot conditions. ChatGPT, Qwen3.7-max, and Gemini 3.5-thinking pass; Gemini 3.5-flash, Qwen 3.6/3.7-plus, GLM-5, and Gemma4 fail. Strict formatting (UP/DOWN/LEFT/RIGHT only) prevents chain-of-thought cheating.
llama.cpp build b9455 with tensor-split achieves 70+ tokens/s on Qwen3.6-27B-UD-Q8_K_XL with 2x3090, matching vllm performance. MTP speculative decoding and flash-attention enabled. Context up to 262K tokens, prefill at 1400+ t/s.
Comparative study of encoder-only Transformer vs LSTM for streamflow prediction in ungauged basins. On NOAA NWM data, LSTM outperforms Transformer. Incorporating downstream information improves all models (+60% median NNSE). Recurrent memory proves better suited for this hydrologic inference task.
MASC (Margin Self-Correction) is a language-model unlearning method that efficiently reduces the logit gap between the original next token and its alternatives, without requiring downstream evaluation. Tested on TOFU, MUSE News, and MUSE Books, it achieves competitive forget-retain trade-offs at a fraction of existing baselines' computational cost.