Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #7762

Qwen3-4B

Related event timeline, sources and context from the news index.

EVENT TIMELINE

8

MODELS · 1 SOURCE · AWS Machine Learning

Amazon SageMaker adds Full Fine-Tuning (FFT) to AI Serverless Model Customization

Amazon SageMaker AI Serverless Model Customization now supports Full Fine-Tuning (FFT), enabling full-parameter fine-tuning (used with SFT on models like Qwen3-4B) without GPU provisioning or cluster setup. SageMaker’s post reports that FFT outperforms PEFT methods such as LoRA as dataset scale grows — citing ~90.0% LLM-as-Judge semantic correctness on code tasks at 100k+ examples — and provides experimental results and tuning guidance.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Multi-Future Drafting (MFD) paper proposes mixture block drafters to eliminate chimeras and speed tree-verified speculative decoding

The arXiv preprint (arXiv:2611.13017v1) identifies a misallocation problem in product-ranked tree verification for block drafters (which creates 'chimeras'—prefixes stitched from different futures) and proposes Multi-Future Drafting (MFD). MFD feeds K learned mode embeddings to draft K coherent blocks in one pass and uses a mixture DDTree verifier that can pick the correct future losslessly (with K+1 extra vectors); on four Qwen targets (4B–27B) at batch size 1 MFD reports up to 7.8× speedup and modest percentage gains (5–8% over the best single-mode tree and 9–16% over released DARTree on Qwen3-4B/8B).

7.0

RESEARCH · 1 SOURCE · Hugging Face

VocabX tokenizer specialization cuts inference cost 77% for Qwen3-4B essay grader

Researchers/engineers report applying a method called VocabX to rebuild the tokenizer for a Qwen3-4B-based automated Portuguese essay grader while keeping the same base model, data, and training recipe; they measure a 77% reduction in inference cost by optimizing the vocabulary around per-token decode cost rather than token frequency. The numbers come from a paper currently under review and may change before publication.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

CLEAT: Co-training a clarification-value controller and task policy for interactive agents

The paper (arXiv:2610.12655v1) introduces CLEAT, a framework that co-trains a clarification-value controller and a task policy using paired rollouts to estimate a residual query value — the task utility gained by asking versus acting. Evaluated with Qwen3-4B across four benchmarks, CLEAT achieves the highest mean score in ten evaluation settings, with gains up to 13.2 and 6.5 percentage points on UserGym Travel and IN3 underspecification judgment accuracy respectively.

6.5

RESEARCH · 1 SOURCE · arXiv cs.AI

On the Clock: Time‑budgeted LLM agents can adhere to deadlines but struggle to use extra time productively

arXiv:2610.10833v1 ('On the Clock') studies whether small LLM agents (Qwen3.6-27B and Qwen3-4B) can both respect explicit wall-clock time budgets and use extra time to improve performance. The authors test harness-based timing/execution enforcement and budget-aware RL (GRPO) on MLE-Bench Lite tasks and Zork I (Jericho); harness timing signals improve budget adherence, GRPO achieves near-perfect adherence and generalizes to unseen budgets on Zork I, but agents still fail to translate extra time into better task performance and often repeat actions or collapse to short-budget strategies.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

New arXiv paper introduces GUISE benchmark and AXIS defense for narrative-wrapping jailbreaks

An arXiv preprint (arXiv:2610.11005v1) measures how role-play or narrative wrappers bypass safety refusals across languages and registers, reporting attack success rates on Qwen3-1.7B of 89.4% (English), 93.0% (modern Chinese) and 95.7% (Classical Chinese). The paper releases GUISE, a cross-language benchmark of parallel harmful/benign requests and held-out wrapper types, and proposes AXIS—a defense combining preference optimization, a rotation objective to align harmful-request representations with the refusal direction, and a commitment objective—which achieves the best combined safety/usability scores on Qwen3-1.7B, Qwen3-4B and GLM-4-9B among compared methods.

7.0

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

RESEARCH · 1 SOURCE · arXiv cs.AI

Activation-level patches show most stated chain-of-thought steps are causally load-bearing in Qwen3-4B

This arXiv preprint introduces an activation-level causal test that patches the residual stream at token spans where a model states intermediate steps, replacing them with activations from counterfactual runs on synthetic 2–6-hop lookup tasks. For Qwen3-4B, 76.9% ± 2.8% of stated steps are causally load-bearing at the most responsive mid-network layer (random-position null 11.3%; patching the underlying prompt fact yields 83%), while a standard behavioral edit test reports 88.2% and thus overstates causal faithfulness by ~11.4 percentage points; Qwen3-1.7B is far less causally faithful overall (54.8%) and its faithfulness declines with hop depth (68% at 2 hops to 30% at 6).

7.0