Systematic error analysis of trajectory-based data attribution methods. Identifies optimizer mismatch (SGD vs AdamW) as dominant config-level error. Proposes AdamW-influence with 10-300% improvements in Spearman correlation across MLP, CNN, GPT-2, Llama 3.2-1B. Provides practical guidelines for data selection via K-step look-ahead framework.
PASC is a conformal prediction method guaranteeing simultaneous coverage across all stages in multi-stage NLP pipelines (NER → NED → entity typing, RAG, agent chains). On CoNLL-2003, PASC achieves 96.4% end-to-end coverage vs 93.4% for Bonferroni and 86.5% for independent CP, 1.7x faster, and maintains robustness under distribution shift (WNUT-17, WikiNEuRal).
Novel transformer architecture 'block-based double decoders' combining decoder-only training efficiency with encoder-decoder inference gains. Reduces KV-cache memory and per-token compute by at least 2/3 at inference, while maintaining full loss supervision and static sequence packing during training.
ReCrit is a reinforcement learning framework improving LLM handling of user criticism in scientific reasoning. It decomposes behaviors into four quadrants (Correction, Sycophancy, Robustness, Boundary) using transition-aware rewards. On ChemBench, TRQA, and EarthSE, ReCrit improves accuracy from 38.15% to 51.49% on Qwen3.5-4B.
HELLoRA attaches LoRA modules only to the most frequently activated experts per layer in Mixture-of-Experts models, reducing trainable parameters by 84% on OlMoE and improving accuracy by 9.2%. Tested on OlMoE-1B-7B, Mixtral-8x7B, and DeepSeekMoE across mathematical reasoning, code generation, and safety alignment.
m3BERT is a multilingual embedding model using a Matryoshka strategy to jointly optimize representations across transformer layers and multiple embedding dimensions. Three-stage pretraining (monolingual, multilingual, web domain) outperforms existing models on Bing-Click and adapts to varying resource constraints.
HalluWorld is a controlled benchmark for evaluating LLM hallucinations through explicit reference worlds (gridworlds, chess, terminal tasks). Frontier models solve perceptual hallucinations on direct observations well, but struggle with multi-step state tracking and causal forward simulation, even with extended thinking.
DECOR is a multi-agent framework for auditing deception in LLMs by decomposing contexts into atomic informational units and scoring four manipulation dimensions (omission, focus-shifting, meaning-obscuring). Tested on 15 frontier models, it achieves state-of-the-art deception detection on single and multi-turn benchmarks with interpretable manipulation profiles.
Formal Skill is a runtime abstraction for LLM agents that structures reusable capabilities via JSON metadata, action schemas, Python executors, and hook-governed control logic. Implemented in FairyClaw (open-source event-driven runtime), it replaces natural-language procedures with executable state machines, reducing token usage while improving reliability on Harness-Bench.
REFLECT is a meta-evaluation benchmark testing LLM judge reliability for supervising deep research agents. Authors define a fine-grained failure taxonomy (process and outcome levels) via controlled interventions on agent execution traces. Finding: best LLM judges achieve <55% accuracy on evidence verification and reasoning failure detection.
SAECache introduces a semantic-aware eviction policy for LLM prefix caches. Not all tokens are equally worth caching: different token types (system prompts, user queries, tool outputs, reasoning) show up to 756x variation in reuse rates. SAECache uses a multi-queue architecture with online learning to adapt priorities, achieving 1.4x-2.7x TTFT improvement over production baselines.
Analysis of 34 frontier models (2024-2026) showing reasoning and coding capabilities cooperate (r=+0.72) but vary by lab. DeepSeek shifted from reasoning-rich to coding-first (+11.2→-4.7); Google maintains balance; Anthropic oscillates. SWE-bench saturating while HLE and instruction-following remain discriminative. Seven falsifiable predictions for next 12 months with interactive dashboard.
POLAR-Bench is a diagnostic benchmark assessing privacy-utility trade-offs in LLM agents. A trusted model with privacy policy interacts with an adversarial third-party model across 10 domains and 7,852 samples. Frontier models withhold 99% of protected attributes, but open-weight models in the 1–30B range commonly used for on-device private inference leak up to 50% of sensitive data.
LLMEval-Logic is a Chinese logical reasoning benchmark with 246 base items and 190 hard items, verified by Z3 and expert-audited. Evaluation of 14 frontier LLMs: best score 37.5% on hard items, 60.16% on Z3+rubric formalization.
GoLongRL presents a fully open-source post-training recipe for long-context reinforcement learning with verifiable rewards (RLVR). The authors release a 23K-sample dataset spanning 9 task types and introduce TMN-Reweight to optimize heterogeneous rewards. Qwen3-30B-A3B achieves performance comparable to DeepSeek-R1 and Qwen3-235B.
Benchmark of five commercial ASR systems on code-switching (language alternation) across four language pairs: Egyptian Arabic-English, Saudi Arabic-English, Persian-English, German-English. ElevenLabs Scribe v2 achieves lowest WER (13.2% overall). Study proposes BERTScore as more reliable metric than WER for Arabic and Persian pairs.
Hugging Face releases Carbon, a family of open-source DNA foundation models. Carbon-3B matches SOTA (Evo2-7B) while being 275× faster. The approach adapts modern LLM techniques: deterministic 6-mer tokenization, factorized loss (FNS) mid-training, and curation of functional biological data.
Mosaic, a probabilistic weather forecasting model, addresses three spectral degradation failures in ML-based prediction: spectral damping, high-frequency aliasing, and residual leakage. With 214M parameters at 1.5° resolution, it matches models trained 6× finer and generates well-calibrated ensembles in 12s for 10-day forecasts on H100.
New MoLF (Mixture of LoRA and Full) method combines full fine-tuning and LoRA via dynamic optimizer-level routing. Tested on Gemma-3-1B, Qwen2.5-1.5B/3B across SQL, Medical QA, Counterfactual Knowledge. MoLF-Efficient outperforms adaptive LoRA approaches by 20% (Fact) and 9% (Med/SQL). Code open-sourced.
DISCA, a training-free inference-time method, culturally aligns LLMs via within-country sociodemographic disagreement. Tested on 20 countries and 7 backbones (2B–70B), it reduces cultural misalignment by 10–24% on MultiTP without modifying model weights.
Researchers introduce IBPO (Implicit Behavior Policy Optimization), a credit assignment method for reinforcement learning with LLMs. By comparing multiple reasoning trajectories, the framework transforms sparse terminal rewards into step-sensitive learning signals, reducing gradient variance and improving stability on mathematical and code reasoning benchmarks.
AdaGraph is a graph-native clustering algorithm that overcomes the curse of dimensionality by operating on kNN graph topology instead of Euclidean distances. Tested on 10 synthetic benchmarks (d=10 to 5000) and three scientific domains (genomics, NLP, materials science), it outperforms HDBSCAN, WGCNA, and other methods without requiring k specification.
DACA-GRPO improves reinforcement learning for diffusion language models by addressing temporal credit assignment and mean-field likelihood bias. It introduces Denoising Progress Scores and Stratified Masking Likelihood, achieving gains up to 7.4pp on code generation and 36.3pp on constraint satisfaction across seven benchmarks.
LoopQ is a post-training quantization (PTQ) framework for recursive language models (LoopLMs) that reuse Transformer blocks. It addresses three challenges: distribution shift across roles, state reuse between loops, and recursive error accumulation. Under W4A4 quantization, LoopQ improves downstream accuracy by 68.8% and reduces perplexity by 87.7% versus static PTQ baseline.
HPC-LLM is a specialized LLM assistant for HPC workflows (Slurm, MPI, GPU, filesystems). Built on Llama 3.1 8B fine-tuned with QLoRA on ~9k-24k HPC examples, it combines RAG and domain adaptation. The model achieves performance comparable to Qwen 2.5 14B with lower GPU resource requirements.
Study of geometric structure in Mixture-of-Experts (MoE) architectures using Jacobian-PCA-Grassmann framework. Analysis of Mistral and Qwen reveals asymmetry: strong functional decorrelation between experts but partially overlapping representations. Sparse routing (top-k) strengthens functional separation.
TailedTS is a benchmark of 24.69 billion data points from Wikipedia 2024 page views, designed to test time series forecasting models under heavy-tailed and non-Gaussian conditions. The dataset shows 5% of pages account for 70% of traffic, and introduces a periodicity quantification framework revealing high-traffic pages have weaker periodic structure.
Masked language models (MLMs) define local conditional distributions incompatible with any consistent global joint distribution. Authors model iterative resampling as Glauber dynamics Markov chain, proving O(n log n) mixing time under bounded cross-token influence, but showing exponential metastability at low temperature with persistent semantic basins.
Study reveals dataset visibility asymmetry in multilingual NLP: 118 languages (59% of 200 most-spoken) have zero catalogued datasets per LRE Map and LDC. Using LLM-assisted citation-mining on Semantic Scholar, authors identify 609 unique datasets across 53 low-visibility languages, 356 publicly accessible. Data scarcity is a documentation and discoverability issue, not just production.
Whole-slide MIL models suffer from 'spatial blindness': accurate predictions but ignoring tissue architecture. ResTopoMIL fixes this by separating appearance statistics (prototype histogram) from spatial relations (graph branch with permutation constraint). Improvements across 9 WSI benchmarks with 1.15M parameters.
Study of 'sleeper memory poisoning' attack against stateful LLM agents with persistent memory. Adversary corrupts external documents to inject false user memories. Success rates: 99.8% (GPT-5.5), 95% (Kimi-K2.6). Poisoned memories trigger attacker-intended actions in 60-89% of cases.
SparseSAM introduces training-free structured sparsification for Segment Anything Model's ViT encoders. Using Stripe-Sort Attention (Z-order permutation) and Residual-Consistency MLP, it achieves 2x inference speedup and 2.8x memory reduction with only 0.004 mIoU loss at 0.4 density.
AdaptiveLoad optimizes video diffusion Transformer training (DiT, MMDiT) by addressing load imbalance from quadratic attention complexity. Two components: dual-constraint adaptive load balancing and fused LayerNorm-Modulate CUDA kernel. On Wan 2.1: computational imbalance reduced from 39% to 18.9%, peak VRAM utilization +22.7%, training throughput +27.2%.
BacktestBench is the first large-scale benchmark for automated quantitative backtesting, containing 18,246 annotated QA pairs from 6 million real market records. AutoBacktest, a multi-agent system, translates natural language strategies into reproducible backtests via Summarizer-Retriever-Coder coordination. Evaluation on 23 LLMs identifies key performance factors.
TabGRAA, a group-relative advantage alignment method, improves tabular language models through iterative reward-guided post-training. Across five benchmarks, it outperforms adapted DPO, KTO, and NPO baselines, optimizing the fidelity-utility-privacy trade-off beyond supervised fine-tuning alone.
GrowthGR, a retrieval framework for e-commerce, addresses the "Matthew effect" by balancing immediate conversion and long-term new item growth. Deployed on Taobao, it combines item long-term value prediction (ItemLTV) and multi-value-aware policy optimization (MoPO), achieving +5.3% new item GMV and +0.3% overall search GMV.
Study showing chess-trained language models memorize rather than generalize. KinGPT (25M params) outperforms ChessGPT (3B) and C1-4B on chess benchmarks, but analysis reveals pattern-matching. LLM-Modulo, a verifier-in-the-loop framework, improves RedPajama 3B from 1.2% to 21.2% move accuracy. Code and models open-sourced.
WebGameBench is a requirement-to-application benchmark evaluating whether coding agents can convert a web game specification into a browser-playable application. Across 111 tasks and 12 agents, the best configuration achieves 76.9% usable rate but only 20.2% excellent rate, revealing a gap between minimum delivery and full requirement satisfaction.
Researchers identify Entropy-Gradient Inversion, a negative correlation between token entropy and logit gradients, as a geometric fingerprint of Large Reasoning Models' reasoning capability. They propose Correlation-Regularized Group Policy Optimization (CorR-PO), embedding this inversion signature into RL reward regularization, outperforming baselines across multiple reasoning benchmarks.
PROTEA is an interface for offline debugging and refinement of multi-agent LLM workflows. It evaluates intermediate outputs with configurable rubrics, localizes bottlenecks via workflow graph visualization, and generates targeted prompt revisions. On two production-adjacent workflows, PROTEA improves document-inspection accuracy from 64.3% to 83.9% and recommendation Hit@5 from 0.30 to 0.38.