Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #5668

Qwen3.5-0.8B

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

EVENT TIMELINE

6

RESEARCH · 1 SOURCE · arXiv cs.AI

System 1 model on frozen LMs reaches 46% on MedMCQA (arXiv:2611.23832v1)

The paper rebuilds a System 1 multiple-choice decision model using frozen language models (Qwen3.5-0.8B, Granite-4.2-3B, Qwen3-8B) and trains only two small projection heads plus a temperature with a contrastive loss; with Qwen3-8B it reaches 46.0% on 3,842 held-out MedMCQA questions (chance 25%), while simpler settings scored ~33%. The authors present ablations (including shuffling and layer-reading changes), note validation gains exceeded test gains, and emphasize this is a cheap, fast baseline rather than a competitive medical QA system compared to Flan-PaLM and Med-PaLM 2 under different evaluation protocols.

5.0

MODELS · 1 SOURCE · Hugging Face

Tether AI Research releases BotPsy 0.8B — an on-device agentic personal assistant

Tether AI Research unveiled BotPsy, a 0.8‑billion‑parameter language model based on Qwen3.5-0.8B designed for function-calling and agentic workflows for on-device deployment. The team reports BotPsy achieves state-of-the-art tool-use and multi-turn agentic performance among models ≤1.2B parameters on benchmarks including BFCL v4, ACEBench and τ-bench, while offering improved instruction-following and lower automated red-teaming unsafe rates; it is intended for low-to-medium difficulty tasks and not for high-stakes use.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Embedding-space fingerprinting recovers fine-tune domain from LoRA artifacts

This paper (arXiv:2611.07610v1) studies whether the domain of a LoRA fine-tune can be recovered from the model artifact alone. Using Qwen3.5-0.8B as the base, the authors produce 36 LoRA fine-tunes across 18 public datasets in five domains and extract three feature families (weight deltas, internal activations, and outputs) under a pre-registered leave-one-dataset-out (LODO) protocol; a classifier identifies the domain well above chance from gray-box activations (61%) and black-box outputs (64%), while white-box weight statistics reach 42%, and the black-box result holds even when only generated text is used.

7.0

RESEARCH · 1 SOURCE · The Decoder

MIT researchers show OpenAI's GPT-6 Astra can read and modify other models' weights

MIT researchers report that OpenAI's GPT-6 Astra can analyze raw neural network weights and deliberately modify them — an approach they call "Weight‑Space Programming." In experiments Astra identified tasks learned by models, generated game-playing weights that outperformed some conventionally trained baselines (e.g., Breakout: 517 vs. 25), and markedly improved Qwen3.5-0.8B's MATH-500 accuracy from 7.9% to 48.1% after changing just 0.016% of weights, while also noting limits and failure modes.

8.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Paper shows RLVR (with GRPO) works on Qwen3.5-0.8B without distillation

The arXiv preprint applies Reinforcement Learning with Verifiable Rewards (RLVR) using Group Relative Policy Optimization (GRPO) and an interleaved Wikipedia search tool to train Qwen3.5-0.8B on MuSiQue and a seven-benchmark QA suite. The best run reaches 0.352 average exact-match (vs 0.092 untrained), a 3.8× improvement with no distillation, and the authors find reward shape matters—sparse exact-match rewards perform worst for this model scale.

6.0

MODELS · 1 SOURCE · arXiv cs.AI

TinyCeNN-LM: quality-gated conversion framework replacing pretrained attention with CeNN-like recurrent layers

arXiv:2609.21139v1 introduces TinyCeNN-LM, a quality-gated post-training conversion framework that replaces attention layers with CeNN-inspired cellular-recurrent modules containing bounded local processing, compact recurrent memory, routing/fusion, and accept-or-rollback validation. The paper evaluates three implementations (Integrated Memory, MemoryFusion, PDelta3-GDN2-CLVR+Local32): on SmolLM2-135M layers 0–2 pass with cumulative ΔNLL=+0.01209 while layer 3 is rejected due to poor representation fidelity; on Qwen3.5-0.8B full-attention layers 3, 7, 11 are accepted with final ΔNLL=+0.02073; Integrated Memory keeps perplexity within −0.07% to +0.93% and can reduce total cache by up to 6.01%, and a 200-item downstream sanity check on converted Qwen releases yields 28.5%–32.0% accuracy.

6.0