The Real Cost of Inaction: How Silos Hurt Productivity for Data Scientists (Sponsored)
Post Content
Uncategorized
Meta AI Releases SAM Audio: A State-of-the-Art Unified Model that Uses Intuitive and Multimodal Prompts for Audio Separation
Meta has released SAM Audio, a prompt driven audio separation model that targets a common editing bottleneck, isolating one sound from a real world mix without building a custom model per sound class. Meta released 3 main sizes, sam-audio-small, sam-audio-base, and sam-audio-large. The model is available to download and to try in the Segment Anything Playground. Architecture SAM Audio uses separate encoders for each conditioning signal, an audio encoder for the mixture, a text encoder for the natural language description, a span encoder for time anchors, and a visual encoder that consumes a visual prompt derived from video plus an object mask. The encoded streams are concatenated into time aligned features, then processed by a diffusion transformer that applies self attention over the time aligned representation and cross attention to the textual feature, then a DACVAE decoder reconstructs waveforms and emits 2 outputs, target audio and residual audio. https://ai.meta.com/blog/sam-audio/ What SAM Audio does, and what ‘segment’ means here? SAM Audio takes an input recording that contains multiple overlapping sources, for example speech plus traffic plus music, and separates out a target source based on a prompt. In the public inference API, the model produces 2 outputs, result.target and result.residual. The research team describes target as the isolated sound, and residual as everything else. That target plus residual interface maps directly to editor operations. If you want to remove a dog bark across a podcast track, you can treat the bark as the target, then subtract it by keeping only residual. If you want to extract a guitar part from a concert clip, you keep the target waveform instead. Meta uses these exact kinds of examples to explain what the model is meant to enable. The 3 prompt types Meta is shipping Meta positions SAM Audio as a single unified model that supports 3 prompt types, and it says these prompts can be used alone or combined. Text prompting: You describe the sound in natural language, for example “dog barking” or “singing voice”, and the model separates that sound from the mixture. Meta lists text prompts as one of the core interaction modes, and the open source repo includes an end to end example using SAMAudioProcessor and model.separate. Visual prompting: You click the person or object in a video and ask the model to isolate the audio associated with that visual object. Meta team describes visual prompting as selecting the sounding object in the video. In the released code path, visual prompting is implemented by passing video frames plus masks into the processor via masked_videos. Span prompting: Meta team calls span prompting an industry first. You mark time segments where the target sound occurs, then the model uses those spans to guide separation. This matters for ambiguous cases, for example when the same instrument appears in multiple passages, or when a sound is present only briefly and you want to prevent the model from over separating. https://ai.meta.com/blog/sam-audio/ Results Meta team positions SAM Audio as achieving cutting edge performance across diverse, real world scenarios, and frames it as a unified alternative to single purpose audio tools. The team publishes a subjective evaluation table across categories, General, SFX, Speech, Speaker, Music, Instr(wild), Instr(pro), with General scores of 3.62 for sam audio small, 3.28 for sam audio base, and 3.50 for sam audio large, and Instr(pro) scores reaching 4.49 for sam audio large. Key Takeaways SAM Audio is a unified audio separation model, it segments sound from complex mixtures using text prompts, visual prompts, and time span prompts. The core API produces two waveforms per request, target for the isolated sound and residual for everything else, which maps cleanly to common edit operations like remove noise, extract stem, or keep ambience. Meta released multiple checkpoints and variants, including sam-audio-small, sam-audio-base, sam-audio-large, plus tv variants that the repo says perform better for visual prompting, the repo also publishes a subjective evaluation table by category. The release includes tooling beyond inference, Meta provides a sam-audio-judge model that scores separation results against a text description with overall quality, recall, precision, and faithfulness. Check out the Technical details and GitHub Page. Feel free to check out our GitHub Page for Tutorials, Codes and Notebooks. Also, feel free to follow us on Twitter and don’t forget to join our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post Meta AI Releases SAM Audio: A State-of-the-Art Unified Model that Uses Intuitive and Multimodal Prompts for Audio Separation appeared first on MarkTechPost.
Gemini 3 Flash arrives with reduced costs and latency — a powerful combo for enterprises
Enterprises can now harness the power of a large language model that’s near that of the state-of-the-art Google’s Gemini 3 Pro, but at a fraction of the cost and with increased speed, thanks to the newly released Gemini 3 Flash. The model joins the flagship Gemini 3 Pro, Gemini 3 Deep Think, and Gemini Agent, all of which were announced and released last month. Gemini 3 Flash, now available on Gemini Enterprise, Google Antigravity, Gemini CLI, AI Studio, and on preview in Vertex AI, processes information in near real-time and helps build quick, responsive agentic applications. The company said in a blog post that Gemini 3 Flash “builds on the model series that developers and enterprises already love, optimized for high-frequency workflows that demand speed, without sacrificing quality. The model is also the default for AI Mode on Google Search and the Gemini application. Tulsee Doshi, senior director, product management on the Gemini team, said in a separate blog post that the model “demonstrates that speed and scale don’t have to come at the cost of intelligence.” “Gemini 3 Flash is made for iterative development, offering Gemini 3’s Pro-grade coding performance with low latency — it’s able to reason and solve tasks quickly in high-frequency workflows,” Doshi said. “It strikes an ideal balance for agentic coding, production-ready systems and responsive interactive applications.” Early adoption by specialized firms proves the model’s reliability in high-stakes fields. Harvey, an AI platform for law firms, reported a 7% jump in reasoning on their internal ‘BigLaw Bench,’ while Resemble AI discovered that Gemini 3 Flash could process complex forensic data for deepfake detection 4x faster than Gemini 2.5 Pro. These aren’t just speed gains; they are enabling ‘near real-time’ workflows that were previously impossible. More efficient at a lower cost Enterprise AI builders have become more aware of the cost of running AI models, especially as they try to convince stakeholders to put more budget into agentic workflows that run on expensive models. Organizations have turned to smaller or distilled models, focusing on open models or other research and prompting techniques to help manage bloated AI costs. For enterprises, the biggest value proposition for Gemini 3 Flash is that it offers the same level of advanced multimodal capabilities, such as complex video analysis and data extraction, as its larger Gemini counterparts, but is far faster and cheaper. While Google’s internal materials highlight a 3x speed increase over the 2.5 Pro series, data from independent benchmarking firm Artificial Analysis adds a layer of crucial nuance. In the latter organization’s pre-release testing, Gemini 3 Flash Preview recorded a raw throughput of 218 output tokens per second. This makes it 22% slower than the previous ‘non-reasoning’ Gemini 2.5 Flash, but it is still significantly faster than rivals including OpenAI’s GPT-5.1 high (125 t/s) and DeepSeek V3.2 reasoning (30 t/s). Most notably, Artificial Analysis crowned Gemini 3 Flash as the new leader in their AA-Omniscience knowledge benchmark, where it achieved the highest knowledge accuracy of any model tested to date. However, this intelligence comes with a ‘reasoning tax’: the model more than doubles its token usage compared to the 2.5 Flash series when tackling complex indexes. This high token density is offset by Google’s aggressive pricing: when accessing through the Gemini API, Gemini 3 Flash costs $0.50 per 1 million input tokens, compared to $1.25/1M input tokens for Gemini 2.5 Pro, and $3/1M output tokens, compared to $ 10/1 M output tokens for Gemini 2.5 Pro. This allows Gemini 3 Flash to claim the title of the most cost-efficient model for its intelligence tier, despite being one of the most ‘talkative’ models in terms of raw token volume. Here’s how it stacks up to rival LLM offerings: Model Input (/1M) Output (/1M) Total Cost Source Qwen 3 Turbo $0.05 $0.20 $0.25 Alibaba Cloud Grok 4.1 Fast (reasoning) $0.20 $0.50 $0.70 xAI Grok 4.1 Fast (non-reasoning) $0.20 $0.50 $0.70 xAI deepseek-chat (V3.2-Exp) $0.28 $0.42 $0.70 DeepSeek deepseek-reasoner (V3.2-Exp) $0.28 $0.42 $0.70 DeepSeek Qwen 3 Plus $0.40 $1.20 $1.60 Alibaba Cloud ERNIE 5.0 $0.85 $3.40 $4.25 Qianfan Gemini 3 Flash Preview $0.50 $3.00 $3.50 Google Claude Haiku 4.5 $1.00 $5.00 $6.00 Anthropic Qwen-Max $1.60 $6.40 $8.00 Alibaba Cloud Gemini 3 Pro (≤200K) $2.00 $12.00 $14.00 Google GPT-5.2 $1.75 $14.00 $15.75 OpenAI Claude Sonnet 4.5 $3.00 $15.00 $18.00 Anthropic Gemini 3 Pro (>200K) $4.00 $18.00 $22.00 Google Claude Opus 4.5 $5.00 $25.00 $30.00 Anthropic GPT-5.2 Pro $21.00 $168.00 $189.00 OpenAI More ways to save But enterprise developers and users can cut costs further by eliminating the lag most larger models often have, which racks up token usage. Google said the model “is able to modulate how much it thinks,” so that it uses more thinking and therefore more tokens for more complex tasks than for quick prompts. The company noted Gemini 3 Flash uses 30% fewer tokens than Gemini 2.5 Pro. To balance this new reasoning power with strict corporate latency requirements, Google has introduced a ‘Thinking Level’ parameter. Developers can toggle between ‘Low’—to minimize cost and latency for simple chat tasks—and ‘High’—to maximize reasoning depth for complex data extraction. This granular control allows teams to build ‘variable-speed’ applications that only consume expensive ‘thinking tokens’ when a problem actually demands PhD-level lo The economic story extends beyond simple token prices. With the standard inclusion of Context Caching, enterprises processing massive, static datasets—such as entire legal libraries or codebase repositories—can see a 90% reduction in costs for repeated queries. When combined with the Batch API’s 50% discount, the total cost of ownership for a Gemini-powered agent drops significantly below the threshold of competing frontier models “Gemini 3 Flash delivers exceptional performance on coding and agentic tasks combined with a lower price point, allowing teams to deploy sophisticated reasoning costs across high-volume processes without hitting barriers,” Google said. By offering a model that delivers strong multimodal performance at a more affordable price, Google is making the case that enterprises concerned with controlling their AI spend should choose its models, especially Gemini 3 Flash.
AI, Committee, Notizie, Uncategorized
Why Reinforcement Fine-Tuning Enables MLLMs Preserve Prior Knowledge Better: A Data Perspective
arXiv:2506.23508v3 Announce Type: replace Abstract: Post-training algorithms such as Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT) are widely used to adapt multimodal large language models to downstream tasks. While effective at task adaptation, their impact on prior knowledge remains unclear. In this paper, we introduce jigsaw puzzles as a novel task absent from existing pretraining corpora and systematically study the behavior of SFT and RFT on open-source multimodal model, Qwen2.5-VL series. Our experiments reveal a sharp trade-off: SFT enables rapid task acquisition but leads to catastrophic forgetting, whereas RFT learns more slowly but maintains prior knowledge. We study this phenomenon through learning dynamics by examining both the magnitude and direction of how training data influence prior knowledge. Our analysis shows that RFT mainly reinforces correct samples naturally aligned with the base model’s probability landscape, leading to weaker interference with prior knowledge. Moreover, training on RFT-simulated rollouts, which exert a small magnitude of influence and are well aligned in direction to prior knowledge, allows SFT to preserve prior knowledge better while rapidly learning new tasks. These findings suggest that distribution of training data, rather than algorithmic differences, plays a central role in forgetting, and highlight RFT’s potential for stable continual learning in multimodal large language models.
AI, Committee, Notizie, Uncategorized
Estimating Privacy Leakage of Augmented Contextual Knowledge in Language Models
arXiv:2410.03026v3 Announce Type: replace Abstract: Language models (LMs) rely on their parametric knowledge augmented with relevant contextual knowledge for certain tasks, such as question answering. However, the contextual knowledge can contain private information that may be leaked when answering queries, and estimating this privacy leakage is not well understood. A straightforward approach of directly comparing an LM’s output to the contexts can overestimate the privacy risk, since the LM’s parametric knowledge might already contain the augmented contextual knowledge. To this end, we introduce *context influence*, a metric that builds on differential privacy, a widely-adopted privacy notion, to estimate the privacy leakage of contextual knowledge during decoding. Our approach effectively measures how each subset of the context influences an LM’s response while separating the specific parametric knowledge of the LM. Using our context influence metric, we demonstrate that context privacy leakage occurs when contextual knowledge is out of distribution with respect to parametric knowledge. Moreover, we experimentally demonstrate how context influence properly attributes the privacy leakage to augmented contexts, and we evaluate how factors — such as model size, context size, generation position, etc. — affect context privacy leakage. The practical implications of our results will inform practitioners of the privacy risk associated with augmented contextual knowledge.
Estimating Privacy Leakage of Augmented Contextual Knowledge in Language Models Leggi l'articolo »
AI, Committee, Notizie, Uncategorized
RePo: Language Models with Context Re-Positioning
arXiv:2512.14391v1 Announce Type: cross Abstract: In-context learning is fundamental to modern Large Language Models (LLMs); however, prevailing architectures impose a rigid and fixed contextual structure by assigning linear or constant positional indices. Drawing on Cognitive Load Theory (CLT), we argue that this uninformative structure increases extraneous cognitive load, consuming finite working memory capacity that should be allocated to deep reasoning and attention allocation. To address this, we propose RePo, a novel mechanism that reduces extraneous load via context re-positioning. Unlike standard approaches, RePo utilizes a differentiable module, $f_phi$, to assign token positions that capture contextual dependencies, rather than replying on pre-defined integer range. By continually pre-training on the OLMo-2 1B backbone, we demonstrate that RePo significantly enhances performance on tasks involving noisy contexts, structured data, and longer context length, while maintaining competitive performance on general short-context tasks. Detailed analysis reveals that RePo successfully allocate higher attention to distant but relevant information, assign positions in dense and non-linear space, and capture the intrinsic structure of the input context. Our code is available at https://github.com/SakanaAI/repo.
RePo: Language Models with Context Re-Positioning Leggi l'articolo »
AI, Committee, Notizie, Uncategorized
Inflation Attitudes of Large Language Models
arXiv:2512.14306v1 Announce Type: new Abstract: This paper investigates the ability of Large Language Models (LLMs), specifically GPT-3.5-turbo (GPT), to form inflation perceptions and expectations based on macroeconomic price signals. We compare the LLM’s output to household survey data and official statistics, mimicking the information set and demographic characteristics of the Bank of England’s Inflation Attitudes Survey (IAS). Our quasi-experimental design exploits the timing of GPT’s training cut-off in September 2021 which means it has no knowledge of the subsequent UK inflation surge. We find that GPT tracks aggregate survey projections and official statistics at short horizons. At a disaggregated level, GPT replicates key empirical regularities of households’ inflation perceptions, particularly for income, housing tenure, and social class. A novel Shapley value decomposition of LLM outputs suited for the synthetic survey setting provides well-defined insights into the drivers of model outputs linked to prompt content. We find that GPT demonstrates a heightened sensitivity to food inflation information similar to that of human respondents. However, we also find that it lacks a consistent model of consumer price inflation. More generally, our approach could be used to evaluate the behaviour of LLMs for use in the social sciences, to compare different models, or to assist in survey design.
Inflation Attitudes of Large Language Models Leggi l'articolo »
AI, Committee, Notizie, Uncategorized
PAT: Accelerating LLM Decoding via Prefix-Aware Attention with Resource Efficient Multi-Tile Kernel
arXiv:2511.22333v2 Announce Type: replace-cross Abstract: LLM serving is increasingly dominated by decode attention, which is a memory-bound operation due to massive KV cache loading from global memory. Meanwhile, real-world workloads exhibit substantial, hierarchical shared prefixes across requests (e.g., system prompts, tools/templates, RAG). Existing attention implementations fail to fully exploit prefix sharing: one-query-per-CTA execution repeatedly loads shared prefix KV cache, while one-size-fits-all tiling leaves on-chip resources idle and exacerbates bubbles for uneven KV lengths. These choices amplify memory bandwidth pressure and stall memory-bound decode attention. This paper introduces PAT, a prefix-aware attention kernel implementation for LLM decoding that organizes execution with a pack-forward-merge paradigm. PAT packs queries by shared prefix to reduce repeated memory accesses, runs a customized multi-tile kernel to achieve high resource efficiency. It further applies practical multi-stream forwarding and KV splitting to reduce resource bubbles. The final merge performs online softmax with negligible overhead. We implement PAT as an off-the-shelf plugin for vLLM. Evaluation on both real-world and synthetic workloads shows that PAT reduces attention latency by 53.5% on average and TPOT by 17.0-93.1% under the same configurations against state-of-the-art attention kernels. PAT’s source code is publicly available at https://github.com/flashserve/PAT.
AI, Committee, Notizie, Uncategorized
BreakFun: Jailbreaking LLMs via Schema Exploitation
arXiv:2510.17904v2 Announce Type: replace-cross Abstract: The proficiency of Large Language Models (LLMs) in processing structured data and adhering to syntactic rules is a capability that drives their widespread adoption but also makes them paradoxically vulnerable. In this paper, we investigate this vulnerability through BreakFun, a jailbreak methodology that weaponizes an LLM’s adherence to structured schemas. BreakFun employs a three-part prompt that combines an innocent framing and a Chain-of-Thought distraction with a core “Trojan Schema”–a carefully crafted data structure that compels the model to generate harmful content, exploiting the LLM’s strong tendency to follow structures and schemas. We demonstrate this vulnerability is highly transferable, achieving an average success rate of 89% across 13 foundational and proprietary models on JailbreakBench, and reaching a 100% Attack Success Rate (ASR) on several prominent models. A rigorous ablation study confirms this Trojan Schema is the attack’s primary causal factor. To counter this, we introduce the Adversarial Prompt Deconstruction guardrail, a defense that utilizes a secondary LLM to perform a “Literal Transcription”–extracting all human-readable text to isolate and reveal the user’s true harmful intent. Our proof-of-concept guardrail demonstrates high efficacy against the attack, validating that targeting the deceptive schema is a viable mitigation strategy. Our work provides a look into how an LLM’s core strengths can be turned into critical weaknesses, offering a fresh perspective for building more robustly aligned models.
BreakFun: Jailbreaking LLMs via Schema Exploitation Leggi l'articolo »
AI, Committee, Notizie, Uncategorized
SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines
arXiv:2509.21320v3 Announce Type: replace Abstract: We present a scientific reasoning foundation model that aligns natural language with heterogeneous scientific representations. The model is pretrained on a 206B-token corpus spanning scientific text, pure sequences, and sequence-text pairs, then aligned via SFT on 40M instructions, annealed cold-start bootstrapping to elicit long-form chain-of-thought, and reinforcement learning with task-specific reward shaping, which instills deliberate scientific reasoning. It supports four capability families, covering up to 103 tasks across workflows: (i) faithful translation between text and scientific formats, (ii) text/knowledge extraction, (iii) property prediction, (iv) property classification, (v) unconditional and conditional sequence generation and design. Compared with specialist systems, our approach broadens instruction coverage, improves cross-domain generalization, and enhances fidelity. We detail data curation and training and show that cross-discipline learning strengthens transfer and downstream reliability. The model, instruct tuning datasets and the evaluation code are open-sourced at https://huggingface.co/SciReason and https://github.com/open-sciencelab/SciReason.
SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines Leggi l'articolo »
