# qwen

Qwen is a series of large language and large multimodal models from the Qwen Team at Alibaba Group.

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

## Your AI red team may be finding the same prompt injection thousands of times

DevFeed: [Your AI red team may be finding the same prompt injection thousands of times](<https://devfeed.tech/articles/your-ai-red-team-may-be-finding-the-same-prompt-injection-thousands-of-times-145725.md>)

Original publisher: [Read original article](<https://www.helpnetsecurity.com/2026/11/20/ai-prompt-injection-red-team/>)

Author: Mirko Zorz

Published: 2026-11-20T06:30:20Z

Content type: news

Language: en

Sources: [Help Net Security](<https://devfeed.tech/sources/help-net-security.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-red-team](<https://devfeed.tech/tags/ai-red-team.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [automated](<https://devfeed.tech/tags/automated.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [defense](<https://devfeed.tech/tags/defense.md>), [don-t-miss](<https://devfeed.tech/tags/don-t-miss.md>), [features](<https://devfeed.tech/tags/features.md>), [hot-stuff](<https://devfeed.tech/tags/hot-stuff.md>), [llms](<https://devfeed.tech/tags/llms.md>), [news](<https://devfeed.tech/tags/news.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [prompt-injection-attacks](<https://devfeed.tech/tags/prompt-injection-attacks.md>), [red-team](<https://devfeed.tech/tags/red-team.md>), [red-teaming](<https://devfeed.tech/tags/red-teaming.md>), [research](<https://devfeed.tech/tags/research.md>), [team](<https://devfeed.tech/tags/team.md>), [workflow](<https://devfeed.tech/tags/workflow.md>)

### AI overview

In three of four setups, retraining on more successful attacks from automated red-team tools left agents easier to hijack. The researchers found that the tools often repeated a few techniques with altered wording, so attack counts overstated coverage. Their PIPETTE system rewards attacks using new techniques and trains agents to check whether each action serves the user's request. On AgentDojo, standard injection attacks succeeded in under 0.5% of attempts on average against PIPETTE-trained small models; adaptive attackers reached up to 10% on the 40 selected weak-spot tasks. The defense also preserved task completion: hardened Qwen3-4B completed slightly more clean tasks than the original, while PromptGuard blocked somewhat more under one attack but completed the user's task only 16.19% of the time, versus 58.10% for PIPETTE. Results are limited to open-weight models and small task sets; a larger attacker budget may do better, and the researchers have not verified their technique-coverage measure against human judgment or established that narrow coverage alone explains the gap.

### Source excerpt

AI agents retrained on a bigger pile of prompt injection attacks came out easier to hijack in three of four tests run by researchers at the University of Waterloo, Zhejiang University and KTH Royal Institute of Technology. The automated red-teaming tool that found more working attacks produced the weaker defense. Prompt injection purple-teaming workflow (Source: Research paper) Someone hides instructions in an email, a web page or a tool's output, and the AI agent that ... More -> The post Your AI red team may be finding the same prompt injection thousands of times appeared first on Help Net Security.

## Can Computation from Earlier Problems Help LLMs Solve New Ones?

DevFeed: [Can Computation from Earlier Problems Help LLMs Solve New Ones?](<https://devfeed.tech/articles/can-computation-from-earlier-problems-help-llms-solve-new-ones-144949.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.39394>)

Author: Jipei He, Wenhui Tan, Xiaoyi Yu, Enver Sangineto, Fiorenzo Parascandolo, Rita Cucchiara, Ruihua Song

Published: 2026-11-20T05:00:00Z

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>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [code](<https://devfeed.tech/tags/code.md>), [code-generation](<https://devfeed.tech/tags/code-generation.md>), [controller](<https://devfeed.tech/tags/controller.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>), [domain](<https://devfeed.tech/tags/domain.md>), [effect](<https://devfeed.tech/tags/effect.md>), [entries](<https://devfeed.tech/tags/entries.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [generation](<https://devfeed.tech/tags/generation.md>), [help](<https://devfeed.tech/tags/help.md>), [history](<https://devfeed.tech/tags/history.md>), [independent](<https://devfeed.tech/tags/independent.md>), [keys](<https://devfeed.tech/tags/keys.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [llms](<https://devfeed.tech/tags/llms.md>), [problems](<https://devfeed.tech/tags/problems.md>), [solve](<https://devfeed.tech/tags/solve.md>)

### AI overview

Retained conversation history can either help or hurt later-turn accuracy, and the abstract reports recurring, problem-dependent changes in model internal states when preceding history varies. The proposed STAIR method stores attention keys and values from earlier responses in a read-only bank, then uses a learned reflection of current query vectors to redirect attention to that bank during prompt processing. The backbone and stored entries remain fixed; training updates only 12,288 parameters. Across three Qwen models and four reasoning benchmarks, STAIR raises average later-turn accuracy by up to 11.67 percentage points versus the unmodified model with history. A math-trained Qwen3.5-9B controller also improves HumanEval+ code generation without further training. These are abstract-level claims; detailed benchmark results are not supplied.

### Source excerpt

arXiv:2609.39394v2 Announce Type: replace-cross Abstract: Large language models often solve independent problems in the same conversation. Can computation from earlier problems help them solve new ones? Preliminary experiments show that retained history can raise or lower later-turn accuracy, even within the same domain. To trace this effect, we hold each problem fixed and vary its preceding history. The resulting changes in the model's internal states contain a recurring, problem-dependent component. To improve reasoning with retained history, we introduce STAIR (Stale-Token Attention for Inter-query Reuse). It captures attention keys and values from earlier response generation in a read-only bank. STAIR learns to reflect current query vectors to redirect attention over this bank during prompt processing. The backbone and stored entries remain fixed; only 12,288 parameters are trained. Across three Qwen models and four reasoning benchmarks, STAIR improves average later-turn accuracy by up to 11.67 percentage points over the unmodified model with history. The math-trained Qwen3.5-9B controller also improves code generation on HumanEval+ without further training.

## LEDGE: An Efficient Simulator for Distributed LLM Serving on Heterogeneous Edge Devices

DevFeed: [LEDGE: An Efficient Simulator for Distributed LLM Serving on Heterogeneous Edge Devices](<https://devfeed.tech/articles/ledge-an-efficient-simulator-for-distributed-llm-serving-on-heterogeneous-edge-devices-142754.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24085>)

Author: Zihao Wang, Ye Tian, Tajana Rosing, Nalini Venkatasubramanian

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

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

Topics: [LLM Techniques](<https://devfeed.tech/topics/llm-techniques.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [resource orchestration](<https://devfeed.tech/topics/resource-orchestration.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [distributed](<https://devfeed.tech/tags/distributed.md>), [edge](<https://devfeed.tech/tags/edge.md>), [edge-devices](<https://devfeed.tech/tags/edge-devices.md>), [jetson](<https://devfeed.tech/tags/jetson.md>), [latency](<https://devfeed.tech/tags/latency.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-serving](<https://devfeed.tech/tags/llm-serving.md>), [pipeline-parallelism](<https://devfeed.tech/tags/pipeline-parallelism.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [raspberry](<https://devfeed.tech/tags/raspberry.md>), [scaling](<https://devfeed.tech/tags/scaling.md>), [simulation](<https://devfeed.tech/tags/simulation.md>), [simulator](<https://devfeed.tech/tags/simulator.md>), [source-code](<https://devfeed.tech/tags/source-code.md>)

### AI overview

LEDGE simulates distributed LLM serving on heterogeneous edge devices without executing tensor operations. It builds event graphs for non-uniform tensor and pipeline parallelism, estimates execution costs with reusable performance models calibrated from end-to-end measurements, and models shared-link contention. On a 10-device Jetson and Raspberry Pi testbed, it covered 17 model-precision combinations across four LLM families. For Qwen deployments, request-level MAPE was below 3.6%, with 135.3x aggregate speedup over physical inference. Models calibrated on one family achieved 7.44% zero-shot end-to-end latency MAPE across three others. The authors also report simulations showing effects of shared-link contention on placement and network topology on scaling; 26,773 runs across four edge environments took 83.8 minutes.

### Source excerpt

arXiv:2611.24085v1 Announce Type: cross Abstract: Evaluating distributed large language model (LLM) serving on heterogeneous edge devices requires exploring diverse model partitions, device placements, and network conditions, making repeated physical experiments costly. An efficient simulator for edge LLM serving must support diverse, fine-grained placement configurations, capture runtime behavior, and accurately predict execution times while keeping simulation overhead low. We present LEDGE, a lightweight simulator for fast, fine-grained evaluation of distributed LLM serving on heterogeneous edge devices without executing tensor operations. LEDGE constructs event graphs supporting non-uniform tensor and pipeline parallelism, estimates execution costs using reusable performance models calibrated from end-to-end measurements, and simulates shared-link contention. We evaluate LEDGE on a heterogeneous testbed of 10 Jetson and Raspberry Pi devices, covering 17 model--precision combinations across 4 LLM families. On Qwen deployments, LEDGE achieves request-level mean absolute percentage errors below 3.6% and a 135.3 times aggregate speedup over physical inference. Performance model calibrated on one LLM family yield 7.44% zero-shot end-to-end latency MAPE across three other families. Further simulations show how shared-link contention affects placement decisions and network topology constrains scaling. Across four edge environments, LEDGE completes 26,773 simulation runs in 83.8 minutes, enabling efficient exploration of deployment strategies under diverse resource conditions and serving objectives. The source code is publicly available at https://github.com/gdfwj/LEDGE.

## Assessing Vision-Language Models for Early Prediction of Unsafe Situations in Robotics

DevFeed: [Assessing Vision-Language Models for Early Prediction of Unsafe Situations in Robotics](<https://devfeed.tech/articles/assessing-vision-language-models-for-early-prediction-of-unsafe-situations-in-robotics-142205.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24146>)

Author: Muhammad Farooq, Aitor Arrieta, Paolo Arcaini, Shaukat Ali, Thomas Peyrucain

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [Monitoring & Alerting](<https://devfeed.tech/topics/monitoring-alerting.md>)

Tags: [classification](<https://devfeed.tech/tags/classification.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [early-warning](<https://devfeed.tech/tags/early-warning.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [history](<https://devfeed.tech/tags/history.md>), [llava](<https://devfeed.tech/tags/llava.md>), [monitoring](<https://devfeed.tech/tags/monitoring.md>), [performance](<https://devfeed.tech/tags/performance.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [robot-safety](<https://devfeed.tech/tags/robot-safety.md>), [robotics](<https://devfeed.tech/tags/robotics.md>), [support](<https://devfeed.tech/tags/support.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [video](<https://devfeed.tech/tags/video.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>), [vision-language-models-vlms](<https://devfeed.tech/tags/vision-language-models-vlms.md>)

### AI overview

The abstract reports an empirical evaluation of Qwen, InternVL, and LLaVA-OneVision for early robot-safety monitoring across HuRoN, Pal Robotics, and CrowdBot, using video and odometry to classify situations as safe, potentially unsafe, or unsafe and assess uncertainty. Qwen had the strongest overall classification and prediction-reliability performance; InternVL showed stronger early-warning capability in several configurations. The effect of robot-state information and temporal history varied by model and evaluation objective, so the preferred configuration depends on whether accurate safety-state classification, reliable uncertainty estimates, or timely warning is the priority.

### Source excerpt

arXiv:2611.24146v1 Announce Type: cross Abstract: Vision-Language Models (VLMs) process visual information and textual instructions and, by using their ability to understand and reason about context, are well-suited for predicting robot failures before they occur. However, evidence supporting this remains limited. To address this, we present an empirical evaluation of VLMs for predicting robot failures. In particular, we experiment with Qwen, InternVL, and LLaVA-OneVision across three robotic case studies (HuRoN, Pal Robotics, and CrowdBot), analyzing both video and odometry data for early failure warning (i.e., before an unsafe situation occurs). Specifically, we evaluate their performance in classifying situations as safe, potentially unsafe, and unsafe, together with uncertainty quantified using various metrics, and study the relationship between uncertainty and performance. Our results showed that Qwen provides the strongest overall classification and prediction-reliability performance, while InternVL shows stronger early-warning capability in several configurations. The results further show that the effect of robot-state information and temporal history depends on the model and evaluation objective. Overall, our findings indicate that general-purpose VLMs can support early robot-safety monitoring, but the most suitable configuration depends on whether the priority is accurate assessment of safety-state, reliable confidence and uncertainty, or timely warning before unsafe events.

## What Does a Bias Head Measure? Non-Identifiability in Head-Level Bias Attribution

DevFeed: [What Does a Bias Head Measure? Non-Identifiability in Head-Level Bias Attribution](<https://devfeed.tech/articles/what-does-a-bias-head-measure-non-identifiability-in-head-level-bias-attribution-141863.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23218>)

Author: S. M Seefat Alam

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

Content type: article

Language: en

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

Topics: [Chinchilla scaling law](<https://devfeed.tech/topics/chinchilla-scaling-law.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [llama](<https://devfeed.tech/topics/llama.md>)

Tags: [ablation](<https://devfeed.tech/tags/ablation.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attention](<https://devfeed.tech/tags/attention.md>), [attribution](<https://devfeed.tech/tags/attribution.md>), [bias](<https://devfeed.tech/tags/bias.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [llama-3-1](<https://devfeed.tech/tags/llama-3-1.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [qwen](<https://devfeed.tech/tags/qwen.md>)

### AI overview

On Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct, the audited head-level bias map could not be identified as measuring bias: although it strongly correlated with a label-blind response-length map, breaking the length-label coupling showed that length explained only 1% of the statistic's variance, while prompt identity explained 64%. Within a fixed prompt, fairness label, response length and protected-group-token presence were highly collinear, leaving observational attribution unresolved. The authors also report that bias selectivity was not concentrated in late layers and that none of the tested head sets reduced unfairness when ablated. Ablating the complementary set instead reliably damaged refusal behaviour, suggesting the pipeline detects effects but not specifically bias.

### Source excerpt

arXiv:2611.23218v1 Announce Type: cross Abstract: Methods that localise social bias in large language models to individual attention heads and remove it by ablation are validated by their effect on a fairness metric, never by evidence that the statistic selecting the heads measures bias. We audit this on Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct. The head map correlates at $\rho = -0.950$ and $-0.768$ with a label-blind response-length map, apparently identifying a surface confound. An intervention that breaks the length-label coupling rejects this reading: with the prompt held fixed, response length explains 1% of the statistic's variance. What does explain it is prompt identity, at 64% of variance, in an ordering that is not monotonic in length. Within any fixed prompt the fairness label, response length and the presence of a protected-group token are collinear at $\phi = 0.95$, partly by definition of the labelling rule, so no observational analysis can attribute the map to any of them, and no method in this family performs an experiment that could. The substantive claims fail independently: bias selectivity is not late-layer-concentrated on either model under either reference reading, and no head set we tested reduces the unfairness rate when ablated. Its complement, scored on the opposite label, reliably destroys refusal behaviour ($p = 2.7\times10^{-5}$), demonstrating that the pipeline detects effects at this 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.

## Probing the Trajectories of Reasoning Traces in Large Language Models

DevFeed: [Probing the Trajectories of Reasoning Traces in Large Language Models](<https://devfeed.tech/articles/probing-the-trajectories-of-reasoning-traces-in-large-language-models-138736.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2601.23163>)

Author: Marthe Ballon, Brecht F. Verbeken, Vincent Ginis, Andres Algaba

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

Content type: article

Language: en

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

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Chain-of-thought](<https://devfeed.tech/topics/chain-of-thought.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [gpt-oss](<https://devfeed.tech/topics/gpt-oss.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>)

Tags: [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [deepseek-r1](<https://devfeed.tech/tags/deepseek-r1.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>), [gpt-oss](<https://devfeed.tech/tags/gpt-oss.md>), [gpt-oss-20b](<https://devfeed.tech/tags/gpt-oss-20b.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [protocol](<https://devfeed.tech/tags/protocol.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reasoning-model](<https://devfeed.tech/tags/reasoning-model.md>), [traces](<https://devfeed.tech/tags/traces.md>)

### AI overview

The proposed trajectory-probing protocol measures how a model's answer distribution changes as it receives progressively longer portions of its own reasoning trace: generate a full trace, truncate it at token percentiles, then feed each partial trace back and measure the induced answers. Across seven open-source models and four benchmarks, accuracy and decision commitment generally rise with the amount of reasoning provided. Length, form, and token-identity controls indicate that the gains come from instance-specific semantic content, not simply longer context or generic reasoning style. In cross-model probes, even stronger models often reproduce errors in a source model's partial trace, while allowing additional reasoning substantially improves recovery. The abstract presents the method as model-agnostic and capable of exposing failure modes that aggregate accuracy can miss.

### Source excerpt

arXiv:2601.23163v2 Announce Type: replace Abstract: Large language models (LLMs) increasingly solve difficult problems by producing reasoning traces before emitting a final response. However, it remains unclear how accuracy and decision commitment evolve along a reasoning trajectory, and whether intermediate trace segments provide answer-relevant information beyond generic length or stylistic effects. Here, we propose a trajectory-probing protocol for evaluating how a reasoning model's predicted answer evolves along a partial reasoning trace. The protocol (1) generates a model's full reasoning trace, (2) truncates it at fixed token-percentiles, and (3) injects each partial trace back into the model, measuring the model's induced answer distribution. We apply the protocol to seven open-source reasoning models (Qwen3-4B/8B/14B, gpt-oss-20b/120b, DeepSeek-R1-Distill-Llama-8B and Gemma-4-31B-it) across four benchmarks (GPQA Diamond, MMLU-Pro, Omni-MATH-2 and CompMath). The protocol reveals that accuracy and decision commitment generally increase as the percentage of provided reasoning tokens grows, and length, form, and token-identity controls confirm that gains stem from instance-specific semantic content rather than context length or generic reasoning style effects. Furthermore, cross-model probing shows that even stronger models frequently repeat the source model's errors when given its partial reasoning trace, while allowing further reasoning substantially improves recovery. Together, these results provide a reproducible, model-agnostic measurement of reasoning trace dynamics that generalizes across answer formats and can reveal model-specific failure modes invisible to aggregate accuracy.

## The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models

DevFeed: [The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models](<https://devfeed.tech/articles/the-granularity-axis-a-micro-to-macro-latent-direction-for-social-roles-in-language-models-137249.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.06196>)

Author: Chonghan Qin, Xiachong Feng, Ziyun Song, Xiaocheng Feng, Jing Xiong, Lingpeng Kong

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

Content type: article

Language: en

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

Topics: [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [llama](<https://devfeed.tech/topics/llama.md>)

Tags: [axis](<https://devfeed.tech/tags/axis.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [social](<https://devfeed.tech/tags/social.md>), [steering](<https://devfeed.tech/tags/steering.md>)

### AI overview

The study reports a continuous hidden-state direction that tracks social-role scale, from individual roles to institutions. Defined as the difference between mean macro- and micro-role states, the axis closely matches the leading role-space component in Qwen3-8B (cosine 0.972; PC1 explains 52.6% of variance). Across 75 roles at five levels, projections rose monotonically and remained stable across layers, prompts, and response-quality filters. The pattern transferred to Llama-3.1-8B-Instruct and larger Qwen3 and Llama models. Rater validation and controls for wording, affect, and formality supported scale-related variation. Positive steering shifted Llama's macro-scale score from 2.00 to 3.17 on prompts allowing local responses, though steering strength and stability depended on the model's default operating regime.

### Source excerpt

arXiv:2605.06196v2 Announce Type: replace Abstract: Large language models (LLMs) are prompted to adopt social roles from individuals to institutions, but whether their representations encode this difference in social granularity remains unclear. We discover a contrast-based Granularity Axis, the difference between mean macro- and micro-role hidden states, that aligns with the dominant geometry of role space: in Qwen3-8B, its cosine with PC1 is $0.972$, and PC1 explains $52.6\%$ of the variance. Across 75 roles at five levels, with 90,000 role-conditioned and 1,200 default responses per model, projections increase monotonically and remain stable across layers, prompts, and response-quality filters. The structure transfers to Llama-3.1-8B-Instruct and scales to Qwen3-32B and Llama-3.1-70B-Instruct, whose PC1 shares are $45.99\%$ and $31.23\%$. Independent five-rater validation ($\alpha=0.961$) preserves the ordering; identical continuations and affect/formality controls confirm scale-related variation beyond generated wording. The axis is also behaviorally causal: positive steering moves Llama from $2.00$ to $3.17$ on a five-point macro scale for prompts admitting local responses. Steering strength and stability depend on the model's default operating regime. Together, these findings establish social granularity as a representational primitive: a continuous, ordered, causally manipulable direction organizing role-conditioned generation across model families.

## 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.

## Do Pretraining Optimizer Gains Transfer to KL Distillation?

DevFeed: [Do Pretraining Optimizer Gains Transfer to KL Distillation?](<https://devfeed.tech/articles/do-pretraining-optimizer-gains-transfer-to-kl-distillation-137446.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19419>)

Author: Lavinia Ghita, Dhruv Desai

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

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: [Model Optimizer](<https://devfeed.tech/topics/model-optimizer.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [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>), [distillation](<https://devfeed.tech/tags/distillation.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [muon](<https://devfeed.tech/tags/muon.md>), [overhead](<https://devfeed.tech/tags/overhead.md>), [pretraining](<https://devfeed.tech/tags/pretraining.md>), [quality](<https://devfeed.tech/tags/quality.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [training](<https://devfeed.tech/tags/training.md>), [validation](<https://devfeed.tech/tags/validation.md>)

### AI overview

Across four dense Qwen student-teacher pairs trained with a shared forward-KL objective and fixed budget, the evaluated matrix-structured optimizers did not consistently beat AdamW and diagonal AdaGrad on validation quality versus trainer time. AdamW generally achieved the lowest validation KL; AdaGrad could be slightly faster at larger scale. SOAP often matched AdamW's KL but added substantial trainer-time overhead, while the two Shampoo reductions changed relative ordering across model pairs. The findings indicate that optimizer gains seen in pretraining do not automatically carry over to short-horizon KL distillation and need to be reassessed for the target objective and systems setup.

### Source excerpt

arXiv:2611.19419v1 Announce Type: cross Abstract: Matrix-structured optimizers can improve language-model pretraining, but it is unclear whether these gains transfer to short-horizon KL distillation from an instruction-tuned initialization. We compare AdamW and diagonal AdaGrad with two tractable Shampoo reductions, SOAP, and All-2D Muon across four dense Qwen student-teacher pairs under a common forward-KL objective and fixed training budget. Across scales, none of the evaluated matrix-structured recipes consistently improves the quality versus trainer-time trade-off over the elementwise references: AdamW generally attains the lowest validation KL, while AdaGrad can be slightly faster at larger scale. SOAP often tracks AdamW in KL but incurs substantial trainer-time overhead, and the two Shampoo reductions reverse ordering across model pairs, showing that conclusions depend on the feasible recipe. These results suggest that pretraining optimizer gains do not transfer automatically to KL distillation and should be re-established for the target objective and systems regime.

## LABEL: Literature Asymmetry between Efficacy and Limits in Drug Knowledge and Language Models

DevFeed: [LABEL: Literature Asymmetry between Efficacy and Limits in Drug Knowledge and Language Models](<https://devfeed.tech/articles/label-literature-asymmetry-between-efficacy-and-limits-in-drug-knowledge-and-language-models-137409.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19166>)

Author: S M Rayeed, Charles V. Stewart, Yao Ma

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

Content type: news

Language: en

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

Topics: [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [clinical-ai](<https://devfeed.tech/tags/clinical-ai.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ir](<https://devfeed.tech/tags/cs-ir.md>), [drugs](<https://devfeed.tech/tags/drugs.md>), [language-models](<https://devfeed.tech/tags/language-models.md>)

### AI overview

Across 574 drugs, the study finds that Wikipedia articles and PubMed abstracts mention indications about twice as often as contraindications (within-drug odds ratio 2.19). Qwen2.5 models from 0.5B to 14B improve by 18.3 points on indication accuracy as they scale, while contraindication accuracy stays near random. Yet models use safety facts as effectively as benefits when given a passage containing the answer, suggesting the gap reflects missing safety information in general prose. Routing safety questions to official FDA labels raises accuracy by 15.8 points when the label states the fact. The authors recommend using regulatory labels and structured databases for medication-safety questions; these results are reported for the studied sources and setup.

### Source excerpt

arXiv:2611.19166v1 Announce Type: new Abstract: Every medication comes with two types of medical facts: indications, which describe what conditions the drug treats (benefits), and contraindications, which describe who must avoid it to prevent serious harm (safety limits). Although contraindications are critical for patient safety and FDA regulations require listing them as rigorously as indications, we find that contraindications are far less documented than indications in general prose. In this paper, we measure how often these texts contain both facts across 574 drugs from PrimeKG by comparing a drug's indications against its own contraindications. Across Wikipedia articles and PubMed abstracts, we find that reference prose names a drug's benefits about twice as often as its safety limits (within-drug odds ratio 2.19). We call this documentation gap LABEL (Literature Asymmetry between Efficacy and Limits). Language models mirror this gap. As we scale Qwen2.5 from 0.5B to 14B, accuracy on indications rises by 18.3 points, whereas accuracy on contraindications remains flat near random guessing. However, when we provide models with a retrieved passage that states the answer, they use safety facts just as effectively as benefits. Furthermore, routing safety questions to official FDA labels boosts accuracy by 15.8 points whenever the label states the fact. Overall, our findings show that models fail on medication safety not because they cannot reason about risk, but because free-text corpora lack safety facts. Clinical AI systems should therefore route medication-safety questions to official regulatory labels and structured databases rather than relying on general public records. Code and data are available at https://github.com/SMRayeed/LABEL.

## Before the Word: Name Information in Description-Derived States of Small Language Models, With and Without Naming Failure

DevFeed: [Before the Word: Name Information in Description-Derived States of Small Language Models, With and Without Naming Failure](<https://devfeed.tech/articles/before-the-word-name-information-in-description-derived-states-of-small-language-models-with-and-without-naming-failure-138897.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19609>)

Author: Fernando Cardenas Piepereit

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

Content type: article

Language: en

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

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [failure](<https://devfeed.tech/tags/failure.md>), [information](<https://devfeed.tech/tags/information.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [small-language-models](<https://devfeed.tech/tags/small-language-models.md>), [state](<https://devfeed.tech/tags/state.md>)

### AI overview

Across Qwen2-0.5B, SmolLM2-1.7B and Phi-3-mini, similarity read-outs and multi-position activation patching supported name-related information in states derived from name-free descriptions; Patchscopes stayed near chance. Robustness checks qualify the result: single-position patching at 50% depth showed weak signals. Whether these states are less consistent across paraphrases when naming fails was mixed: SmolLM2's preregistered deficit was not robust, Phi-3 met non-inferiority criteria although its bootstrap interval included zero, and Qwen2-0.5B was inconclusive. The study used 200 items and was internally preregistered without an external registry. Its TOT framing is an analogy, and reported signals concern information-carrying vehicles, not meaning; a capacity explanation was not tested.

### Source excerpt

arXiv:2611.19609v1 Announce Type: new Abstract: The tip-of-the-tongue (TOT) state, knowing an object but not its name, motivates this study as an analogy only. We ask whether small language models form a name-predictive state from a name-free description, and whether it differs when naming fails but description selection succeeds. In an internally pre-registered study (a frozen document, no external registry) of Qwen2-0.5B, SmolLM2-1.7B and Phi-3-mini with 200 items, we combine a similarity read-out, a Patchscopes-style read-out and activation patching with matched controls. H1 (name information in the state) was supported in all nine model-by-description-type combinations; the agreement rule was met by the similarity measure together with patching, while Patchscopes stayed near chance. Post-hoc robustness checks after internal review (logit lens, single-position patch, manipulation check) temper this reading: name information in middle layers is clearest under the multi-position patch and the similarity measure, and weak under single-position patching at 50% depth. H2' (description-paraphrase consistency of name-free states not lower in failure items) is mixed: the SmolLM2 deficit test is significant under the pre-registered specification but not robust; Phi-3 non-inferiority holds across margins and a word-overlap covariate, although its bootstrap interval includes zero; Qwen2-0.5B is inconclusive. A capacity explanation remains untested. We report vehicle-level signals only, not meaning.

## Calpha: Toward Native Biomolecular Intelligence in a General-Purpose Language Model

DevFeed: [Calpha: Toward Native Biomolecular Intelligence in a General-Purpose Language Model](<https://devfeed.tech/articles/calpha-toward-native-biomolecular-intelligence-in-a-general-purpose-language-model-137641.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20986>)

Author: Yong Liu, Zhanpeng Shi, Peicong Lin, Zhenyi Zhong, Yizhou Dang, Wendi Wu, Haoling Li, Shuangjia Zheng

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

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

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [proteins](<https://devfeed.tech/tags/proteins.md>), [q-bio-qm](<https://devfeed.tech/tags/q-bio-qm.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [structure](<https://devfeed.tech/tags/structure.md>)

### AI overview

Calpha demonstrates a general-purpose LLM producing single-chain protein structures from its own residue states: Qwen3.5-9B reads sequences enriched with ESM2 features and can invoke a pretrained SimpleFold head. Under a fixed folding request, mean TM-scores were 0.808 on FoldBench-334, 0.635 on CASP16 domains, and 0.900 on a CAMEO-style PDB set. Compared with an ESM2-3B encoder and fine-tuned SimpleFold-3B baseline, results for 100-299-residue chains had paired confidence intervals that included zero; for 300-600 residues, reported mean differences were −0.044, −0.160, and −0.025. The systems also differed in trainable weights, training duration, and crop length, limiting direct attribution. Four English request types triggered folding on all evaluated examples, while protein training lowered accuracy on a 350-item MMLU-Pro subset from 77.4% to 70.3%. Complex prediction, structure understanding, and language reasoning's effect on coordinates remain future work.

### Source excerpt

arXiv:2611.20986v1 Announce Type: cross Abstract: Biomolecular foundation models predict the structures of proteins and their complexes accurately, and large language models (LLMs) reason in natural language, but the two remain separate: the abilities of the former are not part of a general-purpose LLM, and the reasoning of the latter does not enter a structure prediction. Calpha aims at one general model that folds proteins and complexes, understands protein sequence and structure, and lets its language reasoning influence the structures it predicts. This report covers the first step, in which an LLM outputs single-chain protein structures. Qwen3.5-9B reads a sequence enriched with ESM2 residue features. It can respond in text or emit a token that invokes a pretrained SimpleFold structure head, which receives information about the protein only through Qwen's residue states. Under a fixed folding request, the head achieves mean TM-scores of 0.808 on the FoldBench-334 development set, 0.635 on CASP16 domains and 0.900 on a CAMEO-style PDB set. We compare these predictions with a baseline in which ESM2-3B encodes the sequence and a fine-tuned SimpleFold-3B head predicts the coordinates. The two systems also differ in trainable weights, training duration and crop length. For chains of 100--299 residues, all three paired confidence intervals include zero. At 300--600 residues, the mean differences are -0.044, -0.160 and -0.025, respectively. In a separate test, four English request types triggered folding on every evaluated example. Protein training reduced accuracy on a 350-item MMLU-Pro subset from 77.4% to 70.3%. An LLM can therefore output single-chain structures from its own states, with lower long-chain scores in the current comparisons and reduced accuracy on the evaluated MMLU-Pro subset. Complex prediction, structure understanding and the effect of language reasoning on the predicted coordinates remain future work.

## Out of Sight, Into Memory: Compact Visual Latents Retain Local Evidence in Ultra-High-Resolution Remote Sensing

DevFeed: [Out of Sight, Into Memory: Compact Visual Latents Retain Local Evidence in Ultra-High-Resolution Remote Sensing](<https://devfeed.tech/articles/out-of-sight-into-memory-compact-visual-latents-retain-local-evidence-in-ultra-high-resolution-remote-sensing-137700.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21457>)

Author: Meng Tang, Siyuan Du, Yuxin Liu, Shuaishuai Cao, Di Wang, Haonan Guo

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

Content type: article

Language: en

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

Topics: [Image](<https://devfeed.tech/topics/image.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [image recognition](<https://devfeed.tech/topics/image-recognition.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [memory](<https://devfeed.tech/tags/memory.md>), [remote-sensing](<https://devfeed.tech/tags/remote-sensing.md>), [retention](<https://devfeed.tech/tags/retention.md>), [tokens](<https://devfeed.tech/tags/tokens.md>)

### AI overview

The study finds that 64 continuous visual latents can preserve useful local evidence from high-resolution remote-sensing crops after the crop tokens are discarded. With frozen backbones, this memory compressed retained crop tokens by 7.6-9.9x and improved accuracy over a downsampled global image by 2.5-4.8 percentage points on LRS-GRO, with consistent behavior on XLRS-Bench. Replacing memory with another image's or another region's memory reduced accuracy by 5.8-7.4 points, to below the no-memory result, indicating that predictions relied on the encoded regional evidence. Restoring the source crop reduced accuracy losses from zeroing or removing memory. The reported findings concern evidence retention and content dependence; the abstract does not establish broader performance beyond these evaluations.

### Source excerpt

arXiv:2611.21457v1 Announce Type: cross Abstract: Question answering over ultra-high-resolution (UHR) remote sensing scenes poses a tension for multimodal large language models: global overviews obscure small objects and local attributes, whereas detailed crops consume substantial context. Compressing inspected crops into short sequences of continuous visual latents offers a compact memory once crop tokens are discarded, yet diagnostic studies on natural images find that models often ignore latent content, leaving unclear whether compact latents can retain fine-grained targets lost in downsampled UHR remote sensing scenes. To isolate evidence retention from region selection, we encode high-resolution crops into latent memory tokens for frozen backbones and evaluate both their predictive utility and content dependence. Across two generations of Qwen-VL on LRS-GRO, a 64-token memory achieves 7.6- to 9.9-fold compression of retained crop tokens and improves accuracy by 2.5 to 4.8 percentage points over the downsampled global image alone, recovering most of the high-resolution crop gain with consistent behavior on XLRS-Bench. Substituting memory from another image or another region of the same scene reduces accuracy by 5.8 to 7.4 percentage points and yields lower accuracy than removing memory altogether, demonstrating that predictions depend on the encoded regional evidence. Restoring the source crop within the same model reduces the accuracy loss caused by zeroing or removing memory. These findings establish that compact visual latents retain useful local evidence in UHR remote sensing and identify evidence availability as a key factor governing their contribution.

## Internal Verdicts Track Evidence: A Template-Controlled Neutrosophic Reading of Epistemic States in Large Language Models via the Jacobian Lens

DevFeed: [Internal Verdicts Track Evidence: A Template-Controlled Neutrosophic Reading of Epistemic States in Large Language Models via the Jacobian Lens](<https://devfeed.tech/articles/internal-verdicts-track-evidence-a-template-controlled-neutrosophic-reading-of-epistemic-states-in-large-language-models-via-the-jacobian-lens-138749.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.18602>)

Author: Maikel Yelandi Leyva-Vazquez, Alexis Matheu Perez, Florentin Smarandache

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

Content type: article

Language: en

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

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [audit](<https://devfeed.tech/topics/audit.md>), [LLM observability](<https://devfeed.tech/topics/llm-observability.md>)

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [inside](<https://devfeed.tech/tags/inside.md>), [internals](<https://devfeed.tech/tags/internals.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [lens](<https://devfeed.tech/tags/lens.md>), [measurement](<https://devfeed.tech/tags/measurement.md>), [models](<https://devfeed.tech/tags/models.md>), [qwen3-5-4b](<https://devfeed.tech/tags/qwen3-5-4b.md>), [template](<https://devfeed.tech/tags/template.md>), [templates](<https://devfeed.tech/tags/templates.md>)

### AI overview

Using the Jacobian lens on Qwen3.5-4B, the authors read layer-wise truth, indeterminacy, and falsity signals from support, refutation, and hedging lexicons. Initial patterns under conflicting evidence and false belief largely disappeared under matched prompt templates, indicating sensitivity to prompt grammar. A verdict balance, B = log10(F) − log10(T), remained associated with evidence polarity under identical templates: roughly three orders of magnitude of separation across all nine reported layers, with Cohen's d of 2.3-2.5. It also residually distinguished false from true first-person beliefs. The abstract presents this two-battery comparison as a reusable protocol for auditing lens-based readings of model internals.

### Source excerpt

arXiv:2611.18602v1 Announce Type: cross Abstract: For twenty-five years, neutrosophic logic has assigned independent degrees of truth (T), indeterminacy (I), and falsity (F) to epistemic states that classical and fuzzy semantics cannot express. Previous applications measured these components on outputs: answers, judgments, expert evaluations. This paper reports, to our knowledge, the first neutrosophic reading of components inside the internal representations of a large language model. Using the recently released Jacobian lens on Qwen3.5-4B, we project intermediate-layer readouts onto lexicons of support, refutation, and hedging, obtaining layer-wise (T, I, F) profiles under conflicting evidence, first-person false belief, and factual control (n = 20 each). A first battery yielded three striking signatures: a surge of refutation mass under conflict, sustained indeterminacy under false belief, and an apparent "verdict collapse" before the output layer. A second, template-controlled battery showed that all three, as initially stated, were largely artifacts of prompt grammar: matched templates with agreeing evidence inflate the same lexicon masses, true beliefs elicit the same hedging, and the model verbalizes its verdict when allowed to generate. What survives the controls is stronger than what died: the neutrosophic verdict balance B = log10(F) - log10(T) tracks evidence polarity under identical templates (about 3 orders of magnitude of separation, Cohen's d = 2.3-2.5, p < 10^-6 in 9 of 9 layers; paired Wilcoxon: median difference +2.4 orders, 19/20 items, p = 3.8x10^-6) and residually discriminates false from true first-person beliefs. We argue that the independence of the neutrosophic components is what keeps the common-mode channel observable, making both the measurement and its self-correction possible, and we distill the two-battery design into a reusable protocol for auditing language model internals with lens-based readouts.

## Confidence Without Cause: Predicting and Localizing Physics-Shortcut Sensitivity in Language Model Hidden States

DevFeed: [Confidence Without Cause: Predicting and Localizing Physics-Shortcut Sensitivity in Language Model Hidden States](<https://devfeed.tech/articles/confidence-without-cause-predicting-and-localizing-physics-shortcut-sensitivity-in-language-model-hidden-states-137444.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19403>)

Author: Shubham Mishra, Atharv Sharma, Dhruv Kumar

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

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: [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Language models](<https://devfeed.tech/topics/language-models.md>), [Ground truth / benchmark quality](<https://devfeed.tech/topics/ground-truth-benchmark-quality.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [machine learning overfitting](<https://devfeed.tech/topics/machine-learning-overfitting.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [hidden-states](<https://devfeed.tech/tags/hidden-states.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [layer](<https://devfeed.tech/tags/layer.md>), [linear](<https://devfeed.tech/tags/linear.md>), [physics](<https://devfeed.tech/tags/physics.md>)

### AI overview

DECOY uses 155 families of four physics prompts, with a symbolic solver certifying that each changed detail leaves the answer unchanged, to measure sensitivity to irrelevant distractors. Among families with labels, Qwen2.5-7B-Instruct was sensitive on 50.0% and Llama-3.1-8B-Instruct on 22.9%; effects were largest when distractor values were numerically close to the answer. A Qwen hidden-state probe predicted sensitivity before generation (AUROC 0.760 with whole families held out), performing on par with a fitted token-entropy baseline (0.727) and achieving the best tested Brier score (0.204 vs. 0.214). On 15 families built after training, a frozen probe reached 0.820 pooled AUROC, but only 0.667-0.6875 within cue types. Knocking out four layer-16 attention heads canceled sensitivity on a 20-family panel. The results suggest detectability and a targeted intervention, with limited evidence for generalization across cue types.

### Source excerpt

arXiv:2611.19403v1 Announce Type: new Abstract: A language model's confidence on a physics problem deserves trust only if it ignores details irrelevant to the answer. We introduce DECOY (Distractor-Embedded Certified cOunterfactuals for phYsics): 155 families of four prompts differing in one irrelevant detail. A symbolic solver certifies that each edit preserves the answer. Families are labelled by the largest drop in the model's log-probability of the certified answer, using a threshold fixed before probe training. Qwen2.5-7B-Instruct is sensitive on 50.0% of the 140 families yielding labels, and Llama-3.1-8B-Instruct on 22.9%. Effects are largest when distractor values lie numerically close to the answer. A linear probe on Qwen's layer-18 hidden state, averaged over four renderings, predicts sensitivity before generation at AUROC 0.760 with whole families held out. A separately tuned single-rendering probe reaches 0.747. A frozen layer-16 probe reaches 0.820 pooled AUROC on 15 families built after training, though within-cue-type AUROC is lower (0.667-0.6875). Under a paired family bootstrap, the probe performs on par with a fitted token-entropy baseline (AUROC 0.727) and achieves the lowest Brier score among tested predictors (0.204 versus 0.214). Knocking out four attention heads at layer 16 cancels sensitivity on a 20-family panel, localising it to a narrow mechanism. These results show that Qwen's confidence can track irrelevant details, that this failure is detectable before generation, and that targeted intervention can suppress it. DECOY provides certified ground truth and paired-tested baselines for evaluating hidden-state monitors.

## FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic Language Models

DevFeed: [FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic Language Models](<https://devfeed.tech/articles/flashreinforce-critic-free-single-rollout-asynchronous-rl-for-agentic-language-models-138753.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.18627>)

Author: Jian Hu, Yifan Zhang, Hao Zhang, Binfeng Xu, Shaokun Zhang, Hongqing Peng, Zhiding Yu, Pavlo Molchanov, Jan Kautz, Yi Dong

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

Content type: article

Language: en

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

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [grpo](<https://devfeed.tech/topics/grpo.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [asynchronous](<https://devfeed.tech/tags/asynchronous.md>), [batch](<https://devfeed.tech/tags/batch.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [deepseek-r1](<https://devfeed.tech/tags/deepseek-r1.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [framework](<https://devfeed.tech/tags/framework.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [math](<https://devfeed.tech/tags/math.md>), [model](<https://devfeed.tech/tags/model.md>), [moe](<https://devfeed.tech/tags/moe.md>), [one](<https://devfeed.tech/tags/one.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [pass](<https://devfeed.tech/tags/pass.md>), [policy](<https://devfeed.tech/tags/policy.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [python](<https://devfeed.tech/tags/python.md>), [rl](<https://devfeed.tech/tags/rl.md>)

### AI overview

FlashREINFORCE is a critic-free asynchronous RL method that uses one rollout per prompt to increase prompt coverage and avoid group synchronization, while addressing noisier updates on stale, variable-length trajectories with One-Batch REINFORCE, a Sequence Trust Region, and Sample-Mean Optimization. The abstract reports stable training through 6,000 updates at roughly four updates of policy lag on DeepSeek-R1-Distill-Qwen-1.5B, and stable Python-tool training at lag eight on Qwen3-30B-A3B. On Qwen2.5-Math-1.5B it scores 38.0 mean accuracy across five benchmarks, 1.7 points above the reported GRPO baseline using 256k rather than 512k rollouts. In ALFWorld, it retains tool use where GRPO stops making calls and reports 98.3% seen and 96.5% unseen success on Qwen2.5-7B-Instruct.

### Source excerpt

arXiv:2611.18627v1 Announce Type: cross Abstract: Asynchronous reinforcement learning suits long-horizon agents with irregular rollout times, but group-relative updates require repeated rollouts of each prompt. However, group-based sampling limits prompt coverage at a fixed rollout budget and introduces group synchronization barriers. One rollout per prompt avoids these constraints but removes the variance-reducing group baseline, making updates noisier and more prone to instability on stale, variable-length trajectories. We present FlashREINFORCE, a critic-free, one-pass framework built on One-Batch REINFORCE, a Sequence Trust Region, and Sample-Mean Optimization. These components preserve prompt coverage and signed feedback while controlling policy drift and length-amplified negative updates. In our experiments, FlashREINFORCE remains stable through 6,000 updates on DeepSeek-R1-Distill-Qwen-1.5B at a policy lag of approximately four updates. On Qwen2.5-Math-1.5B, it reaches 38.0 mean accuracy across five benchmarks, outperforming the reported GRPO baseline by 1.7 percentage points with half the rollouts (256k versus 512k). Python-tool experiments extend to the 30B MoE model Qwen3-30B-A3B, with stable training at policy lag eight. On Qwen2.5-7B-Instruct, FlashREINFORCE maintains tool use where GRPO stops making tool calls and achieves 98.3%/96.5% seen/unseen ALFWorld success.

## GRAFT: Cross-Backbone LoRA Reuse for Continual Language and Multimodal Learning

DevFeed: [GRAFT: Cross-Backbone LoRA Reuse for Continual Language and Multimodal Learning](<https://devfeed.tech/articles/graft-cross-backbone-lora-reuse-for-continual-language-and-multimodal-learning-138757.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.18638>)

Author: Ha Pham Thien Dinh, Ming Liu, Jiesi Guo, Matteo Vergani, Kewen Liao

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

Content type: article

Language: en

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

Topics: [lora](<https://devfeed.tech/topics/lora.md>), [peft](<https://devfeed.tech/topics/peft.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Model Development](<https://devfeed.tech/topics/model-development.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [adapter](<https://devfeed.tech/tags/adapter.md>), [adapters](<https://devfeed.tech/tags/adapters.md>), [bridge](<https://devfeed.tech/tags/bridge.md>), [computer-science](<https://devfeed.tech/tags/computer-science.md>), [continual-learning](<https://devfeed.tech/tags/continual-learning.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [feature](<https://devfeed.tech/tags/feature.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [library](<https://devfeed.tech/tags/library.md>), [linear](<https://devfeed.tech/tags/linear.md>), [lora](<https://devfeed.tech/tags/lora.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [multimodal-learning](<https://devfeed.tech/tags/multimodal-learning.md>), [multimodal-models](<https://devfeed.tech/tags/multimodal-models.md>), [reuse](<https://devfeed.tech/tags/reuse.md>)

### AI overview

GRAFT reuses LoRA adapters across different model backbones by retrieving a relevant adapter from a modality-aware library and mapping it to the target model with a learned linear bridge between feature spaces. The evaluation transfers Qwen2.5 adapters to Qwen2.5-VL on ten Super-NaturalInstructions text tasks and directly reuses Qwen2.5-VL adapters on five multimodal tasks. Compared with randomly initialized LoRA adapters, GRAFT improves performance on most evaluated tasks; the supplied excerpt does not specify the size of those gains.

### Source excerpt

arXiv:2611.18638v1 Announce Type: new Abstract: In continual learning, parameter-efficient fine-tuning enables large language models to acquire new capabilities through lightweight adapters such as LoRA, yet the reuse of pretrained adapters is typically restricted to models sharing the same backbone. This limits the transfer of learned capabilities from language models to multimodal models. We propose GRAFT (Generalist Reuse of Adapters via Feature-space Transfer), a framework for cross-backbone LoRA reuse. GRAFT retrieves the most relevant pretrained LoRA adapter from a modality-aware library and maps it into the target backbone through a learned linear bridge between corresponding source and target feature spaces. We evaluate GRAFT on ten Super-NaturalInstructions text tasks, transferring Qwen2.5 LoRAs to Qwen2.5-VL, and five multimodal tasks, where Qwen2.5-VL LoRAs are directly reused. Compared with randomly initialized LoRA adapters, GRAFT improves performance on most evaluated tasks. These results show that GRAFT provides a simple and effective mechanism for reusing learned capabilities across language and multimodal models in continual learning.

## Alibaba launches Qwen4-Max with training, attention optimizations

DevFeed: [Alibaba launches Qwen4-Max with training, attention optimizations](<https://devfeed.tech/articles/alibaba-launches-qwen4-max-with-training-attention-optimizations-136251.md>)

Original publisher: [Read original article](<https://siliconangle.com/2026/11/16/alibaba-launches-qwen4-max-with-training-attention-optimizations/>)

Author: Maria Deutscher

Published: 2026-11-16T21:21:09Z

Content type: news

Language: en

Sources: [SiliconANGLE](<https://devfeed.tech/sources/siliconangle.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [LLM Techniques](<https://devfeed.tech/topics/llm-techniques.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [algorithm](<https://devfeed.tech/tags/algorithm.md>), [alibaba](<https://devfeed.tech/tags/alibaba.md>), [alibaba-group](<https://devfeed.tech/tags/alibaba-group.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [attention](<https://devfeed.tech/tags/attention.md>), [gated-deltanet](<https://devfeed.tech/tags/gated-deltanet.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [launches](<https://devfeed.tech/tags/launches.md>), [llm](<https://devfeed.tech/tags/llm.md>), [neural-network](<https://devfeed.tech/tags/neural-network.md>), [news](<https://devfeed.tech/tags/news.md>), [optimizations](<https://devfeed.tech/tags/optimizations.md>), [qwen4-max](<https://devfeed.tech/tags/qwen4-max.md>), [the-latest](<https://devfeed.tech/tags/the-latest.md>), [top-story-4](<https://devfeed.tech/tags/top-story-4.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

Alibaba says Qwen4-Max combines a 2.4-trillion-parameter backbone that activates 71 billion parameters at a time with 328 billion N-gram cache parameters. Its Qwen4 architecture uses four gated gradient channels, which Alibaba says reduce training errors and costs while improving output quality. For prompt processing, Gated DeltaNet reduces hardware needs for large prompts, and QSA skips data deemed unnecessary to answer. In Alibaba's internal comparison with Qwen3.8-Max, Qwen4-Max scored higher and used up to 35% fewer tokens on knowledge-work tasks; token savings reached 49% on Terminal-Bench 4.0 coding challenges. The model is available through Alibaba Cloud. Alibaba also plans smaller open-source models and is open-sourcing a plugin that delivers an initial draft for long tasks, then incorporates improvements and user feedback iteratively.

### Source excerpt

Alibaba Group Holding Ltd. today introduced Qwen4-Max, its newest and most capable large language model. The algorithm not only generates higher-quality output than its predecessor but also costs less. Those improvements are the result of a new architecture, Qwen4, that made the LLM's training run more efficient. Additionally, Alibaba has equipped the model with an [...] The post Alibaba launches Qwen4-Max with training, attention optimizations appeared first on SiliconANGLE.

## With Qwen4-Max, Alibaba puts the open source jacket back on the hanger

DevFeed: [With Qwen4-Max, Alibaba puts the open source jacket back on the hanger](<https://devfeed.tech/articles/with-qwen4-max-alibaba-puts-the-open-source-jacket-back-on-the-hanger-136130.md>)

Original publisher: [Read original article](<https://thenewstack.io/qwen4-max-alibaba-open-weights/>)

Author: Paul Sawers

Published: 2026-11-16T16:54:17Z

Content type: news

Language: en

Sources: [The New Stack](<https://devfeed.tech/sources/the-new-stack.md>)

Topics: [qwen](<https://devfeed.tech/topics/qwen.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-models](<https://devfeed.tech/tags/ai-models.md>), [alibaba](<https://devfeed.tech/tags/alibaba.md>), [api](<https://devfeed.tech/tags/api.md>), [api-pricing](<https://devfeed.tech/tags/api-pricing.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [artificial-analysis](<https://devfeed.tech/tags/artificial-analysis.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [business](<https://devfeed.tech/tags/business.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [code](<https://devfeed.tech/tags/code.md>), [codex](<https://devfeed.tech/tags/codex.md>), [coding](<https://devfeed.tech/tags/coding.md>), [coding-tools](<https://devfeed.tech/tags/coding-tools.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cursor](<https://devfeed.tech/tags/cursor.md>), [model](<https://devfeed.tech/tags/model.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [qwen4-max](<https://devfeed.tech/tags/qwen4-max.md>)

### AI overview

Qwen4-Max is a hosted, proprietary model; Alibaba announced no downloadable weights or release commitment for it, though it said smaller Qwen4 models would have open weights without specifying which or when. The model uses a mixture-of-experts architecture, supports image and text input, and has a one-million-token context window. Alibaba offers it through QwenCloud and other products, with API pricing starting at $1.50 per million input tokens and $4.50 per million output tokens. Alibaba reports improved benchmark performance and lower token use and task costs versus Qwen3.8-Max; independent Artificial Analysis results place it above the median for its price category but below several leading competitors, with higher-than-median output-token use. The article contrasts this release with Qwen3.8-Max, whose downloadable weights had a custom licence, restrictions for some commercial services, and fewer capabilities than the hosted version.

### Source excerpt

Alibaba has officially unveiled Qwen4-Max, the first model in its next-generation Qwen4 family, with a focus on complex professional tasks The post With Qwen4-Max, Alibaba puts the open source jacket back on the hanger appeared first on The New Stack.

## Alibaba Officially Releases Qwen4-Max: Tokyo Region Input Pricing Is $1.413 per Million Tokens

DevFeed: [Alibaba Officially Releases Qwen4-Max: Tokyo Region Input Pricing Is $1.413 per Million Tokens](<https://devfeed.tech/articles/alibaba-officially-releases-qwen4-max-tokyo-region-input-pricing-is-1-413-per-million-tokens-135496.md>)

Original publisher: [Read original article](<https://joho-todai.com/en/qwen4-max-alibaba-tokyo-region-pricing/>)

Author: Toru Misaki

Published: 2026-11-16T12:31:23Z

Content type: news

Language: en

Sources: [Joho no Todai](<https://devfeed.tech/sources/joho-no-todai.md>)

Topics: [qwen](<https://devfeed.tech/topics/qwen.md>), [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>), [API Business Models](<https://devfeed.tech/topics/api-business-models.md>), [function calling](<https://devfeed.tech/topics/function-calling.md>)

Tags: [alibaba](<https://devfeed.tech/tags/alibaba.md>), [api](<https://devfeed.tech/tags/api.md>), [pricing](<https://devfeed.tech/tags/pricing.md>), [qwen4-max](<https://devfeed.tech/tags/qwen4-max.md>), [release](<https://devfeed.tech/tags/release.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [tokyo](<https://devfeed.tech/tags/tokyo.md>)

### AI overview

Qwen4-Max is available through QwenCloud's Tokyo region at $1.413 per million input tokens, $4.24 per million output tokens, and $0.071 per million cached input tokens. Its MoE model has 2.4 trillion total parameters and activates 71 billion per token; the article describes attention and embedding mechanisms aimed at handling long inputs with lower compute. The API supports a roughly 1-million-token context, multimodal input, tool use, and OpenAI- and Anthropic-compatible modes. Tokyo workspaces can use a Japan-only deployment scope, while Global may include mainland China nodes. Tokyo access is real-time only; batch processing, fine-tuning, and evaluation are limited to Singapore. Alibaba reports improved Work360 scores and lower task costs, but those are its own results; Artificial Analysis scores the model 43, slightly above its evaluated-model median but below the leading group. The flagship weights have not been released.

### Source excerpt

The Qwen team announced Qwen4-Max on November 16. A preview had been available only to subscription plans since November 11, and the official release now brings metered API access through QwenCloud.

## Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards

DevFeed: [Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards](<https://devfeed.tech/articles/alibaba-touts-62-lower-task-costs-with-qwen4-max-adds-cyber-safeguards-136232.md>)

Original publisher: [Read original article](<https://www.computerworld.com/article/4249020/alibaba-touts-62-lower-task-costs-with-qwen4-max-adds-cyber-safeguards-3.html>)

Author: Gyana Swain

Published: 2026-11-16T12:24:39Z

Content type: news

Language: en

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

Topics: [qwen](<https://devfeed.tech/topics/qwen.md>), [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>), [cost-optimization](<https://devfeed.tech/topics/cost-optimization.md>), [cloud security](<https://devfeed.tech/topics/cloud-security.md>), [audit](<https://devfeed.tech/topics/audit.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [alibaba](<https://devfeed.tech/tags/alibaba.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [artificial-intelligence-generative-ai](<https://devfeed.tech/tags/artificial-intelligence-generative-ai.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [cost](<https://devfeed.tech/tags/cost.md>), [costs](<https://devfeed.tech/tags/costs.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [edition-us](<https://devfeed.tech/tags/edition-us.md>), [enterprises](<https://devfeed.tech/tags/enterprises.md>), [generative-ai](<https://devfeed.tech/tags/generative-ai.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [mixture-of-experts](<https://devfeed.tech/tags/mixture-of-experts.md>), [model](<https://devfeed.tech/tags/model.md>), [offensive](<https://devfeed.tech/tags/offensive.md>), [performance](<https://devfeed.tech/tags/performance.md>), [prices](<https://devfeed.tech/tags/prices.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [qwen4-max](<https://devfeed.tech/tags/qwen4-max.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reduction](<https://devfeed.tech/tags/reduction.md>), [security](<https://devfeed.tech/tags/security.md>), [storytype-news](<https://devfeed.tech/tags/storytype-news.md>), [tasks](<https://devfeed.tech/tags/tasks.md>)

### AI overview

Alibaba says Qwen4-Max cuts average task cost to $1.67 from $4.43 for Qwen3.8-Max on its Work360 benchmark, using 35% fewer tokens at high reasoning effort. The company attributes savings to lower token use and reduced prices; the model retains a 2.4-trillion-parameter mixture-of-experts backbone, activates about 71 billion parameters per token, and supports a one-million-token context. Alibaba also says training-data filters and security checks improved defensive coding results, with the share of functions passing both functionality and security tests rising from 22% to 31.8%. These are vendor-reported results, not guaranteed enterprise savings or independently validated safeguards. Analysts recommend measuring the cost of successful business tasks, including retries and review, and independently testing cyber protections. For sensitive deployments, they advise examining data processing, retention, jurisdiction, and agent permissions; fine-tuned open-weight versions need renewed security testing.

### Source excerpt

Alibaba has released Qwen4-Max, a 2.4-trillion-parameter flagship model, claiming it reduces the average cost of completing professional tasks by 62% compared with its predecessor while introducing additional safeguards against offensive cybersecurity capabilities and other high-risk uses. "According to Work360 -- a proprietary benchmark evaluating performance on challenging industry tasks across 721 specialized fields -- Qwen4-Max outperforms its predecessor, Qwen3.8-Max, overall while using 35% fewer tokens at High effort reasoning mode and reducing the average cost per task by around 60%," Alibaba said in a blog post. The Qwen team's announcement puts the average task cost at $1.67 for Qwen4-Max, compared with $4.43 for Qwen3.8-Max, representing a 62% reduction. The model keeps the 2.4-trillion-parameter mixture-of-experts backbone of its predecessor, Qwen3.8-Max, but activates about 71 billion parameters per token instead of about 95 billion, the company said. It supports a 1-million-token context window. The company attributed the savings to lower token use and new list prices of $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6. Anthropic lists Claude Opus 5.5 at $20 per million output tokens. Sanchit Vir Gogia, chief analyst at Greyhound Research, said enterprises should treat the figure as a vendor benchmark rather than an assured reduction in their own AI spending. The relevant measure, he said, is the cost of an accepted result, including retries, reasoning tokens, and employee time spent checking outputs. "Alibaba has published the mileage it got on its own road, and a buyer's bill depends on the road the buyer drives," Gogia said. Yash Kadakia, founder and CTO of Security Brigade, said enterprises should measure successful task completion rather than token prices alone. "An enterprise should benchmark what it costs to complete 1,000 real business tasks successfully, not what it costs to generate a million tokens," Kada

## Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards

DevFeed: [Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards](<https://devfeed.tech/articles/alibaba-touts-62-lower-task-costs-with-qwen4-max-adds-cyber-safeguards-136000.md>)

Original publisher: [Read original article](<https://www.infoworld.com/article/4248996/alibaba-touts-62-lower-task-costs-with-qwen4-max-adds-cyber-safeguards.html>)

Author: Gyana Swain

Published: 2026-11-16T12:17:29Z

Content type: news

Language: en

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

Topics: [qwen](<https://devfeed.tech/topics/qwen.md>), [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [anthropic](<https://devfeed.tech/topics/anthropic.md>)

Tags: [alibaba](<https://devfeed.tech/tags/alibaba.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [artificial-intelligence-development-tools-generative-ai-software-development](<https://devfeed.tech/tags/artificial-intelligence-development-tools-generative-ai-software-development.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [claude-opus-5-5](<https://devfeed.tech/tags/claude-opus-5-5.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [cost](<https://devfeed.tech/tags/cost.md>), [costs](<https://devfeed.tech/tags/costs.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [development-tools](<https://devfeed.tech/tags/development-tools.md>), [enterprises](<https://devfeed.tech/tags/enterprises.md>), [generative-ai](<https://devfeed.tech/tags/generative-ai.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [mixture-of-experts](<https://devfeed.tech/tags/mixture-of-experts.md>), [model](<https://devfeed.tech/tags/model.md>), [offensive](<https://devfeed.tech/tags/offensive.md>), [performance](<https://devfeed.tech/tags/performance.md>), [prices](<https://devfeed.tech/tags/prices.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [qwen4-max](<https://devfeed.tech/tags/qwen4-max.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reduction](<https://devfeed.tech/tags/reduction.md>), [risk](<https://devfeed.tech/tags/risk.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>)

### AI overview

Alibaba says Qwen4-Max averaged $1.67 per professional task on its proprietary Work360 benchmark, versus $4.43 for Qwen3.8-Max, using 35% fewer tokens at high reasoning effort. It attributes the reduction to lower token use and reduced list prices; the model retains a 2.4-trillion-parameter mixture-of-experts backbone, activates about 71 billion parameters per token, and supports a one-million-token context. Analysts caution that actual enterprise cost depends on successful accepted results, retries, and employee review. Alibaba says it filtered training data to retain defensive cybersecurity tasks and remove offensive ones, and reports an increase in code passing functionality and security tests from 22% to 31.8%. Experts question whether those skills can be separated and call for independent testing. They also advise reviewing prompt retention, data jurisdiction, agent permissions, and re-testing customized open-weight models, since fine-tuning may not preserve safeguards.

### Source excerpt

Alibaba has released Qwen4-Max, a 2.4-trillion-parameter flagship model, claiming it reduces the average cost of completing professional tasks by 62% compared with its predecessor while introducing additional safeguards against offensive cybersecurity capabilities and other high-risk uses. "According to Work360 -- a proprietary benchmark evaluating performance on challenging industry tasks across 721 specialized fields -- Qwen4-Max outperforms its predecessor, Qwen3.8-Max, overall while using 35% fewer tokens at High effort reasoning mode and reducing the average cost per task by around 60%," Alibaba said in a blog post. The Qwen team's announcement puts the average task cost at $1.67 for Qwen4-Max, compared with $4.43 for Qwen3.8-Max, representing a 62% reduction. The model keeps the 2.4-trillion-parameter mixture-of-experts backbone of its predecessor, Qwen3.8-Max, but activates about 71 billion parameters per token instead of about 95 billion, the company said. It supports a 1-million-token context window. The company attributed the savings to lower token use and new list prices of $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6. Anthropic lists Claude Opus 5.5 at $20 per million output tokens. Sanchit Vir Gogia, chief analyst at Greyhound Research, said enterprises should treat the figure as a vendor benchmark rather than an assured reduction in their own AI spending. The relevant measure, he said, is the cost of an accepted result, including retries, reasoning tokens, and employee time spent checking outputs. "Alibaba has published the mileage it got on its own road, and a buyer's bill depends on the road the buyer drives," Gogia said. Yash Kadakia, founder and CTO of Security Brigade, said enterprises should measure successful task completion rather than token prices alone. "An enterprise should benchmark what it costs to complete 1,000 real business tasks successfully, not what it costs to generate a million tokens," Kada

## Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context

DevFeed: [Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context](<https://devfeed.tech/articles/alibaba-releases-qwen4-max-2-4t-moe-71b-active-1m-context-136412.md>)

Original publisher: [Read original article](<https://www.marktechpost.com/2026/11/16/alibaba-qwen4-max-release/>)

Author: Asif Razzaq

Published: 2026-11-16T09:44:17Z

Content type: release

Language: en

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

Topics: [qwen](<https://devfeed.tech/topics/qwen.md>), [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>), [context window](<https://devfeed.tech/topics/context-window.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>)

Tags: [1-5](<https://devfeed.tech/tags/1-5.md>), [2](<https://devfeed.tech/tags/2.md>), [3](<https://devfeed.tech/tags/3.md>), [about](<https://devfeed.tech/tags/about.md>), [agent](<https://devfeed.tech/tags/agent.md>), [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>), [alibaba](<https://devfeed.tech/tags/alibaba.md>), [applications](<https://devfeed.tech/tags/applications.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [context](<https://devfeed.tech/tags/context.md>), [context-engineering](<https://devfeed.tech/tags/context-engineering.md>), [cost](<https://devfeed.tech/tags/cost.md>), [editors-pick](<https://devfeed.tech/tags/editors-pick.md>), [enterprise-ai](<https://devfeed.tech/tags/enterprise-ai.md>), [first](<https://devfeed.tech/tags/first.md>), [for-devs](<https://devfeed.tech/tags/for-devs.md>), [harness](<https://devfeed.tech/tags/harness.md>), [input](<https://devfeed.tech/tags/input.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [max](<https://devfeed.tech/tags/max.md>), [model](<https://devfeed.tech/tags/model.md>), [moe](<https://devfeed.tech/tags/moe.md>), [new-releases](<https://devfeed.tech/tags/new-releases.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [qwen4-max](<https://devfeed.tech/tags/qwen4-max.md>), [releases](<https://devfeed.tech/tags/releases.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>), [tokens](<https://devfeed.tech/tags/tokens.md>), [vision](<https://devfeed.tech/tags/vision.md>), [with](<https://devfeed.tech/tags/with.md>)

### AI overview

Alibaba's Qwen4-Max is a hosted-only proprietary MoE model with a 2.4T-parameter backbone, 71B active parameters per token, multimodal input, and a 1M-token context. Its design pairs Qwen Sparse Attention with Gated DeltaNet layers, adds gated residual branches, and includes 328B N-gram Embedding parameters. In Alibaba's 16-benchmark comparison, it ranks first on four, including legal-agent work, while lagging on coding and vision measures such as ProgramBench and Chartography; scores were not independently reproduced at launch. Alibaba reports a 62% reduction in average Work360 cost per task, from $4.43 to $1.67, alongside lower API prices ($1.50/M input and $4.50/M output). Access is through Alibaba services and API; there are no open weights or self-hosting option.

### Source excerpt

Alibaba released Qwen4-Max, a proprietary 2.4T-parameter MoE model with 71B active parameters, native vision and a 1M-token context. It costs $1.5/M input tokens and cuts Work360 cost per task by 62%. The post Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context appeared first on MarkTechPost.

[Next page](<https://devfeed.tech/topics/qwen.md?cursor=WyIyMDI2LTExLTE2VDA5OjQ0OjE3KzAwOjAwIiwgIjYyMjJlY2M2LTM1N2QtNDJlNy04MDJhLWNjYzdjOTVlN2UxZCJd>)