Researchers present a systematic empirical study of inference-time scaling across four dimensions (contextual, temporal, structural, parallel) for locally-deployed computer-use agents under hardware constraints. Evaluating Qwen3-VL-8B/30B-A3B, UI-TARS-1.5-7B, and OpenCUA-7B on OSWorld, they find that additional compute often shifts rather than eliminates failure modes—contextual scaling saturates, temporal scaling extends erroneous trajectories, and structural decomposition adds overhead. The findings argue for selective compute allocation and failure-aware control mechanisms tailored to local model capabilities.
A new arXiv paper investigates how language models behave when given alternative (non-canonical) tokenizations of the same input string across 27 languages and six downstream tasks. While prior work showed English models are largely invariant to such perturbations, the study finds this does not generalize: Llama-3.1-8B drops 23.7% on average, Qwen3-8B 11.4%, and Gemma-3-12B 9.9% in relative performance. Languages with higher token fragmentation are systematically more sensitive, and the authors show LoRA fine-tuning on multi-tokenization data—including English-only data—provides meaningful mitigation.
Researchers introduce a paired-prompt benchmark testing whether language models can correctly match diagnostic evidence to causal claims that vary by population, estimand, or identifying assumptions — a task where surface-level cues can mislead. Using linear probes on final-token hidden states from Qwen2.5-7B, Qwen3-8B, and Llama-3.1-8B, they find balanced accuracy of 0.654–0.659 on a 49-pair benchmark spanning nine diagnostic families, exceeding permutation nulls and text-only baselines. The key finding is that hidden states contain linearly decodable information about causal relevance that is not fully captured by output logits or surface features.
Researchers propose MemSFT, a method that decouples domain specialization from backbone parameter updates by training a plug-and-play parametric memory module to imitate a non-parametric retriever over domain data. A learned router dynamically fuses the memory and backbone output distributions at each decoding step, allowing selective invocation of domain expertise. Evaluated across biology, geoscience, and law on models from Qwen3-8B to Qwen3-235B-A22B, MemSFT consistently improves domain performance with negligible general-task degradation, whereas full SFT causes severe catastrophic forgetting. The memory module is reusable across LLM sizes, offering a practical path to modular domain specialization.
Researchers present Nanbeige4.2-3B, a 3-billion non-embedding parameter model designed for agentic tasks including code-agent, office-agent, and complex tool use. The model is pretrained on 28T tokens using a Looped Transformer architecture that reuses layer stacks to increase capacity without adding parameters, and trained with a multi-stage RL pipeline combining mixed-mode RLHF, length-controlled reasoning RL, and agentic RL with outcome and process rewards. Evaluations claim Nanbeige4.2-3B outperforms larger models including Qwen3.5-9B and Gemma4-12B on diverse agentic benchmarks while remaining competitive on reasoning tasks. The result contributes to the ongoing question of how much agentic capability can be packed into compact, locally-deployable models.
Researchers propose Visual Contrastive Self-Distillation (VCSD), a training method for vision-language models that creates an on-policy self-distillation signal purely through input conditioning — specifically by contrasting token distributions from an EMA teacher conditioned on the original image versus a content-erased control. The approach eliminates the need for external teachers, privileged answers, visual evidence signals, or reasoning traces. Evaluated on Qwen3-VL and Qwen3.5 models using the ViRL39K dataset, VCSD consistently outperforms matched on-policy self-distillation baselines, with aggregate benchmark gains of up to ~3.75 percentage points at the 8B scale.
Researchers introduce SelectBench, a benchmark and training set for evaluating whether retrieval-augmented LLMs can selectively adopt valid evidence while rejecting misleading or injected content. They post-train Qwen3.5-4B using DAPO with rule-based and semantic judge rewards, achieving modest but directional improvements on SelectBench-v2 (22.46% to 26.46% strict success). Gains do not survive Holm multiple-comparison correction, and prompt-injection resistance shows no improvement, leaving statistical robustness and injection resistance as open challenges. General capabilities on MMLU and HotpotQA are preserved.
Researchers evaluate multiple training paradigms for legal machine translation, comparing supervised fine-tuning and reinforcement learning with verifiable rewards (RLVR) on small models (Qwen3.5 4B/9B, Gemma 3 12B) against frontier reasoning models. Using the Swiss multilingual legal system as a testbed, they find RLVR outperforms SFT for legal NMT and brings small models close to frontier performance, though a gap remains. The study also observes diminishing returns from re-training as model size increases. Code and models are publicly released.
Researchers use learned soft prefixes (opaque continuous vectors) to probe the logical stability of LLMs on syllogistic reasoning benchmarks without modifying model weights. Across Qwen3.6-35B-A3B MoE, Qwen3-8B, and Gemma 4 31B, successful prefixes redirect correct answers at flip rates of 72–90% for Qwen3.6 MoE and 54–56% for Gemma, far exceeding random controls by 37–99 percentage points. The dominant effect is a broad answer-preference bias rather than symbol-level forcing, and the bias generalizes across unseen logical forms and prompt interfaces. Model-specific differences in how this bias manifests suggest substantial variation in logical robustness across architectures.
A new arXiv paper provides the first direct comparison of model merging versus joint multi-task reinforcement learning training, using Qwen3-8B specialists trained on the AppWorld agent benchmark with the LOOP algorithm. Merging methods (TIES, RAM+) statistically match jointly trained models on task-goal completion. The authors explain this via task vector geometry: specialist task vectors are near-orthogonal (cosine similarity 0.06–0.10) despite ~65% parameter support overlap, causing sign- and support-based merging methods to collapse to near-uniform averaging.
Researchers propose a self-supervised mid-training objective called function-aware fill-in-the-middle (FIM) that exploits the structural isomorphism between a coding agent's action-observation-continuation loop and function call sites in ordinary code. Applied to Qwen2.5-Coder-Instruct (7B/14B) and Qwen3-8B on a 2.6B-token GitHub corpus, the method yields +2.8 to +5.4 point gains on SWE-Bench-Verified and SWE-Bench-Lite across multiple post-training pipelines. Notably, the technique also mitigates capability erosion on non-agent coding and tool-use benchmarks, suggesting the function-call inductive bias generalizes beyond the training domain.
Researchers introduce three contributions to address gaps in multilingual moral reasoning for LLMs: MCLASH, a culturally situated multilingual moral decision-making benchmark; MET, a two-step prompting method grounded in psychological and philosophical moral theory; and MET-D, a self-distillation training variant requiring no external supervision. MET-D improves macro-F1 by an average of 3.71 points on MCLASH and 4.23 on MMoralExceptQA across Qwen3-4B, Qwen3-8B, and Gemma3-4B, with a peak gain of 12.94 points for Malay on Qwen3-8B. The work also finds that native-language reasoning increases by 62 points on average and that beneficial moral grounds differ systematically across cultures.
A new arXiv paper examines whether test-time scaling (TTS) transfers to small open vision-language models using EXAMS-V, a multilingual visual multiple-choice benchmark. The study compares self-consistency, describe-then-reason with PRM-guided beam search, and post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. Key findings: prompt parseability and decoding budget (token limit) dominate gains, while elaborate search/verification methods like PRM-guided beam search underperform plain majority vote at 8x the cost. The best configuration achieves 84.1% on ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.
Researchers from Meta and several French and Spanish institutions released Brain2Qwerty v2, a non-invasive brain-computer interface system that decodes magnetoencephalography (MEG) signals into text using a CNN/conformer encoder, a word-aligner, and a fine-tuned Qwen3-4B language model with per-subject LoRA adapters. The system achieves a 39% word error rate on 9 subjects, down from 43% in v1, trained on 90 hours of MEG recordings. A notable finding is that cross-subject training substantially outperforms single-subject training, suggesting a data-scaling dynamic analogous to LLM pretraining. Training code and v1 data have been open-sourced.
DominoTree is a new training-free speculative decoding method that constructs best-first draft trees scored by Domino's conditional, non-factorized GRU-based correction along each root-to-node path. On Qwen3-4B across eight benchmarks, it achieves up to 6.6x speedup over autoregressive decoding and a mean accept length of up to 10.7 tokens per round, outperforming prior methods including DDTree, CaDDTree, DFlash, and the base Domino decoder. A GPU-native CUDA-graph tree builder provides 9-10% throughput gains over Domino overall, with up to +22% on Alpaca, while maintaining bit-identical acceptance behavior.
Researchers at Stanford and UC Berkeley developed RoboReward, a family of 4B and 8B vision-language reward models designed to provide reward signals for robot reinforcement learning across diverse robot types and tasks. The team built a novel dataset by augmenting successful robot demonstrations with synthetically generated failure examples using GPT-5 mini and Qwen3-4B, then fine-tuned Qwen3-VL models to predict task progress scores. RoboReward 8B outperformed GPT-5, GPT-5 mini, and Gemini Robotics-ER 1.5 on the new RoboRewardBench evaluation, and in real-world robot trials substantially exceeded prior reward model baselines while still falling short of human-assigned rewards. The authors also release RoboRewardBench as a community benchmark for reward model evaluation.
Researchers introduce RECONTEXT, a training-free inference-time method for improving long-context reasoning in LLMs. The approach uses model-internal relevance signals to build a query-conditioned evidence pool that is replayed before final generation, without modifying the original context, external memory, or context pruning. Experiments across eight long-context datasets at 128K context length show consistent improvements on Qwen3-4B, Qwen3-8B, and Llama3-8B. The authors provide a theoretical grounding via associative memory theory, framing attention as cue-trace association and replay as trace reactivation.
Researchers propose NLL-guided layer selection, a training-free technique for hybrid attention models that identifies which layers should use full versus sliding-window attention by measuring negative log-likelihood degradation on answer tokens. On LongMemEval with Qwen3-4B, the method achieves 64.6% accuracy using only 1/4 full-attention layers, matching a 1/2-FA periodic baseline while halving compute, and outperforming a periodic 1/4-FA baseline by 10.4 percentage points. The calibration procedure requires approximately 15 minutes of one-time compute, making it practical for deployment. The work advances the efficiency-accuracy tradeoff for long-context LLM inference without requiring any retraining.
DeepSeek published eagle3_qwen3_4b_ttt7 on Hugging Face, a draft model for EAGLE3 speculative decoding targeting the Qwen3-4B base model. EAGLE3 is DeepSeek's third-generation speculative decoding framework designed to accelerate inference by predicting future tokens with a lightweight draft model. The release is a narrow inference-optimization artifact with zero downloads and likes at time of indexing, suggesting it is very fresh or experimental.
DeepSeek published eagle3_qwen3_8b_ttt7 on Hugging Face, a draft model for EAGLE3 speculative decoding targeting the Qwen3-8B base model. EAGLE3 is DeepSeek's third-generation speculative decoding framework designed to accelerate inference by predicting future tokens with a lightweight draft head. The release is a narrow inference optimization artifact with minimal engagement at time of indexing.
Researchers introduce RiVER (Ranking-induced VERifiable framework), a reinforcement learning approach that trains LLMs on score-based optimization tasks using deterministic execution feedback as continuous rewards, without requiring ground-truth answers. The method addresses two failure modes in group-relative RL with continuous rewards—scale dominance and frequency dominance—via calibrated, instance-wise reward shaping. Applied to Qwen3-8B and GLM-Z1-9B-0414 on competitive programming tasks, RiVER improves ALE rating rank by ~9% and also transfers to exact-solution benchmarks (LiveCodeBench, USACO) with 2-4% absolute gains, unlike raw-score baselines. The result suggests score-based heuristic tasks can serve as general-purpose RL training environments for coding ability.
OPERA (Objective Perplexity-based Reflective Alignment) proposes replacing LLM-as-a-judge reward models with intrinsic rewards derived from perplexity dynamics to stabilize RL training on open-ended tasks like creative writing. The method includes a cold-start data synthesis pipeline generating 20,000 reasoning trajectories using perplexity-prioritized rollouts. Applied to Qwen3-8B, OPERA claims state-of-the-art among open-source models on open-ended tasks, reportedly matching or exceeding Gemini 2.5 and MiniMax-M2.5 on some benchmarks.
Researchers introduce STAGE (Spreadsheet-grounded Text-to-JSON Artifact GEneration), a data generation pipeline that uses LLMs to synthesize training data for structured extraction from long unstructured documents, validating outputs against underlying spreadsheets. Evaluated on STAGE-Eval, an 851-example benchmark, the pipeline substantially improves Qwen3-4B performance, raising exact match from 31.37% to 74.27% and value accuracy from 45.46% to 90.69%. The work targets a practical bottleneck in enterprise document processing: reliably converting financial filings and clinical records into machine-readable JSON.
Researchers introduce G-IdiomAlign, a benchmark anchoring idioms via English glosses from Wiktionary to evaluate cross-lingual idiom equivalence in LLMs. The benchmark supports two evaluation protocols: a multiple-choice task with typed distractors and a gloss-contrastive generation task isolating the effect of explicit semantic pivots. Experiments across diverse LLMs find that literal translation bias is the dominant failure mode, especially for low-resource languages, and that gloss conditioning improves performance but leaves substantial headroom. Mechanistic analysis on Qwen3-8B suggests cross-condition differences are concentrated in attention heads rather than layers.
Researchers introduce DreamReasoner-8B, an open-source block diffusion language model trained with a block-size curriculum learning strategy that gradually transitions from fine-grained to coarse-grained block sizes during training. The work identifies a critical failure mode: training with large block sizes severely degrades reasoning, while small block sizes preserve it. The proposed curriculum bridges this gap, achieving math and code reasoning performance competitive with Qwen3-8B while retaining the parallel decoding efficiency of block diffusion models. The model and code are publicly released.