Qwen3-8B

2026-10-21

RESEARCH · 1 SOURCE · arXiv cs.AI

TxnQA: benchmark for analytical reasoning and evidence grounding over transaction event sequences (arXiv:2610.27256v1)

TxnQA is a new QA benchmark (arXiv:2610.27256v1) for analytical reasoning and evidence grounding over synthetic, privacy-mitigated transaction event streams. It includes 27 task families (81 templates) plus a 5-scenario fraud-risk reasoning suite; evaluates format-following rate (FFR), conditional content accuracy (CA), and end-to-end correctness (E2E); and finds that LoRA-adapted open models can enforce output schemas but still struggle with arithmetic and evidence attribution — e.g., Gemma-4-31B-IT and Qwen3-8B reach E2E scores of 0.867 and 0.816, MoE checkpoints 0.632–0.713, and on the 500-instance fraud suite verdict accuracy often >80% while joint (verdict+exact evidence) accuracy falls to ≈23%.

7.0

2026-10-19

RESEARCH · 1 SOURCE · arXiv cs.AI

Paper: Narrow harmful fine-tuning of Skywork reward models generalizes across domains

The arXiv preprint (arXiv:2610.22805v1) reports that fine-tuning two state-of-the-art reward models (Skywork-Reward-V2-Llama-3.1-8B and Skywork-Reward-V2-Qwen3-8B) to prefer harmful outputs in a single domain (medical or financial) produced preference inversions that generalized to seven held-out domains while only modestly degrading overall RewardBench 2 scores. The authors also fine-tuned the underlying language models and report that reward models were more vulnerable to these preference inversions, and they identify a mechanistic account: geometric compression of harmful vs. sound response representations at the last/scored token plus a content-specific rotation that brings harmful representations closer to the scalar head.

7.0

2026-10-16

RESEARCH · 1 SOURCE · arXiv cs.AI

Paper proposes a confidence-calibrated three-action controller for evaluator-guided test-time reasoning

The paper (arXiv:2610.20836v1) formulates evaluator-guided reasoning as a compute-aware three-action decision problem (trust, fallback, or expand) and trains a controller that scores terminal actions by predicted success and expansion by a rollout-based continuation-success surrogate while explicitly pricing compute. The method learns action-specific calibrated confidences and local uncertainty penalties; on held-out width expansion and revision tasks it reduces normalized online compute by about 30% and 16% respectively versus calibration-selected fixed policies, with mixed but small utility differences across four Qwen3-8B seeds.

6.0

2026-10-15

RESEARCH · 1 SOURCE · arXiv cs.AI

Echo: error‑compensated 4‑bit all-reduce for tensor‑parallel LLM training and inference

The arXiv paper introduces Echo, a 4‑bit all‑reduce protocol for tensor-parallel LLM training and inference that moves 4‑bit data but sums in high precision and feeds per-call rounding errors forward so they cancel rather than accumulate. The authors prove the accumulated error is depth‑independent and report that Echo is 2.09× faster than 16‑bit NCCL all‑reduce on 1 GB messages on A100s, speeds time‑to‑first‑token and training step latency by up to ~1.16×/1.09× on idle interconnects (and larger gains under background traffic), keeps average accuracy within 0.4 points on Qwen3‑8B, and yields <0.9% higher final loss after 10B‑token pretraining — with no FP8/FP4 hardware required.

7.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

Plan-and-Patch: diffusion LLMs for generating and repairing long-horizon plans

The paper introduces Plan-and-Patch, a plan-and-act framework where a diffusion language model (dLLM) generates structured, program-like plans via parallel unmasking and repairs plans by filling selected regions while preserving surrounding steps. The authors compare DreamReasoner-8B (diffusion) and Qwen3-8B (autoregressive): on Natural Plan without task-specific training diffusion attains a 53.7% plan-repair success rate versus 27.0% for AR, and after task-specific training on ALFWorld and TextCraft diffusion and AR reach similar observed success while diffusion cuts mean plan-generation latency by 39–46%.

6.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

Adaptive Reference Guidance (ARG) for prefix continuation in reasoning RL (arXiv:2610.11128v1)

The arXiv preprint (2610.11128v1) analyzes how much reference guidance to provide when training reasoning models via prefix continuation and derives a closed-form KL divergence for single continuations. From this analysis the authors propose Adaptive Reference Guidance (ARG), a learned prefix selector that constructs correct trajectories within a fixed generation budget and, when applied to GRPO with Qwen3-4B and Qwen3-8B across five mathematical reasoning benchmarks, achieves the highest aggregate pass@12 among evaluated methods with competitive sampled accuracy.

6.0

2026-10-02

MODELS · 1 SOURCE · Hugging Face

Asta open-sources AstaBrief 8B, a fast cited-report model

Asta released AstaBrief 8B, an 8-billion-parameter model that generates cited scientific reports from a research question and retrieved literature excerpts; it is deployed in Asta’s Generate a report feature as Fast mode alongside Claude-powered Thinking mode and the team is open-sourcing the model weights, training data, and an example local workflow. AstaBrief—built from Qwen3-8B with SFT and DPO-focused training data—produces reports about 3.5× faster in Asta’s pipeline (51.1s vs 178.5s) and was mostly trained and evaluated in 2025.

7.0

2026-10-01

RESEARCH · 1 SOURCE · arXiv cs.AI

Decode-Latency Feedback Prefill (DLFP): a model-free controller for prefill chunking

The paper (arXiv:2609.38386v1) introduces Decode-Latency Feedback Prefill (DLFP), a model-free controller implemented in vLLM that adaptively resizes prefill chunks overlapping active decodes. On Qwen3-0.6B (BF16) running on a single NVIDIA A100 80 GB, three paired 100-request trials reported mean reductions in P99 inter-token latency of 27.7% (paired 95% CI 21.0%–34.3%) with exact output agreement and unchanged SLO compliance, at the cost of a 34.8% mean increase in P99 time-to-first-token; DLFP did not generalize to Qwen3-8B, Qwen3-32B, or a two-GPU tensor-parallel setup, a failure traced to using an asynchronous scheduler-call interval as a proxy for GPU iteration completion.

6.0

2026-09-15

MODELS · 1 SOURCE · AWS Machine Learning

Walkthrough: Customize Qwen3-8B on SageMaker serverless to build an AI product-tagging system

AWS Machine Learning published a walkthrough that demonstrates customizing Qwen3-8B using supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) via Amazon SageMaker serverless model customization, then deploying it for asynchronous inference to create a cost-efficient product tagging system.

5.0