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.
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.
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.
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.
Comparative study of RL vs SFT on Qwen2.5-3B-Instruct: reinforcement learning better preserves internal circuits of the base model than supervised fine-tuning (SFT), which adapts faster but destroys more prior capabilities. Proposed metric: differential circuit vulnerability at attention head level.
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.
ReasonOps is an unsupervised method for analyzing chain-of-thought traces from LLMs. It identifies 7 recurring reasoning operators (backtracking, inferring, hypothesizing) from 44,662 traces across 12 models on 8 benchmarks. These operators enable source model identification at 70-76% accuracy and predict answer correctness before trace completion.
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.
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.
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.
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.
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.
Anthropic releases Claude Opus 4.8, outperforming GPT-5.5 and Gemini 3.1 Pro on most benchmarks. The model catches its own coding errors 4× better than its predecessor. Anthropic also rolls out dynamic workflows enabling hundreds of parallel sub-agents for codebase-wide migrations.
Wall-OSS-0.5 is a 4B VLA from X Square Robot with open training code. Zero-shot evaluation on 17 real-robot tasks: 4 tasks >80% progress, including Rope Tightening (82%). Post fine-tuning: 60.5% average task progress (+17.5pp vs pi0.5). Mixture-of-Transformers architecture with vision-aligned RVQ tokenizer and distributed DMuon optimizer.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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).
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.
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.
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.
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.
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.
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.
GenSBI is an open-source JAX library for simulation-based inference (SBI). It implements flow matching, score matching, and denoising diffusion with three transformer architectures (SimFormer, Flux1, Flux1Joint). Validation on SBIBM benchmarks: C2ST scores of 0.50-0.56 (ideal=0.50).
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.