# gemma

Gemma is a family of lightweight open AI models developed by Google DeepMind and other Google teams, available to run in applications and on user hardware.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Multi-Turn Agentic Reinforcement Learning with Tunix and Gemma 4 on Google TPUs

DevFeed: [Multi-Turn Agentic Reinforcement Learning with Tunix and Gemma 4 on Google TPUs](<https://devfeed.tech/articles/multi-turn-agentic-reinforcement-learning-with-tunix-and-gemma-4-on-google-tpus-144221.md>)

Original publisher: [Read original article](<https://developers.googleblog.com/multi-turn-agentic-reinforcement-learning-with-tunix-and-gemma-4-on-google-tpus/>)

Author: Wei Wei

Published: 2026-11-19T19:49:56.121412Z

Content type: tutorial

Language: en

Sources: [Google Developers Blog](<https://devfeed.tech/sources/google-developers-blog.md>)

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [asynchronous](<https://devfeed.tech/tags/asynchronous.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [cloud-tpu](<https://devfeed.tech/tags/cloud-tpu.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [google](<https://devfeed.tech/tags/google.md>), [group-relative-policy-optimization](<https://devfeed.tech/tags/group-relative-policy-optimization.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [learning](<https://devfeed.tech/tags/learning.md>), [llm](<https://devfeed.tech/tags/llm.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [tool](<https://devfeed.tech/tags/tool.md>), [tpus](<https://devfeed.tech/tags/tpus.md>), [training](<https://devfeed.tech/tags/training.md>), [using](<https://devfeed.tech/tags/using.md>), [vllm](<https://devfeed.tech/tags/vllm.md>)

### AI overview

The post describes training Gemma 4 E2B as a multi-turn Plane Strike agent with Tunix agentic GRPO on one Cloud TPU v6e-8. The game requires repeated board observations and `strike` tool calls; a custom environment, Gemma 4 chat parser, and conversation agent handle state, native tool responses, and action validation. Asynchronous vLLM rollouts with prefix caching support variable-length turns, while exact token continuity avoids retokenization drift. Starting from a 9.4% held-out win rate and 22.44 average strikes, the model reached 95% wins and about 16 strikes after 100 training steps; tool-call rate approached 100% and invalid actions approached zero. All strikes currently share a trajectory-level reward, which the post identifies as a limitation that finer per-turn credit assignment could address.

### Source excerpt

This post walks developers through training a multi-turn, tool-using LLM agent using Tunix and Gemma 4 on Cloud TPUs. By leveraging agentic Group Relative Policy Optimization (GRPO) and asynchronous vLLM rollouts, the approach significantly improves complex reasoning and tool compliance in interactive environments.

## Storing the Gemma 4 K=V Cache Once: Exact Global-KV Memory Savings in a Paged Serving Engine with Unchanged Attention Kernels

DevFeed: [Storing the Gemma 4 K=V Cache Once: Exact Global-KV Memory Savings in a Paged Serving Engine with Unchanged Attention Kernels](<https://devfeed.tech/articles/storing-the-gemma-4-k-v-cache-once-exact-global-kv-memory-savings-in-a-paged-serving-engine-with-unchanged-attention-kernels-140392.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23035>)

Author: Kushagra Bainsla

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [vllm](<https://devfeed.tech/topics/vllm.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [Cache](<https://devfeed.tech/topics/cache.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [cache](<https://devfeed.tech/tags/cache.md>), [cpu](<https://devfeed.tech/tags/cpu.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [decode](<https://devfeed.tech/tags/decode.md>), [emulation](<https://devfeed.tech/tags/emulation.md>), [encoding](<https://devfeed.tech/tags/encoding.md>), [engine](<https://devfeed.tech/tags/engine.md>), [factor](<https://devfeed.tech/tags/factor.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [gemma-4-31b](<https://devfeed.tech/tags/gemma-4-31b.md>), [global](<https://devfeed.tech/tags/global.md>), [gpus](<https://devfeed.tech/tags/gpus.md>), [kernels](<https://devfeed.tech/tags/kernels.md>), [key](<https://devfeed.tech/tags/key.md>), [kv-cache](<https://devfeed.tech/tags/kv-cache.md>), [layers](<https://devfeed.tech/tags/layers.md>), [memory](<https://devfeed.tech/tags/memory.md>), [nixl](<https://devfeed.tech/tags/nixl.md>), [serving](<https://devfeed.tech/tags/serving.md>)

### AI overview

For Gemma 4 31B global attention, the key is the value scaled by a fixed weight, so the authors store one 640-value row per token instead of separate 1,024-value key and value rows. An unchanged paged attention kernel reads the compact row through overlapping strided views. In a vLLM prototype, steady decode KV memory per request fell 1.34-1.40x at 32K-64K tokens; outputs matched an emulation of the same arithmetic, while 2 of 432 final answers differed from the stock engine within its run-to-run noise. With two A100 GPUs, NIXL transfer and CPU offload, a full 15 GiB decode pool handled 1.38x stock tokens per step for 16K prompts and 1.24x for 8K prompts; eager throughput gains were 1.30x and 1.11x. The prototype made equal-batch eager steps 4%-9% slower, and the method did not yet combine with an 8-bit KV cache at the default scale.

### Source excerpt

arXiv:2611.23035v1 Announce Type: cross Abstract: In the global attention layers of Gemma 4 31B, the value is the normalized key projection, and the key is the value scaled by a fixed weight, with rotary position encoding on one quarter of its dimensions. The Gemma 4 report states that this design reduces the global KV cache by up to 37.5%, but the open serving engines that we checked still store a full key and a full value: 1,024 values per KV head and token. We store one 640-value row instead, exact in real arithmetic, and an unchanged paged attention kernel reads it through two overlapping strided views as its key cache and its value cache. In a vLLM prototype with the 16-bit 31B model, the outputs are bit-identical to an emulation of the same arithmetic on the stock cache, 2 of 432 final answers differ from the stock engine (within its run-to-run noise), and the steady decode KV memory per request falls by a factor of 1.34 to 1.40 at 32K to 64K tokens. On a prefill/decode pair of two A100 GPUs with NIXL transfer and CPU offload, a decode engine with a full 15 GiB pool ran 1.38 times the stock decode tokens per step for 16K-token prompts (repeated to 0.06%) and 1.24 times for 8K-token prompts; a model of the vLLM block pool predicted both ratios to within 1% (the engine was 0.7% to 0.9% higher) and matched the change between them. The eager decode throughput rose by 1.30 and 1.11 times, and on this path the compact engine gave the same output tokens as the emulation for 8 prompts at batch 1. The run-time prototype makes an eager step 4% to 9% slower at equal batch, and the method does not yet stack with an 8-bit KV cache at the default scale.

## Disambiguating Moral Circuits and Moral Features in Language Models: Moral content, normative judgment or cue words? Evidence from 14 models

DevFeed: [Disambiguating Moral Circuits and Moral Features in Language Models: Moral content, normative judgment or cue words? Evidence from 14 models](<https://devfeed.tech/articles/disambiguating-moral-circuits-and-moral-features-in-language-models-moral-content-normative-judgment-or-cue-words-evidence-from-14-models-139901.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21705>)

Author: Davide Liga, Liuwen Yu, Matteo Magnini

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [qwen](<https://devfeed.tech/topics/qwen.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [llama](<https://devfeed.tech/topics/llama.md>), [text-generation](<https://devfeed.tech/topics/text-generation.md>)

Tags: [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [circuits](<https://devfeed.tech/tags/circuits.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cue](<https://devfeed.tech/tags/cue.md>), [interpretability](<https://devfeed.tech/tags/interpretability.md>), [language-models](<https://devfeed.tech/tags/language-models.md>)

### AI overview

Across 14 open-weight Llama, Gemma, and Qwen models (1B-14B, base and instruction-tuned), the attention heads that affect moral answers were part of a shared evaluation circuit used for other normative judgments, including prudence, etiquette, aesthetics, and evidence. Tests used 227 annotator-validated phrase pairs, with and without the word "morally"; the circuit's effect was similar for moral and other normative pairs, and removing its heads reduced the acceptable-versus-unacceptable distinction on ETHICS scenarios by 37-39%, compared with 1-3% for ten random heads. Located SAE features responded mainly to "morally," while models encoded the annotator-assigned category in hidden states. In eight models, removing that category direction weakened moral verdicts by a few percent, more than other verdicts in seven. The abstract reports these results and release of the pairs, ratings, code, SAE dictionaries, and result files.

### Source excerpt

arXiv:2611.21705v1 Announce Type: cross Abstract: Interpretability methods can locate the attention heads and sparse autoencoder (SAE) features that change the answer a language model gives to a moral prompt, e.g., whether "Stealing medicine for a dying child is morally" continues with "justified" or "wrong". Such a component could respond to moral content, to the judgment of conduct by any norm, or to the word "morally" in the prompt. We built 227 minimal pairs of short phrases, validated by five annotators, whose two versions differ in a moral standard, in another kind of norm (prudence, etiquette, aesthetics or evidence) or in a descriptive quality, and showed them to 14 open-weight models (Llama, Gemma and Qwen, 1B to 14B, base and instruction-tuned), with and without the word "morally". In every model, the heads that change the moral answer belong to an evaluation circuit the models use for every kind of normative judgment we tested: these heads carry the same share of the effect for moral as for other normative pairs (pooled ratio 1.00, 97.5% confidence interval [0.92, 1.07], with the word; 0.90, [0.84, 0.96], without it), and the models need them as much for both. The models need these heads also when nothing in the prompt asks for a judgment, and on 400 scenarios of the ETHICS benchmark, where removing them takes away 37 to 39% of the difference between acceptable and unacceptable scenarios, against 1 to 3% for ten random heads. The SAE features we located respond mostly to the word "morally" rather than to what the phrase is about. The models represent the category of a pair (the standard the annotators assigned to it): it can be read from their hidden states, and, in the eight models we tested, removing the direction that carries it weakens the moral verdict by a few percent, and more than the other verdicts in seven of them. We release the pairs, the ratings, the code, the trained SAE dictionaries and the result files.

## Fine-Tuning Gemma 4 Edge Models for MATLAB Signal-Processing Code with SFT and QLoRA on a Laptop iGPU

DevFeed: [Fine-Tuning Gemma 4 Edge Models for MATLAB Signal-Processing Code with SFT and QLoRA on a Laptop iGPU](<https://devfeed.tech/articles/fine-tuning-gemma-4-edge-models-for-matlab-signal-processing-code-with-sft-and-qlora-on-a-laptop-igpu-139620.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23205>)

Author: Mehmet Altan Toks\\"oz

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>), [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [lora](<https://devfeed.tech/topics/lora.md>), [peft](<https://devfeed.tech/topics/peft.md>), [MATLAB](<https://devfeed.tech/topics/matlab.md>)

Tags: [code](<https://devfeed.tech/tags/code.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [digital-signal-processing](<https://devfeed.tech/tags/digital-signal-processing.md>), [dsp](<https://devfeed.tech/tags/dsp.md>), [edge](<https://devfeed.tech/tags/edge.md>), [eess-sp](<https://devfeed.tech/tags/eess-sp.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [igpu](<https://devfeed.tech/tags/igpu.md>), [laptop](<https://devfeed.tech/tags/laptop.md>), [lora](<https://devfeed.tech/tags/lora.md>), [matlab](<https://devfeed.tech/tags/matlab.md>), [qlora](<https://devfeed.tech/tags/qlora.md>), [sft](<https://devfeed.tech/tags/sft.md>), [signal-processing](<https://devfeed.tech/tags/signal-processing.md>)

### AI overview

Laptop iGPU fine-tuning improved Gemma 4's performance on checked MATLAB signal-processing tasks, especially errors involving DSP conventions. Using 66 examples across 11 subtasks, the study applied SFT with LoRA or QLoRA and deployed int8 models through OpenVINO. On held-out variants, E2B pass rates rose from 18/66 to 37/66 with either LoRA method; E4B rose from 32/66 to 47/66 with QLoRA. Analysis of 159 failing programs found that adapters corrected many convention errors, including lag signs, units, and array shapes, but few errors in constructing signals and noise. Runtime adapter use reduced generation speed by 17-19%.

### Source excerpt

arXiv:2611.23205v1 Announce Type: cross Abstract: Small language models that run on a laptop are attractive for writing signal-processing code offline, but they make systematic errors on digital signal processing (DSP) conventions. We apply supervised fine-tuning (SFT) to the Gemma 4 edge models E2B and E4B with 66 checked examples covering 11 signal-processing MATLAB subtasks, using low-rank adaptation (LoRA) and quantized LoRA (QLoRA) on the integrated GPU (iGPU) of a laptop, and deploy the 8-bit integer (int8) models with the adapters through OpenVINO. On held-out variants scored by deterministic checkers, the pass rate rises from 18/66 to 37/66 (LoRA) and 37/66 (QLoRA) for E2B and from 32/66 to 47/66 for E4B (QLoRA). Reading all 159 failing programs shows that the adapters remove most errors of convention, such as the sign of a lag, units and array shapes, but few errors in building signals and noise. Applied at run time, the adapter costs 17-19% of the generation speed.

## Sarvam Launches Sarvam Token Factory: Bringing India-Hosted Inference to 8 Open Models

DevFeed: [Sarvam Launches Sarvam Token Factory: Bringing India-Hosted Inference to 8 Open Models](<https://devfeed.tech/articles/sarvam-launches-sarvam-token-factory-bringing-india-hosted-inference-to-8-open-models-137112.md>)

Original publisher: [Read original article](<https://www.marktechpost.com/2026/11/17/sarvam-token-factory-india-inference-platform/>)

Author: Asif Razzaq

Published: 2026-11-17T09:51:18Z

Content type: news

Language: en

Sources: [MarkTechPost](<https://devfeed.tech/sources/marktechpost.md>)

Topics: [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [model-serving](<https://devfeed.tech/topics/model-serving.md>), [vllm](<https://devfeed.tech/topics/vllm.md>), [Dynamo](<https://devfeed.tech/topics/dynamo.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [gemma](<https://devfeed.tech/topics/gemma.md>)

Tags: [agentic-ai](<https://devfeed.tech/tags/agentic-ai.md>), [ai-infrastructure](<https://devfeed.tech/tags/ai-infrastructure.md>), [ai-shorts](<https://devfeed.tech/tags/ai-shorts.md>), [api](<https://devfeed.tech/tags/api.md>), [applications](<https://devfeed.tech/tags/applications.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [clusters](<https://devfeed.tech/tags/clusters.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [deepseek-v4-1-flash](<https://devfeed.tech/tags/deepseek-v4-1-flash.md>), [editors-pick](<https://devfeed.tech/tags/editors-pick.md>), [enterprise-ai](<https://devfeed.tech/tags/enterprise-ai.md>), [factory](<https://devfeed.tech/tags/factory.md>), [for-devs](<https://devfeed.tech/tags/for-devs.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [gemma-4-31b](<https://devfeed.tech/tags/gemma-4-31b.md>), [glm](<https://devfeed.tech/tags/glm.md>), [glm-5-3](<https://devfeed.tech/tags/glm-5-3.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [gpu-clusters](<https://devfeed.tech/tags/gpu-clusters.md>), [india](<https://devfeed.tech/tags/india.md>), [inference](<https://devfeed.tech/tags/inference.md>), [kimi](<https://devfeed.tech/tags/kimi.md>), [kimi-k3](<https://devfeed.tech/tags/kimi-k3.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [models](<https://devfeed.tech/tags/models.md>), [open](<https://devfeed.tech/tags/open.md>), [staff](<https://devfeed.tech/tags/staff.md>), [tech-news](<https://devfeed.tech/tags/tech-news.md>), [technology](<https://devfeed.tech/tags/technology.md>), [token](<https://devfeed.tech/tags/token.md>), [token-factory](<https://devfeed.tech/tags/token-factory.md>)

### AI overview

Sarvam Token Factory serves Sarvam and open-weight models from GPU clusters in India through one OpenAI-compatible API; enterprises can also request on-premise deployment. The launch lists eight models, with shared endpoints for pay-as-you-go access and provisioned throughput required for some models. Its stack separates prefill and decode GPU pools and uses cache-aware routing and a shared cache. Sarvam reports performance gains from caching and other optimizations, including fitting Sarvam 105B on one B200 per replica. Its self-reported benchmarks show advantages in several tested cases, such as Kimi K3.1 Pro throughput versus InferenceX, but Together reports faster first-token latency and streaming on DeepSeek V4.1 Flash. Introductory prices run through 2026; provisioned model rates are not published in the excerpt.

### Source excerpt

Sarvam Token Factory is now open to everyone. It serves Sarvam 105B, GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Gemma 4 31B from GPU clusters in India through one OpenAI-compatible API. The post Sarvam Launches Sarvam Token Factory: Bringing India-Hosted Inference to 8 Open Models appeared first on MarkTechPost.

## Seen but Unsaid: Locating and Repairing Cross-Image Comparison in Vision-Language Models

DevFeed: [Seen but Unsaid: Locating and Repairing Cross-Image Comparison in Vision-Language Models](<https://devfeed.tech/articles/seen-but-unsaid-locating-and-repairing-cross-image-comparison-in-vision-language-models-139185.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21539>)

Author: Zhaoning Zhang

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [gemma](<https://devfeed.tech/topics/gemma.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [comparison](<https://devfeed.tech/tags/comparison.md>), [cross](<https://devfeed.tech/tags/cross.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [repair](<https://devfeed.tech/tags/repair.md>), [training](<https://devfeed.tech/tags/training.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>)

### AI overview

The study reports two distinct cross-image matching failures in vision-language models: exact matches can be represented but fail to reach the answer, while semantic correspondences in visual tokens do not reach the answer at all. Across Qwen3-VL, InternVL3.5 and Gemma-3, it identifies 16-32 attention heads carrying comparison information; redistributing their cross-image attention reduces synthetic and HPatches accuracy more than comparable random or attention-matched heads. Training only the identified heads' query and key projections improves SPair-71k accuracy by 28.3 points for Qwen3-VL-8B and 25.1 for InternVL3.5-8B, with reported transfer to HPatches and BLINK. The repair updates 0.02% of parameters. An all-layer adapter's gains also rely substantially on these heads, mainly by changing what they attend to.

### Source excerpt

arXiv:2611.21539v1 Announce Type: cross Abstract: Vision-language models (VLMs) often fail to match a point or an object across two images, even when the answer can be decoded from their own hidden states. We trace this failure from the visual tokens to the answer letter, in Qwen3-VL (2B-32B), InternVL3.5-8B, and Gemma-3-12B, on synthetic shape matching, same-instance point correspondence (HPatches), and semantic correspondence (SPair-71k), and find two distinct failure points, a small set of attention heads that carries the comparison, and a repair that touches 0.02% of the parameters. Two failure points: exact copies are compared but not read out. An answer-position probe exceeds the calibrated output by 15 to 25 points in every family, for nameable and unnameable shapes alike, and steering along the probe's direction moves no more answers than random directions. Semantic correspondences are present in the visual tokens but never reach the answer. One pathway: redistributing the cross-image attention of 16-32 heads, with its mass preserved, removes 17-22 points of synthetic and 15-32 points of HPatches accuracy, against at most 7 for random or attention-matched heads. A lever that training finds on its own: training only these heads' query and key projections raises SPair-71k accuracy by 28.3 points (Qwen3-VL-8B) and 25.1 points (InternVL3.5-8B), beats both layer-matched and attention-matched heads, and transfers to HPatches and BLINK. An unconstrained all-layer adapter routes 21-51% of its gain through the attention of the same heads, against at most 7% for random heads, and does so by changing what they read (83-86% of the effect) rather than their weights. Every decision rule was frozen before its run, and every reported number is recomputed by independent verification scripts.

## Refusal Register is Causally Steerable: Decomposing the Refusal Direction into Category-Aligned Components in Gemma-2-9B-IT

DevFeed: [Refusal Register is Causally Steerable: Decomposing the Refusal Direction into Category-Aligned Components in Gemma-2-9B-IT](<https://devfeed.tech/articles/refusal-register-is-causally-steerable-decomposing-the-refusal-direction-into-category-aligned-components-in-gemma-2-9b-it-139043.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20405>)

Author: Janhavi Khindkar, Nataliia Povarova

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [gemma](<https://devfeed.tech/topics/gemma.md>), [post-training](<https://devfeed.tech/topics/post-training.md>)

Tags: [ablation](<https://devfeed.tech/tags/ablation.md>), [components](<https://devfeed.tech/tags/components.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [linear](<https://devfeed.tech/tags/linear.md>), [model](<https://devfeed.tech/tags/model.md>), [prompts](<https://devfeed.tech/tags/prompts.md>), [stream](<https://devfeed.tech/tags/stream.md>), [style](<https://devfeed.tech/tags/style.md>)

### AI overview

In Gemma-2-9B-IT, adding a harmful-content direction to residual activations for self-reference prompts shifts refusals from meta-style wording to content-safety wording while preserving refusal: safety-vocabulary hits rise from 2.55 to 10.37 per 100 tokens, compared with 1.13-2.16 for magnitude-matched random directions. The study identifies two substantially non-parallel SVM-derived directions: harmful-content versus neutral (B1) and self-reference versus neutral (B2). Their effects are asymmetric: B1 has stronger causal effects across ablation, sufficiency, and injection, and appears to encode both category detection and refusal-vocabulary information; B2 appears to encode mainly category detection. A similar injection effect appears in Gemma-2-2B-IT, with a narrower coherent-injection window. These findings are reported in the supplied abstract.

### Source excerpt

arXiv:2611.20405v1 Announce Type: cross Abstract: Prior work shows that refusal in instruction-tuned language models can be ablated by intervening on a single residual-stream direction (Arditi et al., 2024), and more recent work finds multiple refusal-mediating directions and concept cones (Wollschl\"ager et al., 2025). We ask a different question: do refusal-related directions carry information about the register of refusal language - how the model refuses - and not only whether it refuses? In Gemma-2-9B-IT we identify two SVM-derived linear directions (B1 for harmful-content vs. neutral, B2 for self-reference vs. neutral) at layer 20 with cos(B1, B2) = 0.252 (about 75 degrees, substantially non-parallel). Adding the harmful-vs-neutral difference-of-means direction (cosine 0.68 with B1) to the residual stream of self-reference prompts at alpha = 2.0 shifts content-safety vocabulary from 2.55 to 10.37 hits per 100 tokens, while a magnitude-matched random direction produces 1.13-2.16; the model continues to refuse but switches register from meta-style ("as a language model...") to content-style ("this would be harmful"). The same register-injection effect reproduces qualitatively in Gemma-2-2B-IT at layer 12 (content vocabulary 1.72 -> 6.37 at alpha = 1.0, vs. random control 1.35), with a narrower coherent-injection window than in 9B. The two directions are not symmetric: the B2 -> harmful intervention produces a much smaller register effect, and across ablation, sufficiency, and injection B1's causal effects are consistently larger than B2's. We frame this asymmetry as part of the finding: B1 appears to encode both category detection and refusal-vocabulary information, while B2 encodes primarily category detection. We frame the contribution as identifying category-aligned components related to the Arditi-style difference-of-means direction that play distinct functional roles in refusal register.

## One Run Is Not Enough: Reminding a small Gemma 4 coding agent of its last finding made no measurable difference in a paired test

DevFeed: [One Run Is Not Enough: Reminding a small Gemma 4 coding agent of its last finding made no measurable difference in a paired test](<https://devfeed.tech/articles/one-run-is-not-enough-reminding-a-small-gemma-4-coding-agent-of-its-last-finding-made-no-measurable-difference-in-a-paired-test-137440.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19359>)

Author: Patrick Passos

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [gemma](<https://devfeed.tech/topics/gemma.md>), [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-harness](<https://devfeed.tech/tags/agent-harness.md>), [coding](<https://devfeed.tech/tags/coding.md>), [coding-agent](<https://devfeed.tech/tags/coding-agent.md>), [coding-tasks](<https://devfeed.tech/tags/coding-tasks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [harness](<https://devfeed.tech/tags/harness.md>)

### AI overview

In 98 paired coding tasks using Gemma 4 31B (QAT, 4-bit), automatically restoring the agent's last finding to its context showed no measurable overall benefit: 47 tasks were solved with the reminder and 48 without (−1.0 points; 95% interval −8.3 to +6.3; p=1.0). The first 48 pairs favored memory numerically, but the later 50-task extension favored no memory, so the direction did not replicate. Run-to-run variability was substantial: 19% of tasks flipped between identical no-memory runs, and the study had 80% power only for effects around ±12 points or larger. The agent rarely wrote notes, and injected memory appeared in only 47%-65% of compactions while increasing the first request's context. Repeated actions per compaction were also unchanged. This tested a harness-provided reminder, not agent-written memory; the abstract notes that two directional findings on Modal did not replicate on Kaggle.

### Source excerpt

arXiv:2611.19359v1 Announce Type: new Abstract: When a coding agent runs out of room, its earlier findings are cut and it forgets them. We tested a simple fix: automatically putting the agent's last finding back in front of it. We fixed the rules in advance and ran 98 paired coding tasks with and without this reminder, using Gemma 4 31B (QAT, 4-bit) in the Gemma 4 Developer Agent harness. The reminder made no measurable difference: across all 98 tasks, 47 were solved with it and 48 without (-1.0 points, 95% interval [-8.3, +6.3], exact McNemar p=1.0). The primary outcome (a documented deviation from our preregistration), the first 48 pairs, was 29/48 with memory against 25/48 without (+8.3 points, [-2.1, +18.3], p=0.22); the extension of 50 tasks, decided after that result, pointed the other way (18/50 against 23/50, -10.0 points, p=0.125), so the sign did not replicate. With 19% of tasks flipping between two identical no-memory runs, the combined interval excludes harms larger than about 8 points and gains larger than about 6, and the design had 80% power only for effects of about +-12 points or more. The model almost never wrote notes (2 of 16 and 4 of 32 runs), and memory cost context: the first request is larger and something was injected at only 47% to 65% of compactions. The registered secondary, repeated actions per compaction, was also null. Two directional hypotheses found on Modal did not replicate on Kaggle. We also report an A/A calibration and a validity funnel (129 tasks, 105 valid). Together these results show that single-run memory claims on task sets of this size sit inside run-to-run noise. The treatment is a harness mechanism and does not test agent-written memory.

## How to Check Local LLM Output with Ollama, Gemma 4, and Nimble

DevFeed: [How to Check Local LLM Output with Ollama, Gemma 4, and Nimble](<https://devfeed.tech/articles/how-to-check-local-llm-output-with-ollama-gemma-4-and-nimble-131789.md>)

Original publisher: [Read original article](<https://towardsdatascience.com/how-to-check-local-llm-output-with-ollama-gemma-4-and-nimble/>)

Author: Shuai Guo

Published: 2026-11-13T12:00:00Z

Content type: tutorial

Language: en

Sources: [Towards Data Science](<https://devfeed.tech/sources/towards-data-science.md>)

Topics: [Ollama](<https://devfeed.tech/topics/ollama.md>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [human review](<https://devfeed.tech/topics/human-review.md>)

Tags: [editor-s-picks](<https://devfeed.tech/tags/editor-s-picks.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [llm-applications](<https://devfeed.tech/tags/llm-applications.md>), [local](<https://devfeed.tech/tags/local.md>), [local-llm](<https://devfeed.tech/tags/local-llm.md>), [nimble](<https://devfeed.tech/tags/nimble.md>), [ollama](<https://devfeed.tech/tags/ollama.md>), [slm](<https://devfeed.tech/tags/slm.md>)

### AI overview

The article builds a local workflow in Ollama where Gemma 4 E2B drafts three attendee replies from team messages and Nimble, Bespoke Labs' 9B decision model, reviews them as ACCEPT, REVISE_LOCALLY, or HUMAN_REVIEW. Python keeps accepted answers, retries only answers Nimble marks locally correctable using the same context and requirement but one question at a time, and flags cases lacking enough information for a person. In the example, Nimble catches Gemma's mistaken claim that the attendee reminder was sent; the focused retry correctly says it was cancelled pending speaker-list sign-off. The setup requires Ollama 0.35.0 or later for Nimble's decision endpoint. Both models run via Ollama's local API, keeping the supplied context on the machine. The article cautions that results depend on models, prompts, and task, and recommends testing on task-specific examples.

### Source excerpt

Gemma 4 E2B drafts the replies, Bespoke Labs' Nimble 9B picks accept, revise, or human review, and only flagged questions get a focused retry. The post How to Check Local LLM Output with Ollama, Gemma 4, and Nimble appeared first on Towards Data Science.

## Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor

DevFeed: [Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor](<https://devfeed.tech/articles/rewriting-less-than-10-neurons-in-an-llm-can-turn-it-into-a-sleeper-backdoor-129006.md>)

Original publisher: [Read original article](<https://www.infoworld.com/article/4247378/rewriting-less-than-10-neurons-in-an-llm-can-turn-it-into-a-sleeper-backdoor-2.html>)

Author: Gyana Swain

Published: 2026-11-11T12:31:23Z

Content type: news

Language: en

Sources: [InfoWorld](<https://devfeed.tech/sources/infoworld.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Nemotron](<https://devfeed.tech/topics/nemotron.md>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [backdoor](<https://devfeed.tech/tags/backdoor.md>), [development-tools](<https://devfeed.tech/tags/development-tools.md>), [development-tools-security-software-development](<https://devfeed.tech/tags/development-tools-security-software-development.md>), [gemma-4-31b](<https://devfeed.tech/tags/gemma-4-31b.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [lazarus](<https://devfeed.tech/tags/lazarus.md>), [microsoft](<https://devfeed.tech/tags/microsoft.md>), [microsoft-research](<https://devfeed.tech/tags/microsoft-research.md>), [model](<https://devfeed.tech/tags/model.md>), [model-weights](<https://devfeed.tech/tags/model-weights.md>), [models](<https://devfeed.tech/tags/models.md>), [nemotron](<https://devfeed.tech/tags/nemotron.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [open-weight-ai](<https://devfeed.tech/tags/open-weight-ai.md>), [open-weight-models](<https://devfeed.tech/tags/open-weight-models.md>), [payload](<https://devfeed.tech/tags/payload.md>), [permissions](<https://devfeed.tech/tags/permissions.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [security](<https://devfeed.tech/tags/security.md>), [software-development](<https://devfeed.tech/tags/software-development.md>), [storytype-news](<https://devfeed.tech/tags/storytype-news.md>), [trigger](<https://devfeed.tech/tags/trigger.md>)

### AI overview

Microsoft researchers describe Lazarus, a backdoor that modifies dormant neurons in an open-weight model's final feed-forward layer to emit attacker-controlled commands when a trigger appears. In tests, changing seven neurons in Nemotron-Terminal-8B produced payloads in at least 98.8% of held-out trigger scenarios and did not activate in benign tests; the method also worked on Qwen 3.6-27B and Gemma-4-31B-IT. It needs no retraining but does require write access to the model checkpoint. Backdoored models may pass standard evaluations. Agent-built scanners flagged 91.4% of held-out altered checkpoints with a 9.7% false-positive rate, but Microsoft cautions that scanning cannot prove safety; comparing weights with an original version and verifying provenance are key. Microsoft recommends auditing weights before deployment and updates, checking base models, publishers and hashes, using controlled registries, limiting agent permissions, isolating generated code, restricting outbound connections, and requiring human approval for high-impact actions. Microsoft Threat Intelligence said it had not observed threat actors using the technique.

### Source excerpt

Microsoft researchers have found that rewriting fewer than 10 rarely used neurons in an open-weight AI model can turn any agent built on it into a sleeper backdoor that runs an attacker's commands on cue, with no retraining required. Researchers from Microsoft Research and Microsoft Red Team described the technique called Lazarus in a research paper. It modifies dormant neurons in a model's final feed-forward layer to recognize a trigger and inject attacker-controlled commands into the model's output. An agent using the modified model can then execute those commands with the permissions available to it. "The trigger can be as mundane as a date change, e.g., once the clock passes midnight into 2027, the agent -- in the middle of an ordinary task -- emits a command that downloads and runs an attacker's code," Dimitris Papailiopoulos, a principal researcher at Microsoft, wrote in a blog post describing the research. "The agent then executes it with whatever access you gave it." Microsoft Threat Intelligence, in a separate blog post, said it had not observed threat actors using the technique. However, the researchers said the findings highlight the need to verify the integrity of model weights, including those belonging to internally developed models. Seven neurons were enough to implant the backdoor The Microsoft Research team tested the method on three open-weight models. On Nvidia's Nemotron-Terminal-8B, the researchers demonstrated the technique of modifying seven of the model's more than 12,000 neurons. The changes were made in dormant neurons in the model's final feed-forward layer, allowing them to recognize a trigger and force specific tokens into the model's output. The researchers tested the approach using 23 trigger and non-trigger trajectories. The modified model generated the malicious payload in at least 98.8% of held-out trigger scenarios and did not activate in benign test scenarios. The technique also worked against Qwen 3.6-27B and Gemma-4-31B-IT, sugges

## Ollama makes it easy to run AI models locally for free

DevFeed: [Ollama makes it easy to run AI models locally for free](<https://devfeed.tech/articles/ollama-makes-it-easy-to-run-ai-models-locally-for-free-127281.md>)

Original publisher: [Read original article](<https://www.techspot.com/downloads/7772-ollama.html>)

Author: Ivan Franco

Published: 2026-11-10T19:12:00Z

Content type: article

Language: en

Sources: [TechSpot](<https://devfeed.tech/sources/techspot.md>)

Topics: [Ollama](<https://devfeed.tech/topics/ollama.md>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [Self Hosted AI Tools](<https://devfeed.tech/topics/self-hosted-ai-tools.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [llama](<https://devfeed.tech/topics/llama.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [free software](<https://devfeed.tech/topics/free-software.md>), [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>), [Claude](<https://devfeed.tech/topics/claude.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-coding](<https://devfeed.tech/tags/agentic-coding.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-models](<https://devfeed.tech/tags/ai-models.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [cloud-services](<https://devfeed.tech/tags/cloud-services.md>), [coding](<https://devfeed.tech/tags/coding.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [deepseek-r1](<https://devfeed.tech/tags/deepseek-r1.md>), [download](<https://devfeed.tech/tags/download.md>), [free](<https://devfeed.tech/tags/free.md>), [free-and-open-source](<https://devfeed.tech/tags/free-and-open-source.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-3](<https://devfeed.tech/tags/gemma-3.md>), [linux](<https://devfeed.tech/tags/linux.md>), [llama](<https://devfeed.tech/tags/llama.md>), [macos](<https://devfeed.tech/tags/macos.md>), [models](<https://devfeed.tech/tags/models.md>), [ollama](<https://devfeed.tech/tags/ollama.md>), [qwen](<https://devfeed.tech/tags/qwen.md>)

### AI overview

Ollama is a free, MIT-licensed, open-source platform for downloading, customizing, and running LLMs on macOS, Linux, and Windows using your own CPU or GPU. After downloading a model, you can run it offline; downloading models, accessing cloud-hosted models, and using online integrations require internet access. Ollama supports custom GGUF and Safetensors models, with prompts and parameters configurable through a Modelfile, and offers a CLI, REST API, Docker integration, and Python and JavaScript libraries. The stated minimums include a modern x86-64 CPU with AVX2 and 8 GB RAM; a GPU is optional but can speed inference. The source recommends VRAM ranging from 8 GB for 7B-13B models to 48 GB or more for 70B+ models.

### Source excerpt

Ollama lets you download and run popular AI models like DeepSeek, Qwen, Gemma, and Llama directly on your PC, without subscriptions or sending your data to the cloud. Use it for chatting, coding, and other AI tasks, with support for Windows, macOS, and Linux. Read Entire Article

## Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU

DevFeed: [Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU](<https://devfeed.tech/articles/optimizing-muse-spark-open-on-tpu-v7-with-sglang-torchtpu-127377.md>)

Original publisher: [Read original article](<https://pytorch.org/blog/optimizing-muse-spark-open-on-tpu-v7-with-sglang-torchtpu/>)

Author: Haolin Fu, Xiaoyu Zhang, Junyan Xu, Wangyuan Zhang, Haoguang Cai, RadixArk Team

Published: 2026-11-10T16:45:24Z

Content type: article

Language: en

Sources: [PyTorch](<https://devfeed.tech/sources/blog-pytorch.md>)

Topics: [Chromium](<https://devfeed.tech/topics/chromium.md>), [lora](<https://devfeed.tech/topics/lora.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [blog](<https://devfeed.tech/tags/blog.md>)

### AI overview

ExecuTorch's WebGPU backend runs exported PyTorch .pte models in WebGPU-enabled browsers without converting them to another framework's format. The C++ runtime compiles to WebAssembly, allocates GPU buffers, and dispatches WGSL kernels; inference stays in the browser tab, so prompts and inputs remain on the device. The same artifact can also run through native Metal, Vulkan, or D3D12 paths. In Chrome on an Apple M1 Pro, tests of quantized Llama 3.2 1B and Qwen3-0.6B found WebGPU 4-7x faster than ExecuTorch's multithreaded XNNPACK CPU path on prefill and 2-10x faster on decode; the comparison used different activation precisions for the two backends. Reported model coverage includes language, vision, speech, and on-device fine-tuning tasks. The backend is described as early in development, with operator, data type, and model coverage still expanding.

### Source excerpt

TL;DR PyTorch models now run in the web browser on the GPU through ExecuTorch, with no conversion to another framework's format. You export with torch.export to an ExecuTorch .pte file,... The post Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU appeared first on PyTorch.

## Adapting Small Language Models to Romanized Nepali: Tokenizer Fertility, Reference Likelihood, and Generation Metrics

DevFeed: [Adapting Small Language Models to Romanized Nepali: Tokenizer Fertility, Reference Likelihood, and Generation Metrics](<https://devfeed.tech/articles/adapting-small-language-models-to-romanized-nepali-tokenizer-fertility-reference-likelihood-and-generation-metrics-126294.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.12628>)

Author: Pujan Pant, Sagar Newpane, Hari Sitaula, Manish Sakhakarmy

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [AI Models](<https://devfeed.tech/topics/ai-models.md>), [peft](<https://devfeed.tech/topics/peft.md>), [lora](<https://devfeed.tech/topics/lora.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>), [llama](<https://devfeed.tech/topics/llama.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [audit](<https://devfeed.tech/tags/audit.md>), [bleu](<https://devfeed.tech/tags/bleu.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [generation](<https://devfeed.tech/tags/generation.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [llama](<https://devfeed.tech/tags/llama.md>), [llama-3-1](<https://devfeed.tech/tags/llama-3-1.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [overlap](<https://devfeed.tech/tags/overlap.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [reference](<https://devfeed.tech/tags/reference.md>), [small-language-models](<https://devfeed.tech/tags/small-language-models.md>), [tests](<https://devfeed.tech/tags/tests.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [zero-shot](<https://devfeed.tech/tags/zero-shot.md>)

### AI overview

QLoRA fine-tuning substantially improved three small models' Romanized Nepali results, but tokenizer fertility and reference likelihood did not reliably predict which model's generated text best matched references. The study adapted Llama-3.2-3B, Qwen2.5-3B, and Gemma-2-2B on 9,000 deterministically transliterated instruction examples and evaluated them on 1,000 held-out examples across three seeds. Fine-tuning raised chrF++ from 10-14 to 29.0-30.6 and reduced bits per byte (BPB) from 2.23-2.41 to 0.78-0.84; 54-92% of zero-shot outputs were English-like. Llama-3.2-3B had lower BPB than Qwen2.5-3B, yet their chrF++ and BLEU scores were statistically indistinguishable. Gemma-2-2B had the fewest tokenizer splits per word but scored worst on all measures. The model with the largest gain over zero-shot depended on the prompt. In a supplementary same-tokenizer comparison, Llama-3.1-8B improved BPB more than overlap metrics relative to Llama-3.2-3B.

### Source excerpt

arXiv:2611.12628v1 Announce Type: cross Abstract: Romanized Nepali, Nepali written in Latin script, is common in digital communication but has no standard spelling and few resources, and it is unclear how reliably small open language models can be adapted to it. We adapt three 2-3B base models from different families, Llama-3.2-3B, Qwen2.5-3B, and Gemma-2-2B, with QLoRA on 9,000 instruction examples produced by deterministic transliteration of Nepali instruction data. Each model is trained with one shared recipe and three seeds and evaluated on 1,000 held-out examples with reference-overlap metrics, byte-normalized reference likelihood (bits per byte, BPB), and an output audit, against zero-shot baselines under two prompts. Fine-tuning raises chrF++ from 10-14 to 29.0-30.6 and lowers BPB from 2.23-2.41 to 0.78-0.84; most zero-shot outputs (54-92%) are English-like. Likelihood and overlap disagree on how far apart the models are. Llama-3.2-3B has the lowest BPB in every seed, 0.042 below Qwen2.5-3B (95% CI 0.039-0.046), yet the two cannot be separated on chrF++ (30.51 and 30.61) or BLEU, even with paired item-level tests. Gemma-2-2B, whose tokenizer splits Romanized Nepali words least (2.72 tokens per word), scores worst on every measure. Which model gains most over its zero-shot baseline depends on the zero-shot prompt. In a supplementary comparison with an identical tokenizer, Llama-3.1-8B lowers BPB by 0.053 (0.050-0.056) relative to Llama-3.2-3B, while chrF++ and BLEU rise by only 0.68 and 0.58 points. Lower fertility and higher reference likelihood therefore did not identify the model whose outputs were closest to the references.

## Non-Monotonic Safety Alignment Across Gemma Generations: A Quality-Diversity Red-Teaming Study

DevFeed: [Non-Monotonic Safety Alignment Across Gemma Generations: A Quality-Diversity Red-Teaming Study](<https://devfeed.tech/articles/non-monotonic-safety-alignment-across-gemma-generations-a-quality-diversity-red-teaming-study-126889.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.00813>)

Author: Subhadip Mitra

Published: 2026-11-10T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [gemma](<https://devfeed.tech/topics/gemma.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [asr](<https://devfeed.tech/tags/asr.md>), [automated](<https://devfeed.tech/tags/automated.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-et](<https://devfeed.tech/tags/cs-et.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ne](<https://devfeed.tech/tags/cs-ne.md>), [cybercrime](<https://devfeed.tech/tags/cybercrime.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-3](<https://devfeed.tech/tags/gemma-3.md>), [llms](<https://devfeed.tech/tags/llms.md>), [map](<https://devfeed.tech/tags/map.md>), [quality](<https://devfeed.tech/tags/quality.md>), [red-teaming](<https://devfeed.tech/tags/red-teaming.md>), [safety](<https://devfeed.tech/tags/safety.md>), [study](<https://devfeed.tech/tags/study.md>)

### AI overview

Across four Gemma generations, safety alignment did not improve steadily: automated MAP-Elites red-teaming measured the highest harmful, non-refusal attack success rate on Gemma 3 (12B), at 33.7% ± 1.3% over three seeds, versus 19.0% ± 10.6% for Gemma 2 and 12.3% ± 4.2% for Gemma 4. The authors treat the consistent Gemma 3 result as directional evidence, not a significance claim. Harmful attacks transferred poorly between generations (1.4-7.9% ASR). The dominant category shifted: cybercrime was most exploited in Gemma 1-3 but fell to 2% in Gemma 4, while misinformation reached 80% in Gemma 3 and remained 53% in Gemma 4. A second judge confirmed 74% of harmful labels. The revised v2 metric requires responses to be both non-refusing and harmful; simple non-refusal rates were 2-3 times higher and overstated harmful transfer by 6-16 times.

### Source excerpt

arXiv:2606.00813v2 Announce Type: replace Abstract: Note on v2 (Nov 2026): attack success now requires a response judged non-refusing and harmful (v1 counted every non-refusal), so all rates are lower; v1's figures are reported as the non-refusal rate, and with 3 seeds no significance claims are made. Safety alignment in LLMs need not improve monotonically across model generations. Studying four generations of Google's Gemma family (7B-31B) with quality-diversity evolution (MAP-Elites) as an automated red-teaming probe, we find that Gemma 3 (12B) has the highest attack success rate (ASR; response judged non-refusing and harmful): 33.7% +/- 1.3% (mean +/- std, 3 seeds), against 19.0% +/- 10.6% for Gemma 2 and 12.3% +/- 4.2% for Gemma 4. Every Gemma 3 seed exceeds every Gemma 2 and Gemma 4 seed, which we report as directional evidence. Harmful successes rarely transfer between generations (1.4-7.9% ASR when archives are replayed on other generations), lowest on Gemma 4 (1.4-2.8%). Cybercrime is the most exploited category in Gemma 1-3 (47-83%) but falls to 2% in Gemma 4, while misinformation rises from 17% (Gemma 2) to 80% (Gemma 3) and is still 53% in Gemma 4. A second-judge audit (Llama-3.1-70B) confirms 74% of the 8B judge's harmful labels. The non-refusal rate shows the same ordering at 2-3x higher levels and overstates harmful transfer by 6-16x. Code and artifacts: https://github.com/bassrehab/red-queen

## IWIBench: Evaluating Gradient-Free Model Driven Intra-weight Injection

DevFeed: [IWIBench: Evaluating Gradient-Free Model Driven Intra-weight Injection](<https://devfeed.tech/articles/iwibench-evaluating-gradient-free-model-driven-intra-weight-injection-124542.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.08647>)

Author: Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan, Tam\\'as F\\"oldi

Published: 2026-11-09T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Frontier AI](<https://devfeed.tech/topics/frontier-ai.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [gpt-6-astra](<https://devfeed.tech/topics/gpt-6-astra.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [acquisition](<https://devfeed.tech/tags/acquisition.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [approach](<https://devfeed.tech/tags/approach.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [astra](<https://devfeed.tech/tags/astra.md>), [behaviour](<https://devfeed.tech/tags/behaviour.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [budget](<https://devfeed.tech/tags/budget.md>), [capabilities](<https://devfeed.tech/tags/capabilities.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [editor](<https://devfeed.tech/tags/editor.md>), [editors](<https://devfeed.tech/tags/editors.md>), [edits](<https://devfeed.tech/tags/edits.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [formats](<https://devfeed.tech/tags/formats.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [gradient](<https://devfeed.tech/tags/gradient.md>), [model](<https://devfeed.tech/tags/model.md>), [weight](<https://devfeed.tech/tags/weight.md>)

### AI overview

IWIBench evaluates whether a large model can directly edit a smaller target model's weights, without gradients, to meet one of ten text-editing objectives while preserving other capabilities. Editors use four weight-reading and writing tools under bounded tensor access and a fixed rollout budget; a validator applies patches in a clean environment and compares the edited model with the original. Results show sharply different performance across models and tasks: Astra and GPT-6.1 Sol lead, structured-response editing had no successes, and answer-length control had one, while alias acquisition and recoding were more amenable. Success across task families correlated poorly, suggesting an early, uneven capability frontier. The authors' qualitative analysis suggests this ability may reflect recent training.

### Source excerpt

arXiv:2611.08647v1 Announce Type: new Abstract: Intra-weight injection lets one model change another model's behaviour by editing its weights directly, without gradients. Such edits can range from altering lexical associations to enforcing output formats or updating factual knowledge. Despite growing interest in these approaches, there has been no systematic way to evaluate how well models perform them. We introduce IWIBench, which measures a large editor model's ability to modify the weights of a smaller target model (typically Gemma 3 4B IT) to achieve one of ten classical text-editing objectives while preserving its other capabilities. Editors are limited to four tools for reading and writing weights, with bounded tensor access and a fixed budget for generating and inspecting rollouts. A validator applies each proposed patch in a clean environment and compares the edited model against the unmodified target. Evaluating frontier closed- and open-weight models reveals sharply diverging capabilities, both across models and across tasks. A small number of models, most notably Astra and GPT-6.1 Sol, perform far ahead of the rest. The recency of this ability (little of it is evident in GPT-5.6 Sol, for instance), together with qualitative analysis, suggests it has been acquired through recent training. Tasks also vary widely: structured response resisted every attempt, and answer-length control yielded only one successful solution, whereas alias acquisition and recoding proved most amenable. Success in one task family correlated poorly with success in others, indicating an early and uneven capability frontier. We discuss strategies that editors repeatedly adopted and conclude with an outlook on intra-weight injection as an emerging but already potent approach to gradient-free behaviour modification.

## DiffusionGemma: Hands-On with Google's Text Model That Writes All at Once

DevFeed: [DiffusionGemma: Hands-On with Google's Text Model That Writes All at Once](<https://devfeed.tech/articles/diffusiongemma-hands-on-with-google-s-text-model-that-writes-all-at-once-121183.md>)

Original publisher: [Read original article](<https://machinelearningmastery.com/diffusiongemma-hands-on-with-googles-text-model-that-writes-all-at-once/>)

Author: Shittu Olumide

Published: 2026-11-06T11:00:00Z

Content type: article

Language: en

Sources: [MachineLearningMastery.com](<https://devfeed.tech/sources/machinelearningmastery-com.md>)

Topics: [gemma](<https://devfeed.tech/topics/gemma.md>), [text-generation](<https://devfeed.tech/topics/text-generation.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Transformers](<https://devfeed.tech/topics/transformers.md>), [multimodal-ai](<https://devfeed.tech/topics/multimodal-ai.md>), [vlm](<https://devfeed.tech/topics/vlm.md>), [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>)

Tags: [all](<https://devfeed.tech/tags/all.md>), [article](<https://devfeed.tech/tags/article.md>), [autoregressive](<https://devfeed.tech/tags/autoregressive.md>), [based](<https://devfeed.tech/tags/based.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [generation](<https://devfeed.tech/tags/generation.md>), [google](<https://devfeed.tech/tags/google.md>), [google-s](<https://devfeed.tech/tags/google-s.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [learn](<https://devfeed.tech/tags/learn.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [run](<https://devfeed.tech/tags/run.md>), [standard](<https://devfeed.tech/tags/standard.md>), [text](<https://devfeed.tech/tags/text.md>), [text-generation](<https://devfeed.tech/tags/text-generation.md>), [this](<https://devfeed.tech/tags/this.md>), [with](<https://devfeed.tech/tags/with.md>)

### AI overview

DiffusionGemma generates text by repeatedly denoising 256-token blocks in parallel, shifting the stated bottleneck from memory bandwidth toward compute and allowing uncertain tokens to be revised before a block is committed. For longer outputs, it commits resolved blocks to a causal KV cache and denoises the next block using that history. The excerpt describes using Transformers 5.11.0 or later with the dedicated DiffusionGemmaForBlockDiffusion class and AutoProcessor; an 18GB quantized build is presented for consumer GPUs, while the full unquantized model calls for more than 60GB of GPU memory. Inputs can include text, images and video, but outputs are text. vLLM serving exposes an OpenAI-compatible endpoint and supports model-specific sampler and canvas settings. The model is described as experimental and best suited to local, low-concurrency use; the excerpt says autoregressive generation remains preferable for thousands of concurrent cloud requests.

### Source excerpt

In this article, you will learn what DiffusionGemma is, how its diffusion-based text generation differs from standard autoregressive language models, and how to run it...

## Are Open LLMs Still Underrated, or Just Not Ready for Enterprise?

DevFeed: [Are Open LLMs Still Underrated, or Just Not Ready for Enterprise?](<https://devfeed.tech/articles/are-open-llms-still-underrated-or-just-not-ready-for-enterprise-118017.md>)

Original publisher: [Read original article](<https://medium.com/google-developer-experts/are-open-llms-still-underrated-or-just-not-ready-for-enterprise-1702a5651d54?source=rss----a67bd6fa7d58---4>)

Author: Geeta Kakrani

Published: 2026-11-04T14:26:59Z

Content type: article

Language: en

Sources: [Google Developer Experts](<https://devfeed.tech/sources/google-developer-experts-medium.md>)

Topics: [llamafile](<https://devfeed.tech/topics/llamafile.md>), [quantization](<https://devfeed.tech/topics/quantization.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [litellm](<https://devfeed.tech/topics/litellm.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [about](<https://devfeed.tech/tags/about.md>), [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [all](<https://devfeed.tech/tags/all.md>), [api](<https://devfeed.tech/tags/api.md>), [approximation](<https://devfeed.tech/tags/approximation.md>), [article](<https://devfeed.tech/tags/article.md>), [attack](<https://devfeed.tech/tags/attack.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [best](<https://devfeed.tech/tags/best.md>), [bill](<https://devfeed.tech/tags/bill.md>), [budget](<https://devfeed.tech/tags/budget.md>), [but](<https://devfeed.tech/tags/but.md>), [cascade](<https://devfeed.tech/tags/cascade.md>), [cases](<https://devfeed.tech/tags/cases.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [google](<https://devfeed.tech/tags/google.md>), [google-gemini](<https://devfeed.tech/tags/google-gemini.md>), [google-gemma](<https://devfeed.tech/tags/google-gemma.md>), [just](<https://devfeed.tech/tags/just.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llms](<https://devfeed.tech/tags/llms.md>), [open](<https://devfeed.tech/tags/open.md>)

### AI overview

Open LLMs are suited to routine, high-volume work as one layer in a routed AI system, but the evidence does not support replacing frontier models across the board. The article recommends handling repeat requests with caches or rules, routing routine tasks to small open models, reserving frontier models for hard reasoning, and setting token budgets. Reported routing research shows large potential cost reductions, but savings depend on workload and hosting utilization. Self-hosting shifts costs to compute, engineering, quality checks, latency tuning, and security review; cited break-even and performance figures are third-party estimates or benchmark results, not universal comparisons. The article advises testing models in shadow mode on an organization's own tasks, measuring cost per task and utilization, and starting with a narrow pilot. Its survey is small and mostly North American, and its energy-savings evidence is described as thin and largely modeled.

### Source excerpt

Geeta Kakrani Open models already run in production at 42% of enterprises, yet they take only about 5% of what those enterprises spend on AI. That gap, not model quality, is the real story. In Futurum's September 2026 survey, production users said open weights now carry about 34% of their AI tokens, up from 23% a year earlier. Google's Gemma runs in 55% of production accounts, up from 35%. But the money did not follow: the tokens move to open weights, and the budget reappears as servers, compute and engineering time. The survey is small (200 enterprises, 83 production users, mostly North America), so read it as a signal, not a verdict. So are open LLMs underrated, or just not ready? The short answer is neither. They are ready for the routine layer of an AI system, and they pay off only when they sit behind a router with a budget on top. This article covers what open models really cost (hosting, quality, latency, skills), and how to recover from each cost. The promise: stop paying frontier prices for routine work Most agent traffic is routine: tagging, extraction, formatting, simple lookups. Sending all of it to the biggest model is the main reason API bills climb. Open models attack exactly this part, and the research on routing supports the idea. FrugalGPT reports matching the best single LLM (for example GPT-4) with up to 98% lower cost, using prompt adaptation, model approximation and a cascade of models. RouteLLM reports cutting cost by more than 2x in some cases without compromising response quality. In the Futurum survey, low-complexity workloads grew from 39% to 51% of production users' open-weight usage, so enterprises are already using open models where it is safest. These are best-case, benchmark-style results. Your savings depend on how much of your traffic is truly routine, and on what the open model costs to run, which is the next section. The hidden bill: four costs the API price hides An open model has no per-token fee, but it is not free. The cost mo

## Google's first AI data center in space is up and running

DevFeed: [Google's first AI data center in space is up and running](<https://devfeed.tech/articles/google-s-first-ai-data-center-in-space-is-up-and-running-116361.md>)

Original publisher: [Read original article](<https://www.computerworld.com/article/4244229/googles-first-data-center-in-space-is-up-and-running.html>)

Author: Viktor Eriksson

Published: 2026-11-03T11:08:00Z

Content type: news

Language: en

Sources: [Computerworld](<https://devfeed.tech/sources/computerworld.md>)

Topics: [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [data centers](<https://devfeed.tech/topics/data-centers.md>), [gemma](<https://devfeed.tech/topics/gemma.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-chip](<https://devfeed.tech/tags/ai-chip.md>), [ai-chips](<https://devfeed.tech/tags/ai-chips.md>), [ai-data-center](<https://devfeed.tech/tags/ai-data-center.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [artificial-intelligence-data-center-emerging-technology-google-vendors-and-providers](<https://devfeed.tech/tags/artificial-intelligence-data-center-emerging-technology-google-vendors-and-providers.md>), [data-center](<https://devfeed.tech/tags/data-center.md>), [deepmind](<https://devfeed.tech/tags/deepmind.md>), [edition-us](<https://devfeed.tech/tags/edition-us.md>), [emerging-technology](<https://devfeed.tech/tags/emerging-technology.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [google](<https://devfeed.tech/tags/google.md>), [google-deepmind](<https://devfeed.tech/tags/google-deepmind.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [productivity](<https://devfeed.tech/tags/productivity.md>), [satellite](<https://devfeed.tech/tags/satellite.md>), [space](<https://devfeed.tech/tags/space.md>), [storytype-news-brief](<https://devfeed.tech/tags/storytype-news-brief.md>), [tensor-processing-units](<https://devfeed.tech/tags/tensor-processing-units.md>), [tpus](<https://devfeed.tech/tags/tpus.md>), [vendors-and-providers](<https://devfeed.tech/tags/vendors-and-providers.md>)

### AI overview

Google's first Suncatcher experimental data center satellite is operational and has begun communicating with Earth after launch and several weeks of testing. It carries four Tensor Processing Units (TPUs); its maximum one-kilowatt output can keep one AI chip active for about 20 minutes. In an initial test, Google DeepMind's Gemma 4 wrote a haiku about life in space, and the satellite reported orbital conditions. Google will next test the chip's response to space temperatures and radiation. If those tests go well, the company plans to launch two more satellites in 2027, toward a longer-term goal of a satellite network for advanced AI computing.

### Source excerpt

Google's first experimental data center satellite as part of the Suncatcher project is now operational. The satellite was launched in October, and after several weeks of testing, it has begun communicating with Earth. On board are four of Google's AI chips, Tensor Processing Units (TPUs), which are used to run AI models. The satellite can generate up to one kilowatt of electricity, which is enough to keep an AI chip active for about 20 minutes, TechCrunch reports. In the first test, the AI model Gemma 4 from Google DeepMind was tasked with writing a haiku about life in space. The satellite has also reported on conditions in its orbit. Google will now test how the chip handles temperatures and radiation in space. If the tests go well, Google plans to launch two more satellites in 2027. The long-term goal is to create a network of satellites capable of performing advanced AI computations. This article originally appeared on Computer Sweden.

## Geometry of Confabulation and Refusal in Small Language Models

DevFeed: [Geometry of Confabulation and Refusal in Small Language Models](<https://devfeed.tech/articles/geometry-of-confabulation-and-refusal-in-small-language-models-114516.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.01272>)

Author: Vladimir Kazei

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [llama](<https://devfeed.tech/topics/llama.md>), [Audit-then-score](<https://devfeed.tech/topics/audit-then-score.md>)

Tags: [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [hidden](<https://devfeed.tech/tags/hidden.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [linear](<https://devfeed.tech/tags/linear.md>), [llama](<https://devfeed.tech/tags/llama.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [models](<https://devfeed.tech/tags/models.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [optimized](<https://devfeed.tech/tags/optimized.md>), [portable](<https://devfeed.tech/tags/portable.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [signal](<https://devfeed.tech/tags/signal.md>), [small-language-models](<https://devfeed.tech/tags/small-language-models.md>), [speech](<https://devfeed.tech/tags/speech.md>), [state](<https://devfeed.tech/tags/state.md>), [vector](<https://devfeed.tech/tags/vector.md>)

### AI overview

ConfabQA, a 784-item benchmark designed to separate lack of knowledge from knowledge-cutoff effects, shows that hidden states before generation can help distinguish refusal from confabulation, with results varying by model. Probes improve over prompt-feature baselines on Qwen3-1.7B and Gemma 2 2B; the ConfabQA margin on Llama 3.2 3B is within fold-to-fold spread, though refusal probes on PopQA and TriviaQA improve by 18-25 percentage points. On Qwen, pushes along three substantially different candidate directions often change confabulations into refusal-style openings, but forcing the opening phrase alone achieves the same one-in-three judged-refusal rate. Opposite pushes mostly turn refusals into confabulations while leaving originally correct answers intact, suggesting these interventions affect the opening speech act rather than underlying knowledge. The abstract reports no concentrated confabulation feature or token signature on Qwen, and says correctness signal transfers across datasets only for Llama 3.2 3B.

### Source excerpt

arXiv:2611.01272v1 Announce Type: new Abstract: Faced with an unanswerable question, small language models can either refuse or confabulate: confidently produce a wrong answer. Can the hidden state before the first generated token provide a clue? Can we flip confabulations into refusals? I introduce ConfabQA, a 784-item benchmark whose obscure-pre-cutoff category breaks the typical correctness/knowledge-cutoff confound and map the geometry of the hidden state with linear probes, a sparse autoencoder (SAE), and direct optimization of intervention directions. On Qwen3-1.7B and Gemma 2 2B the probe separates upcoming refusals from confabulations well beyond prompt features (+7.4/+16.3 pp); on Llama 3.2 3B the ConfabQA margin is within fold-to-fold spread, but a direct refusal probe on PopQA and TriviaQA recovers +18-25 pp. On Qwen I closely inspect three candidate directions for the refusal vector: probe-recovered, Qwen-Scope SAE feature, and directly optimized. Although the three directions are pairwise 50-81 degrees apart, a push at or below the state's own norm along any of them flips most confabulations into refusal openings dominated by the "As..." family. Forcing the opener without touching the hidden state yields the same one-in-three judged-refusal rate as any push. The opposite push collapses refusals mostly into confabulations, yet leaves originally correct answers intact, so the direction gates the opening speech act rather than the knowledge behind it. Confabulations do not resolve into a concentrated feature in a dictionary or a token signature on Qwen3-1.7B. Scoring every probe on the margin over the strongest of four prompt-feature baselines extracts the hidden state's own contribution. Correctness signal is model-dependent and portable only for Llama 3.2 3B: 81% of the probe's margin over test majority survives transfer between ConfabQA, PopQA, and TriviaQA. Benchmark, judge, and code are released.

## Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA

DevFeed: [Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA](<https://devfeed.tech/articles/answer-correction-with-verifiable-rewards-a-gemma-family-study-of-multi-view-driving-qa-114455.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.00828>)

Author: T N Rickesh, Pranav Vinodh, Sowmya Kamath S

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [grpo](<https://devfeed.tech/topics/grpo.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [rlvr](<https://devfeed.tech/topics/rlvr.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [autonomous vehicles](<https://devfeed.tech/topics/autonomous-vehicles.md>)

Tags: [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [driving](<https://devfeed.tech/tags/driving.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [qa](<https://devfeed.tech/tags/qa.md>), [reward](<https://devfeed.tech/tags/reward.md>), [safety](<https://devfeed.tech/tags/safety.md>), [study](<https://devfeed.tech/tags/study.md>)

### AI overview

In the matched-harness Gemma-4-12B base/GRPO comparison, accuracy rose from 73.78% to 75.43% (1.65 percentage points; scene-cluster bootstrap interval +0.84 to +2.41). Of 215 improvements, 213 began with an extractable but incorrect base answer, while two recovered previously unextractable responses, supporting an answer-selection gain for this recorded tuning configuration. Category results show a safety tradeoff: routine perception improved by 5.93 points, while safety-critical hazard prediction declined by 1.82 points. The study cautions that unweighted outcome rewards can favor aggregate accuracy over rare edge cases. A separate archived Gemma-31B thinking comparison found no detected average improvement, but uncertainty does not establish equivalence. The results do not identify a specific grounding mechanism or establish superiority over matched supervised tuning.

### Source excerpt

arXiv:2611.00828v1 Announce Type: cross Abstract: Multi-camera driving question answering provides a discrete test of vision-language models, but post-training gains can reflect changed selected answers or improved answer extraction. We investigate this distinction within the Gemma model family using archived, task-linked WaymoQA evaluations. A zero-shot diagnostic on 5,519 test questions spans effective-small, mixture-of-experts, unified, and dense checkpoints, with serving differences disclosed rather than interpreted as architecture effects. The strongest controlled evidence is a matched-harness Gemma-4-12B base/GRPO pair: accuracy rises from 73.78% to 75.43%, a gain of 1.65 percentage points with a scene-cluster bootstrap interval of [+0.84, +2.41]. There are 215 improvements and 124 regressions; 213 improvements start from an extractable but incorrect base answer, and only two recover previously unextractable responses. Crossing correctness with extraction state shows why correct noncompliant outputs cannot be counted as prediction errors. Category counts reveal a critical safety asymmetry: routine perception gains (+5.93 pp) mask a regression in safety-critical hazard prediction (-1.82 pp), demonstrating that unweighted outcome rewards can trade off rare edge cases for aggregate accuracy. Reward saturation describes sampled training dynamics rather than an architectural ceiling. A separate archived Gemma-31B thinking comparison yields no detected average improvement, with uncertainty that does not establish equivalence. These results support a modest answer-selection gain for the recorded tuning configuration and a reproducible diagnostic accounting, without attributing corrected labels to specific grounding mechanisms or claiming superiority over matched supervised tuning.

## BYOD: Build Your Own DLM - Efficient Autoregressive-to-Diffusion Language Model Conversion

DevFeed: [BYOD: Build Your Own DLM - Efficient Autoregressive-to-Diffusion Language Model Conversion](<https://devfeed.tech/articles/byod-build-your-own-dlm-efficient-autoregressive-to-diffusion-language-model-conversion-111600.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.41574>)

Author: Ruurd J. A. Kuiper, Robert A. Bagheri

Published: 2026-11-02T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Language models](<https://devfeed.tech/topics/language-models.md>), [lora](<https://devfeed.tech/topics/lora.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [text-generation](<https://devfeed.tech/topics/text-generation.md>), [llama](<https://devfeed.tech/topics/llama.md>), [gemma](<https://devfeed.tech/topics/gemma.md>)

Tags: [adapters](<https://devfeed.tech/tags/adapters.md>), [attention](<https://devfeed.tech/tags/attention.md>), [autoregressive](<https://devfeed.tech/tags/autoregressive.md>), [conversion](<https://devfeed.tech/tags/conversion.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [generation](<https://devfeed.tech/tags/generation.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [lora](<https://devfeed.tech/tags/lora.md>)

### AI overview

BYOD converts autoregressive models into diffusion language models using LoRA adapters while keeping all base-model parameters frozen. The same recipe works on Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B; adapters are 4.1-5.8% of final parameters, and each 25,000-iteration run used at most 102.4M tokens and took 3.1-4.0 hours on one Google Colab GPU. The conversions generate from fully masked sequences, use bidirectional context in a controlled test, and retain substantial parent-model knowledge. BYOD-Gemma averaged 50.6% across nine tasks, versus 51.9% for LLaDA 8B Instruct in the same harness. Open-generation quality at low NFE remains the main limitation: for 128 tokens generated in 64 steps, BYOD perplexity was 8.3-11.1, compared with 4.1 for LLaDA and 2.2-3.0 for the autoregressive parents. The supplied abstract also reports a training and evaluation implementation and interactive demos.

### Source excerpt

arXiv:2610.41574v1 Announce Type: cross Abstract: Unlike autoregressive (AR) models, diffusion language models (DLMs) permit parallel token updates, flexible generation, and bidirectional attention, but leading systems rely on costly pretraining or full-model fine-tuning. We present Build Your Own DLM (BYOD), a generalizable LoRA-only recipe for parameter-efficient AR-to-DLM conversion. Without architecture-specific tuning, BYOD induces diffusion generation in Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B; to our knowledge, this is the first reported AR-to-DLM conversion of a Mistral model. The adapters comprise 4.1-5.8% of final parameters while every base-model parameter remains frozen; each 25,000-iteration conversion needs at most 102.4M tokens, processed in 3.1-4.0 hours on one Google Colab GPU. All four models generate from fully masked sequences, use bidirectional context in a controlled attention test, and retain substantial parent-model knowledge. BYOD-Gemma is the strongest BYOD conversion on our nine-task evaluation, reaching 50.6% average accuracy compared with 51.9% for LLaDA 8B Instruct in the same harness. Open-generation quality at low network function evaluation (NFE) budgets remains the principal limitation: BYOD perplexity is 8.3-11.1 versus 4.1 for LLaDA and 2.2-3.0 for the AR parents when generating 128 tokens in 64 steps. To enable further experimentation and inspection, we provide a standalone training and evaluation implementation and interactive demos for all four conversions. BYOD thus provides a practical route for extending diffusion generation to other existing model families while identifying the gap that efficient conversion must next close.

## How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum

DevFeed: [How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum](<https://devfeed.tech/articles/how-fast-should-a-model-commit-to-supervision-training-reasoning-models-on-the-tsallis-loss-continuum-111393.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2604.25907>)

Author: Chu-Cheng Lin, Eugene Ie

Published: 2026-11-02T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [rlvr](<https://devfeed.tech/topics/rlvr.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [machine learning overfitting](<https://devfeed.tech/topics/machine-learning-overfitting.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [bias](<https://devfeed.tech/tags/bias.md>), [computer](<https://devfeed.tech/tags/computer.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [flow](<https://devfeed.tech/tags/flow.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [monte-carlo](<https://devfeed.tech/tags/monte-carlo.md>), [noise](<https://devfeed.tech/tags/noise.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reasoning-models](<https://devfeed.tech/tags/reasoning-models.md>), [rlvr](<https://devfeed.tech/tags/rlvr.md>), [sft](<https://devfeed.tech/tags/sft.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

The paper proposes a Tsallis-loss continuum that interpolates between RLVR at q=0 and latent-trajectory log-marginal-likelihood training at q=1. Its gradient-flow analysis says the RLVR endpoint is robust to label noise but takes Ω(1/p₀) time to escape cold start; the density-estimation endpoint escapes in Θ(log(1/p₀)) time but can memorize noisy labels. Two Monte Carlo methods, Gradient-Amplified RL (GARL) and Posterior-Attenuated Fine-Tuning (PAFT), optimize fixed intermediate q without annotated rationales or external verifiers. They share a stated bias bound but differ in variance and stability. On FinQA, HotPotQA, and MuSiQue, the abstract reports that sufficiently high-q GARL escapes cold start where GRPO fails under Qwen 3 and Gemma 4.

### Source excerpt

arXiv:2604.25907v3 Announce Type: replace Abstract: The standard recipe for training reasoning models---SFT on rationale annotations followed by RLVR---relies on rationale annotations being available and clean. When no annotated rationales exist, SFT cannot start at all, and when the model is too unaligned, RLVR alone stalls at cold start; when annotations are weak, SFT memorizes the errors. Using the Tsallis $q$-logarithm, we define a loss family $J_Q$ that interpolates between RLVR (at $q{=}0$, the exploitation pole) and the log-marginal-likelihood over latent trajectories (at $q{=}1$, the density-estimation pole). The standard pipeline corresponds to a stepwise $q{=}1 \to q{=}0$ schedule, and intermediate $q$ enables training in regimes where the two-step pipeline cannot apply. All members share the same gradient direction, differing only by a per-instance amplification $P_{\theta}^{-q}$ that reweights each instance independently of the learning rate. Under gradient flow analysis, we show that the exploitation pole requires $\Omega(\frac{1}{p_0})$ time to escape cold start but is robust to label noise, while the density-estimation pole escapes in $\Theta\big(\log(\frac{1}{p_0})\big)$ but memorizes label noise. This separation explains why SFT ($q{=}1$) precedes RLVR ($q{=}0$) in the standard pipeline. We further derive two Monte Carlo estimators that directly optimize fixed-$q$ on the $J_Q$ continuum, without annotated rationales or external verifiers: Gradient-Amplified RL (GARL) and Posterior-Attenuated Fine-Tuning (PAFT), with shared bias $O\big(\frac{q}{M P_{\theta}^q}\big)$ but different variance and stability properties. On FinQA, HotPotQA, and MuSiQue, GARL at sufficiently high $q$ escapes cold start where GRPO fails entirely under Qwen 3 and Gemma 4 architectures.

## Google's First Data Center Satellite Is Online and Writing Poetry

DevFeed: [Google's First Data Center Satellite Is Online and Writing Poetry](<https://devfeed.tech/articles/google-s-first-data-center-satellite-is-online-and-writing-poetry-109672.md>)

Original publisher: [Read original article](<https://science.slashdot.org/story/26/10/30/2344228/googles-first-data-center-satellite-is-online-and-writing-poetry>)

Author: BeauHD

Published: 2026-10-31T07:00:00Z

Content type: news

Language: en

Sources: [Slashdot](<https://devfeed.tech/sources/slashdot.md>)

Topics: [gemma](<https://devfeed.tech/topics/gemma.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [communications](<https://devfeed.tech/topics/communications.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [communications](<https://devfeed.tech/tags/communications.md>), [data-center](<https://devfeed.tech/tags/data-center.md>), [deepmind](<https://devfeed.tech/tags/deepmind.md>), [google](<https://devfeed.tech/tags/google.md>), [haiku](<https://devfeed.tech/tags/haiku.md>), [inference](<https://devfeed.tech/tags/inference.md>), [orbit](<https://devfeed.tech/tags/orbit.md>), [poetry](<https://devfeed.tech/tags/poetry.md>), [power](<https://devfeed.tech/tags/power.md>), [satellite](<https://devfeed.tech/tags/satellite.md>), [space](<https://devfeed.tech/tags/space.md>), [stress](<https://devfeed.tech/tags/stress.md>), [tensor-processing-units](<https://devfeed.tech/tags/tensor-processing-units.md>), [tpu](<https://devfeed.tech/tags/tpu.md>), [tpus](<https://devfeed.tech/tags/tpus.md>)

### AI overview

Google's experimental satellite data center has run Google DeepMind's Gemma 4 in orbit, producing a haiku from a prompt sent by engineers on the ground. The spacecraft carries four Tensor Processing Units and generates up to 1 kW, but power limits allow only one TPU to run for about 20 minutes at a time, complicating communication with the ground. Testing will assess the chips under thermal stress and space radiation. If they prove capable, Google plans two more satellites in 2027, with higher power generation and inter-satellite optical links.

### Source excerpt

Google's first experimental space-based data center is operational after launching earlier this month, successfully running its Gemma 4 AI model aboard a satellite and generating a haiku about being in orbit. TechCrunch reports: The satellite, which carries four Google Tensor Processing Units (TPUs) and generates up to 1 kW of electricity, is intended to pave the way for future networks of close-flying satellites performing large inference workloads in orbit. However, the power demands of the TPU mean it can only turn one on for about 20 minutes at a time, complicating the already tricky business of space-to-ground communications. After a commissioning process that tested out the satellite's subsystems over about three weeks, the spacecraft was ready to work. When engineers on the ground sent their first prompt this week, they asked the model pre-loaded onto the satellite, Google DeepMind's Gemma 4, to write a haiku about being in space. And thus: Blue marble below Data streams in silent grace Starlight feeds my core. The ongoing test campaign will examine how the chips perform under thermal stress and in the space radiation environment. [...] If the TPUs onboard this spacecraft prove capable, the next step will be launching two more satellites equipped with the advanced chips in 2027. Those spacecraft will support higher power generation and demonstrate inter-satellite optical links. Read more of this story at Slashdot.

## Google's first data center satellite is online and writing poetry

DevFeed: [Google's first data center satellite is online and writing poetry](<https://devfeed.tech/articles/google-s-first-data-center-satellite-is-online-and-writing-poetry-109113.md>)

Original publisher: [Read original article](<https://techcrunch.com/2026/10/30/googles-first-data-center-satellite-is-online-and-writing-poetry/>)

Author: Tim Fernholz

Published: 2026-10-30T17:00:00Z

Content type: news

Language: en

Sources: [TechCrunch | AI News & Artificial Intelligence](<https://devfeed.tech/sources/techcrunch-ai-news-artificial-intelligence.md>)

Topics: [Google](<https://devfeed.tech/topics/google.md>), [gemma](<https://devfeed.tech/topics/gemma.md>)

Tags: [about](<https://devfeed.tech/tags/about.md>), [accelerators](<https://devfeed.tech/tags/accelerators.md>), [advanced](<https://devfeed.tech/tags/advanced.md>), [ai](<https://devfeed.tech/tags/ai.md>), [data](<https://devfeed.tech/tags/data.md>), [data-center](<https://devfeed.tech/tags/data-center.md>), [deepmind](<https://devfeed.tech/tags/deepmind.md>), [exclusive](<https://devfeed.tech/tags/exclusive.md>), [first](<https://devfeed.tech/tags/first.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [google](<https://devfeed.tech/tags/google.md>), [google-deepmind](<https://devfeed.tech/tags/google-deepmind.md>), [google-s](<https://devfeed.tech/tags/google-s.md>), [model](<https://devfeed.tech/tags/model.md>), [online](<https://devfeed.tech/tags/online.md>), [poetry](<https://devfeed.tech/tags/poetry.md>), [running](<https://devfeed.tech/tags/running.md>), [satellite](<https://devfeed.tech/tags/satellite.md>), [space](<https://devfeed.tech/tags/space.md>), [space-data-centers](<https://devfeed.tech/tags/space-data-centers.md>), [tpu](<https://devfeed.tech/tags/tpu.md>), [writing](<https://devfeed.tech/tags/writing.md>)

### AI overview

Google's Suncatcher prototype is online: its single TPU has run Google DeepMind's Gemma 4 model, answered a prompt with a haiku, and reported orbital conditions. The satellite can generate up to 1 kW, but its power demands limit the TPU to about 20 minutes of operation at a time, complicating ground communications. The current test campaign will measure chip performance under thermal stress and space radiation, including radiation-related bit flips that can corrupt data. Earlier ground tests suggested radiation could affect training more than inference, but the in-orbit tests will check that estimate in practice. If the TPU performs well, Google plans to launch two more chip-equipped satellites in 2027 to test higher power generation and optical links between spacecraft.

### Source excerpt

The first TPU in space is a running a Google DeepMind Gemma 4 model and writing cheesy poetry.

[Next page](<https://devfeed.tech/topics/gemma.md?cursor=WyIyMDI2LTEwLTMwVDE3OjAwOjAwKzAwOjAwIiwgImU2MDk4NjJiLTNkNjktNDExYi04MjRiLWIzNjY2YTkxNzgzMyJd>)