# prompt

A prompt is input provided to a foundational model for processing and response generation.

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

## Can you make models believe false facts in the prompt?

DevFeed: [Can you make models believe false facts in the prompt?](<https://devfeed.tech/articles/can-you-make-models-believe-false-facts-in-the-prompt-142727.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23938>)

Author: Ananya Sutradhar, Fabien Roger

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

Content type: article

Language: en

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

Topics: [peft](<https://devfeed.tech/topics/peft.md>), [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [generalization in machine learning](<https://devfeed.tech/topics/generalization-in-machine-learning.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [audits](<https://devfeed.tech/tags/audits.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [finetuning](<https://devfeed.tech/tags/finetuning.md>), [jacobian](<https://devfeed.tech/tags/jacobian.md>), [llms](<https://devfeed.tech/tags/llms.md>), [models](<https://devfeed.tech/tags/models.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [rl](<https://devfeed.tech/tags/rl.md>), [scale](<https://devfeed.tech/tags/scale.md>), [sdf](<https://devfeed.tech/tags/sdf.md>), [small-models](<https://devfeed.tech/tags/small-models.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>), [system](<https://devfeed.tech/tags/system.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

Prompt-conditioned Synthetic Document Finetuning trains one model on synthetic documents covering thousands of facts, conditioning each document on a fact placed in a <facts> prompt tag. The authors report that belief depth on unseen facts scales roughly log-linearly with the number of training facts, suggesting a potentially inexpensive way to scale false-belief evaluations. In preliminary alignment-audit experiments on the small model studied, results were negative because it often already acted as if it believed facts in ordinary honeypot system prompts. Adding RL to reduce verbalized compliance with prompt falsehoods generalized to held-out evaluations and nearly matched SDF on some evaluations. However, training against one belief-depth metric made that evaluation less trustworthy.

### Source excerpt

arXiv:2611.23938v1 Announce Type: new Abstract: Making LLMs believe false facts is useful for alignment evaluations and for building model organisms. Previous work explored doing so with Synthetic Document Finetuning (SDF). But SDF requires fine-tuning one new model for each set of injected false beliefs. To enable larger scale SDF-powered alignment evaluations, we attempt to produce a single model that believes any fact placed in a tag in the prompt. To do so, we explore doing prompt-conditioned SDF: fine-tuning a model on synthetic documents covering thousands of facts, with each document conditioned on a fact presented in the prompt. We evaluate how deeply the model believes the facts provided in the prompt using evaluations from prior work and a new Jacobian lens evaluation. We find that the belief depth on unseen facts scales roughly log-linearly with the number of training facts, making this a cheap, promising avenue for making models believe false facts, though preliminary experiments on applying it to alignment audits on the small model we studied showed negative results due to these small models often already acting as if they believed facts provided in regular honeypot system prompts. Using RL to reduce the rate at which models verbalize following a false fact in the prompt, on top of prompt-conditioned SDF, generalizes to held-out evals and almost matches SDF on some evaluations. But training against one of our metrics for belief depth makes it harder to trust those evaluations.

## Tencent Hy4 is live on Together AI

DevFeed: [Tencent Hy4 is live on Together AI](<https://devfeed.tech/articles/tencent-hy4-is-live-on-together-ai-146600.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/tencent-hy4>)

Author: Sonny Khan

Published: 2026-11-19T00:00:00Z

Content type: release

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [browser-game](<https://devfeed.tech/topics/browser-game.md>), [Image](<https://devfeed.tech/topics/image.md>), [prompt](<https://devfeed.tech/topics/prompt.md>)

Tags: [browser-game](<https://devfeed.tech/tags/browser-game.md>), [coding](<https://devfeed.tech/tags/coding.md>), [game](<https://devfeed.tech/tags/game.md>), [hy4](<https://devfeed.tech/tags/hy4.md>), [image](<https://devfeed.tech/tags/image.md>), [live](<https://devfeed.tech/tags/live.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [tencent](<https://devfeed.tech/tags/tencent.md>), [together-ai](<https://devfeed.tech/tags/together-ai.md>)

### AI overview

Tencent Hy4 is available on Together Serverless Inference as `tencent/Hy4`, with a 1M-token context window, image input, tool calling, and JSON output. In the described tests, it extracted all three chart values into JSON and rebuilt the chart in a single HTML file, preserving its layout and styling details except for the requested generic font. From one text prompt, it produced a playable browser game in one HTML file without libraries or external assets; a round ended and restarted successfully. Hy4 is a 770B-parameter Mixture-of-Experts model with 49B active per token. Pricing is $0.83 per 1M input tokens, $0.042 per 1M cached input tokens, and $2.50 per 1M output tokens.

### Source excerpt

Tencent Hy4 is live on Together AI. We tested its image input by extracting and rebuilding a chart, then tried its coding with a browser game generated from one prompt.

## Simplicity Goes Far: Auditing Prompt Optimizers with Demonstration-Conditioned Prompt Search

DevFeed: [Simplicity Goes Far: Auditing Prompt Optimizers with Demonstration-Conditioned Prompt Search](<https://devfeed.tech/articles/simplicity-goes-far-auditing-prompt-optimizers-with-demonstration-conditioned-prompt-search-140056.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22605>)

Author: Bohe Chen

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

Content type: article

Language: en

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

Topics: [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [DSPy](<https://devfeed.tech/topics/dspy.md>), [audit](<https://devfeed.tech/topics/audit.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [classification](<https://devfeed.tech/tags/classification.md>), [classification-accuracy](<https://devfeed.tech/tags/classification-accuracy.md>), [control](<https://devfeed.tech/tags/control.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [dspy](<https://devfeed.tech/tags/dspy.md>), [generate](<https://devfeed.tech/tags/generate.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reflection](<https://devfeed.tech/tags/reflection.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [search](<https://devfeed.tech/tags/search.md>), [select](<https://devfeed.tech/tags/select.md>), [selection](<https://devfeed.tech/tags/selection.md>), [structured](<https://devfeed.tech/tags/structured.md>), [systems](<https://devfeed.tech/tags/systems.md>), [token-usage](<https://devfeed.tech/tags/token-usage.md>)

### AI overview

Demonstration-Conditioned Prompt Search (DCPS) isolates prompt generation and candidate selection from heavier optimizer mechanisms. It freezes the generator, conditions proposals on randomly sampled demonstrations, scores them on validation data, and retains the best candidate, without learned updates, reflection, or evolution. In StablePrompt, DCPS matches PPO classification accuracy within 0.5 percentage points at roughly one-third the measured cost. In DSPy-based systems on GPT-4.1-mini, it matches GEPA's macro accuracy under the reported budgets with 55% fewer rollouts and 32% lower average measured API cost. Results vary by task: proposal and selection recover much of heavier methods' performance on structured and instruction-following tasks, while added mechanisms retain advantages on some open-ended and reasoning-heavy tasks. The authors recommend comparing optimizers with generate-and-select controls and reporting rollouts, token use, and dollar costs alongside accuracy.

### Source excerpt

arXiv:2611.22605v1 Announce Type: new Abstract: Automatic prompt optimization is growing more elaborate. Recent methods stack policy-gradient updates, trace reflection, and evolutionary search. These methods share a generate-and-select step; without controlling for it, reported gains may conflate the contributions of optimizer mechanisms and larger sampling budgets. We isolate this step as Demonstration-Conditioned Prompt Search (DCPS), repurposing the Automatic Prompt Engineer (APE) recipe as a mechanism-attribution control: it freezes the generator, conditions it on randomly sampled demonstrations, scores candidates on validation data, and keeps the best---no learned update, reflection, or evolution. Across two optimizer families, DCPS replaces Proximal Policy Optimization (PPO) in StablePrompt, matching classification accuracy within 0.5 pp at roughly one-third the measured cost; in DSPy-based compound AI systems, under the reported budgets, it matches GEPA's macro accuracy on GPT-4.1-mini while using 55% fewer rollouts at 32% lower measured API cost on average. The audit reveals a task-dependent pattern: proposal and selection recover much of heavier optimizers' performance on structured and instruction-following tasks, while additional mechanisms retain advantages in some open-ended and reasoning-heavy settings. We recommend benchmarking prompt optimizers against generate-and-select controls and reporting rollout counts, token usage, and dollar costs alongside accuracy. Code and artifacts: https://github.com/bhc6/DCPS.

## PSM: Prompt Sensitivity Minimization via LLM-Guided Black-Box Optimization

DevFeed: [PSM: Prompt Sensitivity Minimization via LLM-Guided Black-Box Optimization](<https://devfeed.tech/articles/psm-prompt-sensitivity-minimization-via-llm-guided-black-box-optimization-141441.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2511.16209>)

Author: Hussein Jawad, Nicolas J-B Brunel

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

Content type: article

Language: en

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

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [prompt](<https://devfeed.tech/topics/prompt.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [api](<https://devfeed.tech/tags/api.md>), [black-box](<https://devfeed.tech/tags/black-box.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense](<https://devfeed.tech/tags/defense.md>), [framework](<https://devfeed.tech/tags/framework.md>), [hardening](<https://devfeed.tech/tags/hardening.md>), [hidden-instructions](<https://devfeed.tech/tags/hidden-instructions.md>), [lightweight](<https://devfeed.tech/tags/lightweight.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-security](<https://devfeed.tech/tags/llm-security.md>), [methodology](<https://devfeed.tech/tags/methodology.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [privacy](<https://devfeed.tech/tags/privacy.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [risks](<https://devfeed.tech/tags/risks.md>), [robust](<https://devfeed.tech/tags/robust.md>), [security](<https://devfeed.tech/tags/security.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [shield](<https://devfeed.tech/tags/shield.md>), [system](<https://devfeed.tech/tags/system.md>)

### AI overview

PSM hardens system prompts by appending an optimized protective text layer, or SHIELD. It frames prompt defense as a utility-constrained black-box optimization: an LLM optimizer searches for shields that reduce leakage measured across adversarial extraction attacks while keeping output semantic fidelity above a specified threshold. The approach requires API access to the target and optimizer LLMs. The abstract reports that optimized shields reduce leakage across a comprehensive attack suite and outperform baseline defenses without compromising intended functionality, but gives no quantitative results or further evaluation details.

### Source excerpt

arXiv:2511.16209v3 Announce Type: replace Abstract: System prompts are critical for guiding the behavior of Large Language Models (LLMs), yet they often contain proprietary logic or sensitive information, making them a prime target for extraction attacks. Adversarial queries can successfully elicit these hidden instructions, posing significant security and privacy risks. Existing defense mechanisms frequently rely on heuristics, incur substantial computational overhead, or are inapplicable to models accessed via black-box APIs. This paper introduces a novel framework for hardening system prompts through shield appending, a lightweight approach that adds a protective textual layer to the original prompt. Our core contribution is the formalization of prompt hardening as a utility-constrained optimization problem. We leverage an LLM-as-optimizer to search the space of possible SHIELDs, seeking to minimize a leakage metric derived from a suite of adversarial attacks, while simultaneously preserving task utility above a specified threshold, measured by semantic fidelity to baseline outputs. This black-box, optimization-driven methodology is lightweight and practical, requiring only API access to the target and optimizer LLMs. We demonstrate empirically that our optimized SHIELDs significantly reduce prompt leakage against a comprehensive set of extraction attacks, outperforming established baseline defenses without compromising the model's intended functionality. Our work presents a paradigm for developing robust, utility-aware defenses in the escalating landscape of LLM security. The code is made public on the following link: https://github.com/psm-defense/psm

## Can Activation Steering Substitute for Prompt-Based Affective Priming in LLM Digital Twins?

DevFeed: [Can Activation Steering Substitute for Prompt-Based Affective Priming in LLM Digital Twins?](<https://devfeed.tech/articles/can-activation-steering-substitute-for-prompt-based-affective-priming-in-llm-digital-twins-140073.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22738>)

Author: Junyi Fan

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

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>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Prompt Engineering](<https://devfeed.tech/topics/prompt-engineering.md>)

Tags: [activation-steering](<https://devfeed.tech/tags/activation-steering.md>), [computer-science](<https://devfeed.tech/tags/computer-science.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [digital-twins](<https://devfeed.tech/tags/digital-twins.md>), [experimental](<https://devfeed.tech/tags/experimental.md>), [gratitude](<https://devfeed.tech/tags/gratitude.md>), [llm](<https://devfeed.tech/tags/llm.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [stream](<https://devfeed.tech/tags/stream.md>)

### AI overview

In tests with Llama-3.1-8B-Instruct across 100 personas, activation steering did not uniformly substitute for prompt-based affective priming. The real-prompt effects fell between original human and digital-twin effects in 5 of 6 comparisons. Adding a mean-difference direction derived from human writing shifted the manipulation check in a dose-dependent way and matched the real-prompt effect at one strength in the gratitude domain (β ~= 2.7), but stronger injection overshot, including on the proximal measure, and effects differed by direction. A structure-matched no-injection baseline indicated that skipping the writing task raised the proximal measure in both domains, while the injected direction itself moved a distal measure in the lack-of-control domain. The findings suggest substitution depends on the measure; validity for one induction method does not transfer uniformly to another.

### Source excerpt

arXiv:2611.22738v1 Announce Type: new Abstract: Affective priming -- asking participants to write about a recalled event, such as a time they felt grateful or lacked control, and then measuring the induced state and its spillover onto related judgments -- is a standard way to manipulate emotion in the social sciences, and it also shifts the responses of LLM-based digital twins. The writing task, however, requires generating a full text. We test whether activation steering can replace it. Using Llama-3.1-8B-Instruct on 100 personas in two domains (gratitude, lack of control), we compare a real-prompt path replicating the original Twin-2K-500 Mega Study (J2) with an injection path that skips the writing and adds a mean-difference direction, extracted from human writing, to the residual stream at varying strengths. The real-prompt effects fall between the original human and digital-twin effects in 5 of 6 comparisons. Injection shifts the manipulation check in a dose-dependent way and matches the real-prompt effect at one strength in the gratitude domain (beta ~ 2.7), but overshoots beyond it, on the proximal measure as well, and asymmetrically across the two directions. A structure-matched No-Injection baseline separates two sources of difference: omitting the writing task raises the proximal measure in both domains, whereas the injected direction itself moves a distal measure in the lack-of-control domain. Substitution is therefore measure-dependent: validity established for one way of inducing a state does not carry over uniformly to another.

## SkillPE: Creativity-Oriented Cinematic Skill Evolution for Text-to-Video Prompt Enhancement

DevFeed: [SkillPE: Creativity-Oriented Cinematic Skill Evolution for Text-to-Video Prompt Enhancement](<https://devfeed.tech/articles/skillpe-creativity-oriented-cinematic-skill-evolution-for-text-to-video-prompt-enhancement-139822.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.34335>)

Author: Yanwei Huang, Mingxuan Zhu, Shujie Li, Shiyuan Liu, Yuanxing Zhang, Arpit Narechania

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

Content type: article

Language: en

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

Topics: [generative-ai-tools](<https://devfeed.tech/topics/generative-ai-tools.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>)

Tags: [creativity](<https://devfeed.tech/tags/creativity.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [enhancement](<https://devfeed.tech/tags/enhancement.md>), [framework](<https://devfeed.tech/tags/framework.md>), [generation](<https://devfeed.tech/tags/generation.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [skill](<https://devfeed.tech/tags/skill.md>), [skills](<https://devfeed.tech/tags/skills.md>), [text-to-video](<https://devfeed.tech/tags/text-to-video.md>), [video](<https://devfeed.tech/tags/video.md>), [video-generation](<https://devfeed.tech/tags/video-generation.md>)

### AI overview

SkillPE evolves reusable cinematic prompt-enhancement skills to help text-to-video prompts balance user intent with creative variety. It starts with 20 structured seed skills, then uses movie references grouped as resonators, dissonants, and divergents: the first two refine when and how skills apply, while divergents inspire controlled alternative realizations that preserve intent. Candidate skills are evaluated on generated videos for prompt fidelity, cinematic quality, narrative appeal, and creativity. On StoryEval and VBench, the abstract reports gains of up to 1.10 points over seed skills and 1.02 over the strongest external baseline on a seven-point, four-dimensional scale.

### Source excerpt

arXiv:2609.34335v2 Announce Type: replace-cross Abstract: Achieving high-quality cinematic text-to-video generation remains challenging for non-experts, whose prompts often lack professional narrative and creative design. Existing prompt enhancement (PE) methods and static cinematic skills offer limited capabilities for exploring diverse creative realizations. We propose SkillPE, a creativity-oriented PE framework that systematically evolves reusable cinematic skills through reference-guided exploration. Starting from expert-provided cinematic knowledge, SkillPE initializes and normalizes 20 structured seed skills. It then retrieves movie references categorized as resonators (good matches), dissonants (weak matches), and divergents (creatively inspiring near-misses). Resonators and dissonants refine skill applicability and execution, while divergents inspire alternative cinematic realizations through controlled mutations that preserve user intent. Candidate skills are assessed through generated videos along prompt fidelity, cinematic quality, narrative appeal, and creativity to construct the final skill libraries. Experiments on StoryEval and VBench demonstrate improvements of up to 1.10 points over seed skills and 1.02 points over the strongest external baseline on a 7-point four-dimensional evaluation scale. Overall, SkillPE offers a practical approach to balancing fidelity and creativity in cinematic text-to-video generation.

## A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt

DevFeed: [A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt](<https://devfeed.tech/articles/a-conditional-no-hair-theorem-for-transformers-how-language-models-forget-their-prompt-139839.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.24133>)

Author: Shuyang Xiang

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

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>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [text-generation](<https://devfeed.tech/topics/text-generation.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Transformers](<https://devfeed.tech/topics/transformers.md>)

Tags: [about](<https://devfeed.tech/tags/about.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [after](<https://devfeed.tech/tags/after.md>), [alpha](<https://devfeed.tech/tags/alpha.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attention](<https://devfeed.tech/tags/attention.md>), [autoregressive](<https://devfeed.tech/tags/autoregressive.md>), [bias](<https://devfeed.tech/tags/bias.md>), [black](<https://devfeed.tech/tags/black.md>), [black-hole](<https://devfeed.tech/tags/black-hole.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [charges](<https://devfeed.tech/tags/charges.md>), [conditional](<https://devfeed.tech/tags/conditional.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [models](<https://devfeed.tech/tags/models.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

For two prompts followed by the same continuation, transformer hidden-state differences decay approximately as t^-α. Trained models show faster forgetting beyond early layers (α~=0.55-0.7) than untrained models of the same architecture (α~=0.2-0.5). For finite-depth pre-LN softmax transformers with bounded attention logits and non-degenerate pre-norm activations, the authors prove an O(P(log t)^(L−1)/t) bound: growing context dilutes direct prompt influence, while relay through continuation tokens adds logarithmic factors. Conditional on the generated text, prompt influence on the next token therefore vanishes. The bound describes asymptotic behavior but is uninformative at the tested lengths in trained networks. An exactly solvable linear relay model accounts for untrained curves and shows the bound's rate cannot be improved. A prompt logit bias cancels dilution early in training; training suppresses relay by about one billion tokens, before attention sinks form.

### Source excerpt

arXiv:2610.24133v2 Announce Type: replace Abstract: A black hole retains only a few conserved charges of the matter it formed from. We are inspired to make the analogy and we ask how much an autoregressive transformer retains of its prompt after generating $t$ tokens. For two prompts followed by the same continuation, the difference of their hidden states decays approximately as a power law $t^{-\alpha}$, where $\alpha=1$ is the rate at which a fixed prompt's share of uniform attention shrinks. Beyond the first layers, trained models forget faster ($\alpha\approx0.55$--$0.7$) than untrained ones of the same architecture ($\alpha\approx0.2$--$0.5$). For finite-depth pre-LN softmax transformers with bounded attention logits and non-degenerate pre-norm activations, we prove that this offset is $O\big(P(\log t)^{L-1}/t\big)$ for prompt length $P$ and depth $L$: the prompt's direct influence is diluted as the context grows, and \emph{relay} through continuation tokens adds only logarithmic factors. Conditional on the generated text, the prompt's influence on the next token thus vanishes. The bound fixes only the asymptotic shape and is uninformative at the lengths we measure in trained networks; an exactly solvable linear relay model shows that it cannot be improved and accounts for the untrained curves. Cancelling the dilution with a logit bias on the prompt stops the decay early in training, and training suppresses relay by about one billion tokens, before attention sinks form. A toolkit that measures $\alpha$ for any Hugging Face causal language model is publicly available.

## Risk-certified prompt optimization for LLM-as-a-judge guardrails

DevFeed: [Risk-certified prompt optimization for LLM-as-a-judge guardrails](<https://devfeed.tech/articles/risk-certified-prompt-optimization-for-llm-as-a-judge-guardrails-138999.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20156>)

Author: Hussein Jawad, Alexandre Landeau, Guy Fleury Iradukunda, Paul Beaumont, Th\\'eodore Rezel, Sofiane Schaack, Nicolas J-B. Brunel

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

Content type: article

Language: en

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

Topics: [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [Prompt Engineering](<https://devfeed.tech/topics/prompt-engineering.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [certification](<https://devfeed.tech/tags/certification.md>), [confidence-bounds](<https://devfeed.tech/tags/confidence-bounds.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [disclosure](<https://devfeed.tech/tags/disclosure.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [guardrails](<https://devfeed.tech/tags/guardrails.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-as-a-judge](<https://devfeed.tech/tags/llm-as-a-judge.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [prompt-optimization](<https://devfeed.tech/tags/prompt-optimization.md>), [risk](<https://devfeed.tech/tags/risk.md>), [safety](<https://devfeed.tech/tags/safety.md>), [system-prompt-leakage](<https://devfeed.tech/tags/system-prompt-leakage.md>)

### AI overview

The framework selects system prompts for LLM-as-a-judge guardrails by maximizing benign pass rate subject to a user-set limit on missed attacks, with finite-sample risk certification. CPS generates a fixed prompt pool and uses simultaneous upper confidence bounds on a held-out set to choose the highest-utility certified prompt. CRISP uses risk feedback, hard-example-guided rewriting, and a reusable-holdout interface to adapt its search toward feasible candidates; it is particularly useful when a fixed pool yields no certifiable high-utility option. Across four datasets on prompt injection, system-prompt leakage, and sensitive-information disclosure, the authors report that certified selection respects the target missed-attack level on held-out tests, while utility rises as the allowed risk increases.

### Source excerpt

arXiv:2611.20156v1 Announce Type: cross Abstract: LLM-as-a-judge guardrails are often controlled primarily through system prompts, yet standard prompt optimization methods target average performance rather than explicit safety constraints. This paper studies system-prompt selection as a constrained optimization problem: maximize benign pass rate while keeping the missed-attack rate below a user-specified level {\alpha} with finite-sample statistical control. The framework combines LLM-based judging, automatic prompt generation, and distribution-free risk certification. This work instantiates it in two procedures: CPS and CRISP. CPS generates a fixed pool of prompts from training data, then certifies candidates on a held-out evaluation set using simultaneous upper confidence bounds and selects the certified prompt with the highest benign utility. CRISP instead integrates risk feedback into an adaptive search, using hard example-guided rewriting and a reusable-holdout interface to steer generation toward the feasible utility frontier. Experiments on four datasets covering prompt injection, system-prompt leakage, and sensitive-information disclosure show that risk-certified selection consistently respects the target missed-attack level on heldout tests, while utility improves as the permitted risk is relaxed. CRISP is especially useful when fixed-pool search fails to produce certifiable high-utility candidates.

## Do Rule-Based Tools Help LLMs Translate Endangered Uralic Languages in the Prompt or over MCP?

DevFeed: [Do Rule-Based Tools Help LLMs Translate Endangered Uralic Languages in the Prompt or over MCP?](<https://devfeed.tech/articles/do-rule-based-tools-help-llms-translate-endangered-uralic-languages-in-the-prompt-or-over-mcp-137448.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19439>)

Author: Iaroslav Chelombitko, Mika H\\"am\\"al\\"ainen

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

Content type: article

Language: en

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

Topics: [machine-translation](<https://devfeed.tech/topics/machine-translation.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [MSP MCP](<https://devfeed.tech/topics/msp-mcp.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [languages](<https://devfeed.tech/tags/languages.md>), [llms](<https://devfeed.tech/tags/llms.md>), [machine-translation](<https://devfeed.tech/tags/machine-translation.md>), [mcp](<https://devfeed.tech/tags/mcp.md>), [model-context-protocol](<https://devfeed.tech/tags/model-context-protocol.md>), [model-context-protocol-mcp](<https://devfeed.tech/tags/model-context-protocol-mcp.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [tool-calling](<https://devfeed.tech/tags/tool-calling.md>), [tools](<https://devfeed.tech/tags/tools.md>), [translation](<https://devfeed.tech/tags/translation.md>), [translations](<https://devfeed.tech/tags/translations.md>)

### AI overview

Adding rule-based resources improved translation acceptance in the expert-reviewed evaluation: accepted outputs rose from 13 to 43 of 272 when analyses were placed in the prompt, and to 55 with tools attached. The reported gain came from bilingual dictionary glosses. For Udmurt, whose dictionary lacks an English side, prompt-based access produced no gain; for the other three languages, English glosses covered 74-80% of source words. MCP tool use was harder to make work: 14 of 31 open-weight models never called a tool, including 11 unable to produce a usable call, and half of calls named unsupported languages. The evaluation covered 816 sentences across three conditions, and the authors say they release the dataset, outputs, and call logs.

### Source excerpt

arXiv:2611.19439v1 Announce Type: cross Abstract: Komi-Zyrian, Moksha, Erzya and Udmurt have mature rule-based resources, among them finite-state analysers, constraint grammars and bilingual dictionaries, which language models do not use. We ask whether giving a model access to these resources improves machine translation from the four languages into English, and whether it matters how the access is provided: as analyses computed in advance and written into the prompt, or as tools that the model calls itself through the Model Context Protocol (MCP). We evaluate manually: an expert linguist read every English output of the strongest model under three conditions, 816 sentences in all, accepted or rejected each one, and named the main error in every rejected translation. With the analyses in the prompt, the number of accepted translations rises from 13 to 43 out of 272; with the tools attached it rises to 55. The gain comes from the bilingual dictionary. For Udmurt, whose dictionary has no English side, the prompt-based condition brings no gain at all, while for the other three languages an English gloss reaches the prompt for 74 to 80% of the source words. Tool calling is also much harder to get working than prompt injection. Of 31 open-weight models, 14 never call a tool, 11 of them because they cannot produce a usable call, and half of the calls that are made name a language the toolkit does not support. We release the data set, every model output and all tool-call logs.

## Escaping Alignment: A Physical Trap Model of Best-of-N Jailbreaking

DevFeed: [Escaping Alignment: A Physical Trap Model of Best-of-N Jailbreaking](<https://devfeed.tech/articles/escaping-alignment-a-physical-trap-model-of-best-of-n-jailbreaking-137284.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.32116>)

Author: Marco Biroli

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 alignment](<https://devfeed.tech/topics/ai-alignment.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [scaling laws](<https://devfeed.tech/topics/scaling-laws.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attack-surface](<https://devfeed.tech/tags/attack-surface.md>), [computer-science](<https://devfeed.tech/tags/computer-science.md>), [cond-mat-stat-mech](<https://devfeed.tech/tags/cond-mat-stat-mech.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [predictions](<https://devfeed.tech/tags/predictions.md>), [random](<https://devfeed.tech/tags/random.md>), [safety](<https://devfeed.tech/tags/safety.md>), [safety-mechanism](<https://devfeed.tech/tags/safety-mechanism.md>), [scaling](<https://devfeed.tech/tags/scaling.md>), [temperature](<https://devfeed.tech/tags/temperature.md>)

### AI overview

The paper challenges the view that Best-of-N jailbreak success follows a fixed power law: it reports that the exponent drifts with N and that an apparent exponential crossover may reflect a finite adversarial dataset. Its physical barrier model assigns each prompt a baseline safety level and each augmentation a random, thermally activated barrier. Five interpretable parameters describe the attack surface across augmentation count N, completions per augmentation M, and generation temperature T. The authors report extrapolations from N ≤ 100 to N = 10⁴, a shared scaling function across five models, and temperature predictions from a fit at one temperature; they say these results hold with two automated judges and on held-out prompts.

### Source excerpt

arXiv:2609.32116v2 Announce Type: replace Abstract: Best-of-$N$ jailbreaking (BoN) bypasses safeguards of aligned models by drawing $N$ independent augmentations of an unsafe prompt and sampling $M$ completions of each. Previous works have shown that the attack success rate (ASR) seems to follow a power-law in $N$, which we challenge. The exponent drifts with $N$, and the reported exponential crossover is consistent with a finite-size effect of the adversarial dataset. Little work has been done to explore the entire two-budget ($N, M$) attack surface as well as its dependence on the generation temperature $T$. We introduce a simple barrier model where each prompt has a baseline safety level and each augmentation a random thermally activated barrier. Then five numbers, each backed by an interpretable safety mechanism, determine the entire ($N, M$) attack surface. They extrapolate predictions from $N \leq 100$ to $N = 10^4$, collapse five distinct models on the same scaling function and predict ASR at different temperatures from the one they were fitted at. The extrapolation holds under two automated judges and on held-out prompts.

## Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits

DevFeed: [Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits](<https://devfeed.tech/articles/interpreting-and-controlling-model-behavior-via-constitutions-for-atomic-concept-edits-138737.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2602.00092>)

Author: Neha Kalibhat, Zi Wang, Prasoon Bajpai, Drew Proud, Wenjun Zeng, Been Kim, Mani Malek

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

Content type: article

Language: en

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

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [concept-learning](<https://devfeed.tech/topics/concept-learning.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [atomic](<https://devfeed.tech/tags/atomic.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [black-box](<https://devfeed.tech/tags/black-box.md>), [boost](<https://devfeed.tech/tags/boost.md>), [coherence](<https://devfeed.tech/tags/coherence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.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>), [edits](<https://devfeed.tech/tags/edits.md>), [framework](<https://devfeed.tech/tags/framework.md>), [gemini](<https://devfeed.tech/tags/gemini.md>), [generation](<https://devfeed.tech/tags/generation.md>), [gpt](<https://devfeed.tech/tags/gpt.md>), [gpt-5](<https://devfeed.tech/tags/gpt-5.md>), [gpt-image](<https://devfeed.tech/tags/gpt-image.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [input](<https://devfeed.tech/tags/input.md>), [insights](<https://devfeed.tech/tags/insights.md>), [interpretability](<https://devfeed.tech/tags/interpretability.md>), [language](<https://devfeed.tech/tags/language.md>), [mapping](<https://devfeed.tech/tags/mapping.md>), [model](<https://devfeed.tech/tags/model.md>)

### AI overview

The framework uses atomic concept edits (ACEs)--adding, removing, or replacing an interpretable prompt concept--and observes resulting behavior to learn a natural-language "constitution" mapping edits to outcomes. The authors report that GPT-Image emphasizes grammatical adherence while Imagen 4 prioritizes atmospheric coherence in text-to-image generation; distractor variables confuse GPT-5 in mathematical reasoning but leave Gemini 2.5 models and o4-mini largely unaffected. Across the reported tasks, constitution-guided control achieved an average 1.86x boost in success rate over methods without constitutions. These findings are limited to the tasks and comparisons described in the abstract.

### Source excerpt

arXiv:2602.00092v2 Announce Type: replace Abstract: We introduce a black-box interpretability framework that learns a verifiable constitution: a natural language summary of how changes to a prompt affect a model's specific behavior, such as its alignment, correctness, or adherence to constraints. Our method leverages atomic concept edits (ACEs), which are targeted operations that add, remove, or replace an interpretable concept in the input prompt. By systematically applying ACEs and observing the resulting effects on model behavior across various tasks, our framework learns a causal mapping from edits to predictable outcomes. This learned constitution provides deep, generalizable insights into the model. Empirically, we validate our approach across diverse tasks, including mathematical reasoning and text-to-image alignment, for controlling and understanding model behavior. We found that for text-to-image generation, GPT-Image tends to focus on grammatical adherence, while Imagen 4 prioritizes atmospheric coherence. In mathematical reasoning, distractor variables confuse GPT-5 but leave Gemini 2.5 models and o4-mini largely unaffected. Moreover, our results show that the learned constitutions are highly effective for controlling model behavior, achieving an average of 1.86 times boost in success rate over methods that do not use constitutions. Our open-sourced code can be found at https://github.com/google-deepmind/atomic_concept_edits

## Mem2Fair: Online Fairness Control for Text-to-Video Generation via Fair Memory

DevFeed: [Mem2Fair: Online Fairness Control for Text-to-Video Generation via Fair Memory](<https://devfeed.tech/articles/mem2fair-online-fairness-control-for-text-to-video-generation-via-fair-memory-137416.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19215>)

Author: Pengyu Chen, Jiaming Zhang, Xiaohua Feng, Yuyuan Li, Chaochao Chen

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

Content type: article

Language: en

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

Topics: [generative-ai-tools](<https://devfeed.tech/topics/generative-ai-tools.md>), [responsible-ai](<https://devfeed.tech/topics/responsible-ai.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [multimodal-ai](<https://devfeed.tech/topics/multimodal-ai.md>), [prompt](<https://devfeed.tech/topics/prompt.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [black-box](<https://devfeed.tech/tags/black-box.md>), [conditional](<https://devfeed.tech/tags/conditional.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [deployment](<https://devfeed.tech/tags/deployment.md>), [distributions](<https://devfeed.tech/tags/distributions.md>), [effectiveness](<https://devfeed.tech/tags/effectiveness.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [first](<https://devfeed.tech/tags/first.md>), [framework](<https://devfeed.tech/tags/framework.md>), [gender](<https://devfeed.tech/tags/gender.md>), [generate](<https://devfeed.tech/tags/generate.md>), [memory](<https://devfeed.tech/tags/memory.md>), [online](<https://devfeed.tech/tags/online.md>), [text-to-video](<https://devfeed.tech/tags/text-to-video.md>), [video](<https://devfeed.tech/tags/video.md>), [video-generation](<https://devfeed.tech/tags/video-generation.md>)

### AI overview

Mem2Fair addresses demographic imbalance in text-to-video generation when prompts leave protected attributes unspecified. It keeps a concept-conditioned "Fair Memory" and uses videos generated during inference as feedback, selecting prompt-level attribute interventions according to deviation from each concept's prescribed fairness target. The authors frame this as online fairness control that requires neither model training nor separate pre-deployment sampling, avoiding the architecture access required by training-based methods and the advance video-generation cost of sampling-based methods. Experiments report improved fairness over the unmodified generator and existing baselines across demographic attributes, prompt settings, and target distributions, with comparable prompt-video alignment and video quality; the abstract also reports results on commercial black-box models.

### Source excerpt

arXiv:2611.19215v1 Announce Type: new Abstract: Text-to-video (T2V) generation models can produce systematically unfair demographic outcomes when prompts leave protected attributes unspecified. For example, prompts describing nurses without specifying gender may disproportionately generate videos depicting female nurses. Mitigating unfairness in T2V generation is particularly challenging: training-based methods that modify model internals require architecture-specific access and risk disrupting temporal consistency, while sampling-based methods often incur costly pre-deployment video generation. We present the first conditional online approachability formulation of attribute-level fairness for T2V generation. Building on this formulation, we introduce Mem2Fair, a plug-and-play framework that maintains concept-conditioned Fair Memory. Using generated videos at inference time as feedback, Mem2Fair selects prompt-level attribute interventions based on the deviation from each concept's prescribed fairness target, without training or separate pre-deployment sampling. Experiments demonstrate that Mem2Fair consistently improves fairness over the unmodified generator and existing baselines across multiple demographic attributes, prompt settings, and target distributions, while maintaining comparable prompt-video alignment and video quality. We further demonstrate its effectiveness on commercial black-box T2V models.

## LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL

DevFeed: [LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL](<https://devfeed.tech/articles/llm-as-a-tutor-policy-aware-prompt-adaptation-for-non-verifiable-rl-137263.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2607.04412>)

Author: Yujin Kim, Namgyu Ho, Sangmin Hwang, Joonkee Kim, Yongjin Yang, Sangmin Bae, Seungone Kim, Jaehun Jung, Se-Young Yun, Hwanjun Song

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

Content type: article

Language: en

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

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [atomic](<https://devfeed.tech/tags/atomic.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [capability](<https://devfeed.tech/tags/capability.md>), [constraint](<https://devfeed.tech/tags/constraint.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [design](<https://devfeed.tech/tags/design.md>), [external](<https://devfeed.tech/tags/external.md>), [framework](<https://devfeed.tech/tags/framework.md>), [generator](<https://devfeed.tech/tags/generator.md>), [llm](<https://devfeed.tech/tags/llm.md>), [misalignment](<https://devfeed.tech/tags/misalignment.md>), [model](<https://devfeed.tech/tags/model.md>), [original](<https://devfeed.tech/tags/original.md>), [policy](<https://devfeed.tech/tags/policy.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [quality](<https://devfeed.tech/tags/quality.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [results](<https://devfeed.tech/tags/results.md>), [rl](<https://devfeed.tech/tags/rl.md>), [tutor](<https://devfeed.tech/tags/tutor.md>)

### AI overview

LLM-as-a-Tutor addresses a mismatch between fixed training-prompt difficulty and a changing policy in non-verifiable instruction-following RL, where prompts may not produce enough rollout-quality variation for an LLM judge to provide discriminative rewards. One model pairwise-compares rollouts to identify prompts that are no longer challenging, then appends atomic constraints while preserving each original prompt. This raises difficulty with policy capability and creates a self-calibrating signal without external difficulty schedules. Across three complex instruction-following benchmarks, the authors report better average results and higher scores on four of five metrics than policy-unaware baselines and policy-adaptive methods that adjust rubrics or rewrite prompts.

### Source excerpt

arXiv:2607.04412v3 Announce Type: replace Abstract: Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals. While recent methods adapt these rubrics to the evolving policy during training, the training prompts themselves remain static, drawn from fixed corpora. This static approach often results in a critical misalignment between prompt difficulty and policy capability, leaving the judge unable to recover a discriminative reward signal when prompts fail to elicit quality variance among rollouts. To address this misalignment, we introduce LLM-as-a-Tutor, a framework that extends the LLM's role from judge to tutor: a single model serves as an examiner that pairwise-compares policy rollouts to detect non-challenging prompts, and as a generator that appends atomic constraints to them. This append-only design preserves the original prompt while adding constraints to increase difficulty in step with the policy's capability, producing a self-calibrating training signal without external difficulty schedules. On three complex instruction-following benchmarks, our method outperforms both policy-unaware baselines and prior policy-adaptive methods that adapt rubrics or rewrite prompts, on average and on four of five metrics, suggesting prompt adaptation as an underexplored axis of policy-awareness in non-verifiable RL for constraint-heavy instruction following.

## AI Agents May Always Fall for Prompt Injections

DevFeed: [AI Agents May Always Fall for Prompt Injections](<https://devfeed.tech/articles/ai-agents-may-always-fall-for-prompt-injections-139230.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.17634>)

Author: Sahar Abdelnabi, Eugene Bagdasarian

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

Content type: article

Language: en

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

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

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [attack](<https://devfeed.tech/tags/attack.md>), [ci](<https://devfeed.tech/tags/ci.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [defense](<https://devfeed.tech/tags/defense.md>), [privacy](<https://devfeed.tech/tags/privacy.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [research](<https://devfeed.tech/tags/research.md>), [vulnerability](<https://devfeed.tech/tags/vulnerability.md>)

### AI overview

The paper argues that defenses based on separating data from instructions miss prompt injections that manipulate context and can also block appropriate behavior. Using Contextual Integrity (CI), which evaluates information flows against contextual norms, it presents attacks that misrepresent a flow, manipulate norms, or mix flows. The authors argue that an attacker can make a blocked flow appear legitimate in a constructed context, while tightening the rules can block legitimate flows. They propose CI as a framework for evaluating context-sensitive failures and designing alignment for autonomous agents. This overview is based on the supplied abstract.

### Source excerpt

arXiv:2605.17634v2 Announce Type: replace Abstract: Prompt injection is the most critical vulnerability in deployed AI agents. Despite recent progress, we show that the prevailing defense paradigm (data-instruction separation) both fails to detect attacks that operate through contextual manipulation and degrades contextually appropriate behavior. We then recast prompt injection via the lens of Contextual Integrity (CI), a privacy theory that judges information flow compliance with contextual norms. This explains types of attacks that current defenses attempt to patch and predict advanced ones future agents will face. We develop unique benign and attack scenarios that force an agent to violate the norms by (1) misrepresenting the flow, (2) manipulating norms, or (3) mixing multiple flows. This reframing suggests an impossibility argument: an adversary can always construct a context under which a blocked flow appears legitimate, or a defender who tightens norms will block genuinely legitimate flows. Our findings suggest that current research addresses a shrinking fraction of future attack surfaces. Instead, through CI, we offer a principled framework for evaluating context-sensitive failures, and designing CI-aware alignment for frontier autonomous agents.

## Use prompt caching with open-weight models on Amazon Bedrock

DevFeed: [Use prompt caching with open-weight models on Amazon Bedrock](<https://devfeed.tech/articles/use-prompt-caching-with-open-weight-models-on-amazon-bedrock-136125.md>)

Original publisher: [Read original article](<https://aws.amazon.com/blogs/machine-learning/use-prompt-caching-with-open-weight-models-on-amazon-bedrock/>)

Author: Dustin Liu

Published: 2026-11-16T15:42:41Z

Content type: tutorial

Language: en

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

Topics: [Amazon Bedrock](<https://devfeed.tech/topics/amazon-bedrock.md>), [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>), [prompt](<https://devfeed.tech/topics/prompt.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [amazon-bedrock](<https://devfeed.tech/tags/amazon-bedrock.md>), [api](<https://devfeed.tech/tags/api.md>), [best-practices](<https://devfeed.tech/tags/best-practices.md>), [caching](<https://devfeed.tech/tags/caching.md>), [coding](<https://devfeed.tech/tags/coding.md>), [coding-agent](<https://devfeed.tech/tags/coding-agent.md>), [cost](<https://devfeed.tech/tags/cost.md>), [intermediate-200](<https://devfeed.tech/tags/intermediate-200.md>), [latency](<https://devfeed.tech/tags/latency.md>), [long-context](<https://devfeed.tech/tags/long-context.md>), [open-weight-models](<https://devfeed.tech/tags/open-weight-models.md>), [opencode](<https://devfeed.tech/tags/opencode.md>), [prompt-caching](<https://devfeed.tech/tags/prompt-caching.md>), [technical-how-to](<https://devfeed.tech/tags/technical-how-to.md>)

### AI overview

Amazon Bedrock prompt caching can reduce input cost and latency by reusing matching prompt prefixes for open-weight models. Implicit caching works automatically on a best-effort basis; explicit caching uses breakpoints to control which reusable content is stored and avoid writing changing tails that may never be reused. Cache reads cost less than standard input, while some models charge more for writes. The examples show how to configure both modes through the Chat Completions API and verify cache usage in responses, then use implicit caching with OpenCode, which resends its system prompt, tools, and conversation history. Results depend on model and API support, minimum prefix lengths, and cache TTL. The post recommends keeping static content first, checking usage fields and runtime metrics, and keeping frequently reused prefixes warm.

### Source excerpt

Many generative AI applications resend the same long context on every request. Prompt caching lets open-weight models on Amazon Bedrock reuse it to reduce input cost and latency. This post shows implicit and explicit caching through the Chat Completions API and applies caching to an OpenCode coding agent session.

## The Transformer Paradigm: Dynamic Processing and a New Principle of Inference

DevFeed: [The Transformer Paradigm: Dynamic Processing and a New Principle of Inference](<https://devfeed.tech/articles/the-transformer-paradigm-dynamic-processing-and-a-new-principle-of-inference-135154.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17953>)

Author: Marco Giunti, Fabrizia Giulia Garavaglia

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

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

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Transformers](<https://devfeed.tech/topics/transformers.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ar](<https://devfeed.tech/tags/cs-ar.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [cs-et](<https://devfeed.tech/tags/cs-et.md>), [cs-ne](<https://devfeed.tech/tags/cs-ne.md>), [deep-learning](<https://devfeed.tech/tags/deep-learning.md>), [inference](<https://devfeed.tech/tags/inference.md>), [neuromorphic](<https://devfeed.tech/tags/neuromorphic.md>), [parallel](<https://devfeed.tech/tags/parallel.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [token](<https://devfeed.tech/tags/token.md>), [transformer](<https://devfeed.tech/tags/transformer.md>)

### AI overview

The authors frame Transformers as a distinct inference paradigm in which some parameterized transformations are dynamically constructed from the current input. Previously acquired transformations are adapted and then applied; structurally, outputs from some networks determine the weights of others. They call this Sequence-level Interactive Dynamic Parallel Processing (SIDPP) and propose it as a framework for explaining strong prompt sensitivity. The abstract presents interpretability, control, alternative architectures, neuromorphic systems, and possible biological realizations as research directions.

### Source excerpt

arXiv:2611.17953v1 Announce Type: new Abstract: We argue that the Transformer revolution is not adequately understood as a merely quantitative expansion of deep learning, but as the emergence of a new paradigm based on a different principle of inference. We interpret the Transformer as a system in which token representations are transformed by parameterized transformations, or transforming concepts. During inference, some of these transformations are dynamically constructed from the current input: previously acquired transformations are first adapted and then applied. At the mechanical-structural level, this dynamic processing is realized through output-weight interconnections, whereby outputs of some networks determine the weights of others. The resulting mode of processing, which we call Sequence-level Interactive Dynamic Parallel Processing (SIDPP), provides a framework for explaining strong prompt sensitivity and opens research directions concerning interpretability, control, alternative AI architectures, neuromorphic systems, and possible biological realizations of dynamic processing.

## Creating Steering Vectors via Input Optimisation

DevFeed: [Creating Steering Vectors via Input Optimisation](<https://devfeed.tech/articles/creating-steering-vectors-via-input-optimisation-135782.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17481>)

Author: Batuhan Akar

Published: 2026-11-16T05:00:00Z

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

Tags: [activation](<https://devfeed.tech/tags/activation.md>), [activation-steering](<https://devfeed.tech/tags/activation-steering.md>), [behaviour](<https://devfeed.tech/tags/behaviour.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [generation](<https://devfeed.tech/tags/generation.md>), [input](<https://devfeed.tech/tags/input.md>), [language](<https://devfeed.tech/tags/language.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [models](<https://devfeed.tech/tags/models.md>), [optimisation](<https://devfeed.tech/tags/optimisation.md>), [prompts](<https://devfeed.tech/tags/prompts.md>), [steering](<https://devfeed.tech/tags/steering.md>), [vectors](<https://devfeed.tech/tags/vectors.md>), [white-box](<https://devfeed.tech/tags/white-box.md>)

### AI overview

ActGCG approximates activation steering by optimizing an input suffix to match the target residual-stream activations, so the resulting behavior can be elicited without modifying model internals during generation. The authors also describe universal suffixes trained across prompts, a step that translates suffixes into human-readable text during optimization, and soft suffixes that avoid the discrete-token constraint. In their evaluations, discrete suffixes nearly reproduced steering for answering in German and transferred across prompts and models. Anti-refusal steering was less equivalent: it succeeded on a phrase-avoidance metric but only partially under the stricter HarmBench classifier. The reported match therefore depends on the behavior being steered.

### Source excerpt

arXiv:2611.17481v1 Announce Type: cross Abstract: Activation steering influences a language model's behaviour by adding a steering vector to its residual stream during generation, but it requires white-box access to modify the hidden states at every generation step. We introduce ActGCG, a modification of Greedy Coordinate Gradient which optimises a suffix to match the target activations obtained by adding a steering vector to the residual stream. We further introduce universal suffixes trained across many prompts at once, a model-translation step that rewrites the suffix into human-readable text midway through optimisation, and a soft-suffix variant to remove the discrete token space constraint. We study two behaviours: steering a model to answer in German, and anti-refusal steering for jailbreaking. For German, a discrete suffix reproduces the steering vector's behaviour almost perfectly and transfers well across prompts and models. For anti-refusal, the equivalence is weaker, succeeding under a phrase-avoidance metric but only partially under the stricter HarmBench classifier. Our results show that a steering vector's behaviour is reachable from the input, with no access to model internals at generation time, up to a loss that depends on the concept being steered.

## Seedance 2.5 Flash now available on AI Gateway

DevFeed: [Seedance 2.5 Flash now available on AI Gateway](<https://devfeed.tech/articles/seedance-2-5-flash-now-available-on-ai-gateway-136323.md>)

Original publisher: [Read original article](<https://vercel.com/changelog/seedance-2-5-flash-now-available-on-ai-gateway>)

Author: Jerilyn Zheng

Published: 2026-11-16T00:00:00Z

Content type: release

Language: en

Sources: [Vercel News](<https://devfeed.tech/sources/vercel-news.md>)

Topics: [generative-ai-tools](<https://devfeed.tech/topics/generative-ai-tools.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [multimodal-ai](<https://devfeed.tech/topics/multimodal-ai.md>), [prompt](<https://devfeed.tech/topics/prompt.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-gateway](<https://devfeed.tech/tags/ai-gateway.md>), [ai-sdk](<https://devfeed.tech/tags/ai-sdk.md>), [also](<https://devfeed.tech/tags/also.md>), [audio](<https://devfeed.tech/tags/audio.md>), [based](<https://devfeed.tech/tags/based.md>), [bytedance](<https://devfeed.tech/tags/bytedance.md>), [can](<https://devfeed.tech/tags/can.md>), [cost](<https://devfeed.tech/tags/cost.md>), [existing](<https://devfeed.tech/tags/existing.md>), [extend](<https://devfeed.tech/tags/extend.md>), [faster](<https://devfeed.tech/tags/faster.md>), [first](<https://devfeed.tech/tags/first.md>), [flash](<https://devfeed.tech/tags/flash.md>), [follow](<https://devfeed.tech/tags/follow.md>), [formats](<https://devfeed.tech/tags/formats.md>), [gateway](<https://devfeed.tech/tags/gateway.md>), [seedance](<https://devfeed.tech/tags/seedance.md>), [seedance-2-5](<https://devfeed.tech/tags/seedance-2-5.md>), [seedance-2-5-flash](<https://devfeed.tech/tags/seedance-2-5-flash.md>)

### AI overview

ByteDance's Seedance 2.5 Flash is available on AI Gateway as a faster, lower-cost version of Seedance 2.5, with output described as close to the standard model in realism, reference-based editing, and prompt adherence. It generates video with synchronized audio from text or image, video, and audio references; supports 4-30-second clips up to 720p at 24 fps in landscape, square, or portrait formats; and can edit or extend videos or use images to set the first and last frames. The excerpt's AI SDK 7 examples use `bytedance/seedance-2.5-flash`; reference URLs go in `inputReferences`, with image and video prompt labels numbered separately.

### Source excerpt

ByteDance's Seedance 2.5 Flash is now available on AI Gateway. Seedance 2.5 Flash is a faster, lower-cost version of Seedance 2.5. The output is close to the standard model in human realism, reference-based editing, and prompt adherence. Generate videos with synchronized audio from text prompts or image, video, and audio references. Clips can be 4 to 30 seconds long at up to 720p and 24 frames per second, in landscape, square, or portrait formats. You can also edit or extend existing videos, or provide images to set the first and last frames. Use bytedance/seedance-2.5-flash with AI SDK 7 to generate a video with audio from a text prompt: To guide generation with reference images and videos, pass their URLs in inputReferences. Images and videos are numbered separately in the prompt as [Image 1], [Video 1], and so on: Try Seedance 2.5 Flash in the model playground, or follow the video generation quickstart. Read more

## Riley Brown's 4K film: 89 rented GPUs, $20

DevFeed: [Riley Brown's 4K film: 89 rented GPUs, $20](<https://devfeed.tech/articles/riley-brown-s-4k-film-89-rented-gpus-20-135260.md>)

Original publisher: [Read original article](<https://jevdaily.com/issues/2026-11-15>)

Author: Jev Daily

Published: 2026-11-15T17:00:00Z

Content type: news

Language: en

Sources: [Jev Daily](<https://devfeed.tech/sources/jev-daily.md>)

Topics: [blender](<https://devfeed.tech/topics/blender.md>), [3D](<https://devfeed.tech/topics/3d.md>), [Python](<https://devfeed.tech/topics/python.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [GPUDirect](<https://devfeed.tech/topics/gpudirect.md>)

Tags: [4k](<https://devfeed.tech/tags/4k.md>), [agent](<https://devfeed.tech/tags/agent.md>), [blender](<https://devfeed.tech/tags/blender.md>), [built-with-it](<https://devfeed.tech/tags/built-with-it.md>), [fable-5-2](<https://devfeed.tech/tags/fable-5-2.md>), [film](<https://devfeed.tech/tags/film.md>), [gpus](<https://devfeed.tech/tags/gpus.md>), [parallel](<https://devfeed.tech/tags/parallel.md>), [project](<https://devfeed.tech/tags/project.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [python](<https://devfeed.tech/tags/python.md>), [questions](<https://devfeed.tech/tags/questions.md>), [render](<https://devfeed.tech/tags/render.md>), [someone-measured-it](<https://devfeed.tech/tags/someone-measured-it.md>), [state](<https://devfeed.tech/tags/state.md>), [weekend-project](<https://devfeed.tech/tags/weekend-project.md>)

### AI overview

Riley Brown says a single Fable 5.2 prompt generated 12,500 lines of Blender Python for a 2-minute 4K film with 5 characters, 7 sets and 52 shots. Rendering 3,220 frames used 89 rented GPUs and cost a little over $20.

### Source excerpt

Riley Brown says one Fable 5.2 prompt wrote 12,500 lines of Blender Python, then rented 89 GPUs to render 3,220 frames for a little over $20. Also: Allie the...

## Small Language Models for Inference-Capability Routing: Effects of Model Capacity and Prompt Optimization on Downstream Quality and Efficiency

DevFeed: [Small Language Models for Inference-Capability Routing: Effects of Model Capacity and Prompt Optimization on Downstream Quality and Efficiency](<https://devfeed.tech/articles/small-language-models-for-inference-capability-routing-effects-of-model-capacity-and-prompt-optimization-on-downstream-quality-and-efficiency-132231.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.16414>)

Author: Sailendra Kolluru, Maheshkumar Duvvarapu

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

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

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [capacity](<https://devfeed.tech/tags/capacity.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [inference](<https://devfeed.tech/tags/inference.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [latency](<https://devfeed.tech/tags/latency.md>), [prompt-optimization](<https://devfeed.tech/tags/prompt-optimization.md>), [quality](<https://devfeed.tech/tags/quality.md>), [routing](<https://devfeed.tech/tags/routing.md>), [small-language-models](<https://devfeed.tech/tags/small-language-models.md>), [token](<https://devfeed.tech/tags/token.md>)

### AI overview

The study tests prompt-driven Qwen2.5 routers (1.5B-14B parameters) that independently select reasoning, retrieval, web search, and tool use, allowing multiple capabilities per query. Across five prompting strategies, added capacity does not consistently improve classification under minimal prompts. Prompt optimization narrows the capacity gap for reasoning and calculator routing, while RAG and web search retain larger benefits from higher capacity. On 127 questions using GPT-OSS-120B, the 14B router's mean BERTScore F1 (0.8408) is close to enabling all capabilities (0.8418); it cuts average downstream tokens by 68.7% and median downstream latency by 80.5%. These savings exclude router overhead, which was measured separately in a different inference environment and increases with router capacity. The abstract cautions that similar classification F1 can mask different error costs: false positives waste computation, while missed capabilities can lower answer quality.

### Source excerpt

arXiv:2611.16414v1 Announce Type: new Abstract: Optional reasoning, retrieval, web search, and tool use can improve large language model responses but incur unnecessary computation when activated indiscriminately. This study investigates whether prompt optimization can compensate for limited router capacity and whether classification gains translate into downstream efficiency. The proposed framework uses prompt-driven small language models to make four independent capability decisions, allowing multiple capabilities to be selected per query. Qwen2.5 routers ranging from 1.5B to 14B parameters are compared across five prompting strategies, followed by an evaluation of capability-specific optimized configurations using a fixed GPT-OSS-120B downstream model on 127 questions. Under minimal prompting, increasing router capacity does not consistently improve classification performance. Prompt optimization narrows the capacity gap for reasoning and calculator routing, whereas RAG and web search retain larger benefits from higher capacity. Similar classification F1 scores do not necessarily imply equivalent downstream behavior: false-positive activations increase unnecessary computation, while missed capabilities can reduce answer quality. The 14B router achieves a mean BERTScore F1 of 0.8408, compared with 0.8418 when all capabilities are enabled, while reducing average downstream token consumption by 68.7\% (8,837 to 2,764 tokens) and median downstream latency by 80.5\% (9.79 to 1.91 seconds). These savings exclude router overhead, which is measured separately in a different inference environment and increases with router capacity. The findings demonstrate that the relationship between router capacity, prompt design, and downstream efficiency is capability-dependent. Consequently, router selection must consider error profiles and routing overhead alongside classification performance.

## OverServe: Resource Amplification Attacks on SaaS Agents via Tool--Prompt Co-Design

DevFeed: [OverServe: Resource Amplification Attacks on SaaS Agents via Tool--Prompt Co-Design](<https://devfeed.tech/articles/overserve-resource-amplification-attacks-on-saas-agents-via-tool-prompt-co-design-131696.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.14808>)

Author: Ruiheng Yu, Oteo Mamo, Siyu Gan, Weikuan Yu, Mulong Luo

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>), [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [ai-assisted attacks](<https://devfeed.tech/topics/ai-assisted-attacks.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [agentic workflows](<https://devfeed.tech/topics/agentic-workflows.md>), [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [rate-limiting](<https://devfeed.tech/topics/rate-limiting.md>), [Use Cases for Every Cloud Service Model](<https://devfeed.tech/topics/use-cases-for-every-cloud-service-model.md>), [API Business Models](<https://devfeed.tech/topics/api-business-models.md>), [Nemotron](<https://devfeed.tech/topics/nemotron.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [MSP MCP](<https://devfeed.tech/topics/msp-mcp.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-framework](<https://devfeed.tech/tags/agent-framework.md>), [agents](<https://devfeed.tech/tags/agents.md>), [alerts](<https://devfeed.tech/tags/alerts.md>), [api](<https://devfeed.tech/tags/api.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [co-design](<https://devfeed.tech/tags/co-design.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [deployments](<https://devfeed.tech/tags/deployments.md>), [execution](<https://devfeed.tech/tags/execution.md>), [framework](<https://devfeed.tech/tags/framework.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [resource](<https://devfeed.tech/tags/resource.md>), [saas](<https://devfeed.tech/tags/saas.md>), [tool](<https://devfeed.tech/tags/tool.md>)

### AI overview

OverServe is a framework for amplifying resource use by SaaS agents through co-designed MCP status tools and prompts. The tools return accurate status and end normally, while prompts can lead an agent to keep checking an unfinished condition after its original business task is complete. The authors report that tested tool specifications triggered no alerts in the evaluated MCP scanners. On selected tasks from nine benchmarks, Qwen, Nemotron, DeepSeek, and Sonnet runs passed the original task checks; the maximum reported API request amplification was 93.88x while preserving the task goal and collateral checks.

### Source excerpt

arXiv:2611.14808v1 Announce Type: cross Abstract: Advances in large language models (LLMs) are enabling software-as-a-service (SaaS) providers to automate longer business workflows with agents. In subscription-based SaaS deployments, model and tool calls consume provider resources without additional customer payments. Because the agent autonomously decides when to stop, a malicious tenant can send a prompt to the agent, exploiting its completion decisions to extend execution and waste provider resources. In this paper, we present OverServe, a multi-agent framework for constructing such attacks against SaaS agents. OverServe co-designs Model Context Protocol (MCP) status tools and completion-confirmation prompts for the attacker. Each tool call returns accurate status and terminates normally, yet the agent may repeatedly check an unfinished condition to satisfy the prompt. This consumes the SaaS provider's resources even after the original business task is completed. OverServe constructs service-specific tool--prompt pairs and uses separate tool and prompt reviews to guide candidate pair selection. Tested tool specifications trigger no alerts in the evaluated MCP scanners. We evaluate Qwen, Nemotron, DeepSeek, and Sonnet on selected tasks from nine benchmarks. All OverServe runs pass the original task checks. OverServe can reach a maximal amplification of API requests by $93.88\times$ while preserving the original task goal and collateral checks.

## From Noise to Diversity: Random Embedding Injection in LLM Reasoning

DevFeed: [From Noise to Diversity: Random Embedding Injection in LLM Reasoning](<https://devfeed.tech/articles/from-noise-to-diversity-random-embedding-injection-in-llm-reasoning-131912.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.11936>)

Author: Heejun Kim, Seungpil Lee, Jewon Yeom, Jaewon Sok, Seonghyeon Park, Jeongjae Park, Taesup Kim, Sundong Kim

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

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

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Parallel reasoning](<https://devfeed.tech/topics/parallel-reasoning.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Temperature settings in AI](<https://devfeed.tech/topics/temperature-settings-in-ai.md>), [grpo](<https://devfeed.tech/topics/grpo.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [inference](<https://devfeed.tech/tags/inference.md>), [llm](<https://devfeed.tech/tags/llm.md>), [noise](<https://devfeed.tech/tags/noise.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [random](<https://devfeed.tech/tags/random.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

Random Soft Prompts (RSPs) append a freshly sampled sequence of random embedding vectors to an LLM input, avoiding prompt training while reaching accuracy comparable to optimized soft prompts on math reasoning benchmarks in several settings. The vectors are drawn from an isotropic Gaussian fitted to the pretrained embedding table's entrywise mean and variance. The authors propose that attention to the unfamiliar positions flattens the distribution over early generated tokens, branching reasoning trajectories; that effect then dilutes as generation continues. At inference, RSPs increase early token diversity and, with temperature sampling, widen Pass@N--the chance that at least one of N attempts is correct. The authors also report practical gains when applying the effect in DAPO training. The supplied text does not specify benchmark names or numerical results.

### Source excerpt

arXiv:2605.11936v3 Announce Type: replace Abstract: Recent soft prompt research has tried to improve reasoning by inserting trained vectors into LLM inputs, yet whether the gain comes from the learned content or from the act of injection itself has not been carefully separated. We study Random Soft Prompts (RSPs), which drop the training step entirely and append a freshly drawn sequence of random embedding vectors to the input. Each RSP vector is sampled from an isotropic Gaussian fitted to the entrywise mean and variance of the pretrained embedding table; the sequence carries no learned content, and yet reaches accuracy comparable to optimized soft prompts on math reasoning benchmarks in several settings. The mechanism unfolds in two stages: because attention has to absorb a never-seen-before random position, the distribution over the first few generated tokens flattens and reasoning trajectories branch, and as generation continues this influence dilutes naturally so the response commits to a single completion. We show that during inference RSPs lift early-stage token diversity and, combined with temperature sampling, widen Pass@N, the probability that at least one out of N attempts is correct. Beyond inference, we carry the same effect into DAPO training and demonstrate practical gains. Our contributions are: (i) RSP isolates the simplest form of soft prompt -- training-free, freshly resampled -- providing a unified lens for the structural effect of injection that variants otherwise differing in training and form all share; (ii) a theoretical and empirical validation of the underlying mechanism; and (iii) an extension from inference to training.

## How MiniMax-H3-Image Handles High-Resolution AI Image Generation

DevFeed: [How MiniMax-H3-Image Handles High-Resolution AI Image Generation](<https://devfeed.tech/articles/how-minimax-h3-image-handles-high-resolution-ai-image-generation-130189.md>)

Original publisher: [Read original article](<https://hackernoon.com/how-minimax-h3-image-handles-high-resolution-ai-image-generation?source=rss>)

Author: aimodels44

Published: 2026-11-12T06:56:47Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [image editing](<https://devfeed.tech/topics/image-editing.md>), [diffusion-transformers](<https://devfeed.tech/topics/diffusion-transformers.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [Image](<https://devfeed.tech/topics/image.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Image processing](<https://devfeed.tech/topics/image-processing.md>)

Tags: [ai-engineering](<https://devfeed.tech/tags/ai-engineering.md>), [api](<https://devfeed.tech/tags/api.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [customer-success](<https://devfeed.tech/tags/customer-success.md>), [image-editing](<https://devfeed.tech/tags/image-editing.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [minimax](<https://devfeed.tech/tags/minimax.md>), [minimax-h3](<https://devfeed.tech/tags/minimax-h3.md>), [minimax-h3-image](<https://devfeed.tech/tags/minimax-h3-image.md>), [performance](<https://devfeed.tech/tags/performance.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

MiniMax-H3-Image is a unified model for text-to-image generation and editing one or more reference images, with support documented for up to 16 references and output typically targeted around 16 megapixels (effective range about 4-20 MP). The requested WIDTHxHEIGHT determines the returned PNG dimensions. Editing prompts identify images by upload order. Sampling is fixed at 6 steps without classifier-free guidance; negative prompts are unsupported, and prompts are truncated at 4096 tokens. The checkpoint requires a separate MiniMax-H3 text encoder, and its listed size is 68.65 GB (encoder: 66.7 GB). No minimum hardware requirement or independent image-quality/performance benchmarks are provided. The optional prompt expander is a separate version and is not required.

### Source excerpt

MiniMax-H3-Image combines text-to-image generation and reference-image editing. Explore its architecture, capabilities, and limitations.

## Anthropic's Claude Fable 5.2 built a shooter in 1 prompt -- and it only took 8 hours and $550

DevFeed: [Anthropic's Claude Fable 5.2 built a shooter in 1 prompt -- and it only took 8 hours and $550](<https://devfeed.tech/articles/anthropic-s-claude-fable-5-2-built-a-shooter-in-1-prompt-and-it-only-took-8-hours-and-550-129500.md>)

Original publisher: [Read original article](<https://thenewstack.io/claude-fable-game-demos/>)

Author: Amanda Caswell

Published: 2026-11-11T17:52:06Z

Content type: news

Language: en

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

Topics: [Claude](<https://devfeed.tech/topics/claude.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [anthropic](<https://devfeed.tech/topics/anthropic.md>), [implementation](<https://devfeed.tech/topics/implementation.md>), [3D](<https://devfeed.tech/topics/3d.md>)

Tags: [3d](<https://devfeed.tech/tags/3d.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-evaluation](<https://devfeed.tech/tags/ai-evaluation.md>), [ai-models](<https://devfeed.tech/tags/ai-models.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [api](<https://devfeed.tech/tags/api.md>), [approach](<https://devfeed.tech/tags/approach.md>), [build](<https://devfeed.tech/tags/build.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-fable](<https://devfeed.tech/tags/claude-fable.md>), [claude-fable-5-2](<https://devfeed.tech/tags/claude-fable-5-2.md>), [engineer](<https://devfeed.tech/tags/engineer.md>), [fable](<https://devfeed.tech/tags/fable.md>), [first](<https://devfeed.tech/tags/first.md>), [one](<https://devfeed.tech/tags/one.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [run](<https://devfeed.tech/tags/run.md>), [shooter](<https://devfeed.tech/tags/shooter.md>), [stack](<https://devfeed.tech/tags/stack.md>), [vibe-coding](<https://devfeed.tech/tags/vibe-coding.md>)

### AI overview

Claude Fable 5.2 produced a playable browser shooter, Saltline, from one prompt and an eight-hour run, using procedural visuals; engineer Felix Rieseberg estimated the API cost at $550. He later spent about two hours refining weapons, sounds, and gameplay, while keeping the original version available. The excerpt contrasts this with Matthew Berman's Unreal Engine recreation, estimated at nearly $15,000 over several days, and reports other projects that relied on tools such as Tripo and Suno for assets or music. Extended runs can consume enormous token volumes, and cached input reduced Saltline's reported cost substantially. The demos show substantial implementation work can be handed to an agent, but still require human direction and review; a playable build does not establish production quality, reliable calculations, or ease of later maintenance.

### Source excerpt

An Anthropic engineer gave Claude Fable 5.2 one prompt to build a first-person shooter and let it run for eight The post Anthropic's Claude Fable 5.2 built a shooter in 1 prompt -- and it only took 8 hours and $550 appeared first on The New Stack.

[Next page](<https://devfeed.tech/topics/prompt.md?cursor=WyIyMDI2LTExLTExVDE3OjUyOjA2KzAwOjAwIiwgIjY1Y2ZmNGFmLTJmZjMtNGZhNC04Y2MwLTM5MzFkNTUwODU1MyJd>)