Notizie

arXiv:2505.17826v2 Announce Type: replace-cross Abstract: Trinity-RFT is a general-purpose, unified and easy-to-use framework designed for reinforcement fine-tuning (RFT) of large language models. It is built with a modular and decoupled design, consisting of (1) an RFT-core that unifies and generalizes synchronous/asynchronous, on-policy/off-policy, and online/offline modes of RFT; (2) seamless integration for agent-environment interaction with high efficiency and robustness; and (3) systematic data pipelines optimized for RFT. Trinity-RFT can be easily adapted for diverse application scenarios, and serves as a unified platform for development and research of advanced reinforcement learning paradigms at both macroscopic and microscopic levels. This technical report outlines the vision, features, design and implementations of Trinity-RFT, accompanied by extensive examples, applications and experiments that demonstrate its functionalities and user-friendliness.

arXiv:2401.17196v3 Announce Type: replace Abstract: In text classification, creating an adversarial example means subtly perturbing a few words in a sentence without changing its meaning, causing it to be misclassified by a classifier. A concerning observation is that a significant portion of adversarial examples generated by existing methods change only one word. This single-word perturbation vulnerability represents a significant weakness in classifiers, which malicious users can exploit to efficiently create a multitude of adversarial examples. This paper studies this problem and makes the following key contributions: (1) We introduce a novel metric $rho$ to quantitatively assess a classifier’s robustness against single-word perturbation. (2) We present the SP-Attack, designed to exploit the single-word perturbation vulnerability, achieving a higher attack success rate, better preserving sentence meaning, while reducing computation costs compared to state-of-the-art adversarial methods. (3) We propose SP-Defense, which aims to improve r{ho} by applying data augmentation in learning. Experimental results on 4 datasets and BERT and distilBERT classifiers show that SP-Defense improves $rho$ by 14.6% and 13.9% and decreases the attack success rate of SP-Attack by 30.4% and 21.2% on two classifiers respectively, and decreases the attack success rate of existing attack methods that involve multiple-word perturbations.

arXiv:2507.09497v1 Announce Type: new Abstract: Modern enterprise environments demand intelligent systems capable of handling complex, dynamic, and multi-faceted tasks with high levels of autonomy and adaptability. However, traditional single-purpose AI systems often lack sufficient coordination, memory reuse, and task decomposition capabilities, limiting their scalability in realistic settings. To address these challenges, we present textbf{GoalfyMax}, a protocol-driven framework for end-to-end multi-agent collaboration. GoalfyMax introduces a standardized Agent-to-Agent (A2A) communication layer built on the Model Context Protocol (MCP), allowing independent agents to coordinate through asynchronous, protocol-compliant interactions. It incorporates the Experience Pack (XP) architecture, a layered memory system that preserves both task rationales and execution traces, enabling structured knowledge retention and continual learning. Moreover, our system integrates advanced features including multi-turn contextual dialogue, long-short term memory modules, and dynamic safety validation, supporting robust, real-time strategy adaptation. Empirical results on complex task orchestration benchmarks and case study demonstrate that GoalfyMax achieves superior adaptability, coordination, and experience reuse compared to baseline frameworks. These findings highlight its potential as a scalable, future-ready foundation for multi-agent intelligent systems.

arXiv:2507.08771v1 Announce Type: cross Abstract: To alleviate the computational burden of large language models (LLMs), architectures with activation sparsity, represented by mixture-of-experts (MoE), have attracted increasing attention. However, the non-differentiable and inflexible routing of vanilla MoE hurts model performance. Moreover, while each token activates only a few parameters, these sparsely-activated architectures exhibit low chunk-level sparsity, indicating that the union of multiple consecutive tokens activates a large ratio of parameters. Such a sparsity pattern is unfriendly for acceleration under low-resource conditions (e.g., end-side devices) and incompatible with mainstream acceleration techniques (e.g., speculative decoding). To address these challenges, we introduce a novel MoE architecture, BlockFFN, as well as its efficient training and deployment techniques. Specifically, we use a router integrating ReLU activation and RMSNorm for differentiable and flexible routing. Next, to promote both token-level sparsity (TLS) and chunk-level sparsity (CLS), CLS-aware training objectives are designed, making BlockFFN more acceleration-friendly. Finally, we implement efficient acceleration kernels, combining activation sparsity and speculative decoding for the first time. The experimental results demonstrate the superior performance of BlockFFN over other MoE baselines, achieving over 80% TLS and 70% 8-token CLS. Our kernels achieve up to 3.67$times$ speedup on real end-side devices than dense models. All codes and checkpoints are available publicly (https://github.com/thunlp/BlockFFN).

arXiv:2507.08496v1 Announce Type: new Abstract: While large language models (LLMs) have advanced procedural planning for embodied AI systems through strong reasoning abilities, the integration of multimodal inputs and counterfactual reasoning remains underexplored. To tackle these challenges, we introduce LLaPa, a vision-language model framework designed for multimodal procedural planning. LLaPa generates executable action sequences from textual task descriptions and visual environmental images using vision-language models (VLMs). Furthermore, we enhance LLaPa with two auxiliary modules to improve procedural planning. The first module, the Task-Environment Reranker (TER), leverages task-oriented segmentation to create a task-sensitive feature space, aligning textual descriptions with visual environments and emphasizing critical regions for procedural execution. The second module, the Counterfactual Activities Retriever (CAR), identifies and emphasizes potential counterfactual conditions, enhancing the model’s reasoning capability in counterfactual scenarios. Extensive experiments on ActPlan-1K and ALFRED benchmarks demonstrate that LLaPa generates higher-quality plans with superior LCS and correctness, outperforming advanced models. The code and models are available https://github.com/sunshibo1234/LLaPa.

arXiv:2507.08027v1 Announce Type: new Abstract: Recent studies have revealed a consistent liberal orientation in the ethical and political responses generated by most commercial large language models (LLMs), yet the underlying causes and resulting implications remain unclear. This paper systematically investigates the political temperament of seven prominent LLMs – OpenAI’s GPT-4o, Anthropic’s Claude Sonnet 4, Perplexity (Sonar Large), Google’s Gemini 2.5 Flash, Meta AI’s Llama 4, Mistral 7b Le Chat and High-Flyer’s DeepSeek R1 — using a multi-pronged approach that includes Moral Foundations Theory, a dozen established political ideology scales and a new index of current political controversies. We find strong and consistent prioritization of liberal-leaning values, particularly care and fairness, across most models. Further analysis attributes this trend to four overlapping factors: Liberal-leaning training corpora, reinforcement learning from human feedback (RLHF), the dominance of liberal frameworks in academic ethical discourse and safety-driven fine-tuning practices. We also distinguish between political “bias” and legitimate epistemic differences, cautioning against conflating the two. A comparison of base and fine-tuned model pairs reveals that fine-tuning generally increases liberal lean, an effect confirmed through both self-report and empirical testing. We argue that this “liberal tilt” is not a programming error or the personal preference of programmers but an emergent property of training on democratic rights-focused discourse. Finally, we propose that LLMs may indirectly echo John Rawls’ famous veil-of ignorance philosophical aspiration, reflecting a moral stance unanchored to personal identity or interest. Rather than undermining democratic discourse, this pattern may offer a new lens through which to examine collective reasoning.

arXiv:2402.08830v2 Announce Type: replace-cross Abstract: Several popular language models represent local contexts in an input text $x$ as bags of words. Such representations are naturally encoded by a sequence graph whose vertices are the distinct words occurring in $x$, with edges representing the (ordered) co-occurrence of two words within a sliding window of size $w$. However, this compressed representation is not generally bijective: some may be ambiguous, admitting several realizations as a sequence, while others may not admit any realization. In this paper, we study the realizability and ambiguity of sequence graphs from a combinatorial and algorithmic point of view. We consider the existence and enumeration of realizations of a sequence graph under multiple settings: window size $w$, presence/absence of graph orientation, and presence/absence of weights (multiplicities). When $w=2$, we provide polynomial time algorithms for realizability and enumeration in all cases except the undirected/weighted setting, where we show the $#$P-hardness of enumeration. For $w ge 3$, we prove the hardness of all variants, even when $w$ is considered as a constant, with the notable exception of the undirected unweighted case for which we propose XP algorithms for both problems, tight due to a corresponding $W[1]-$hardness result. We conclude with an integer program formulation to solve the realizability problem, and a dynamic programming algorithm to solve the enumeration problem in instances of moderate sizes. This work leaves open the membership to NP of both problems, a non-trivial question due to the existence of minimum realizations having size exponential on the instance encoding.

arXiv:2503.08893v2 Announce Type: replace Abstract: An ideal model evaluation should achieve two goals: identifying where the model fails and providing actionable improvement guidance. Toward these goals for language model (LM) evaluations, we formulate the problem of generating a weakness profile, a set of weaknesses expressed in natural language, given an LM’s performance on every individual instance in a benchmark. We introduce a suite of quantitative assessments to compare different weakness profiling methods. We also introduce a weakness profiling method EvalTree. EvalTree constructs a capability tree where each node represents a capability described in natural language and is linked to a subset of benchmark instances that specifically evaluate this capability; it then extracts nodes where the LM performs poorly to generate a weakness profile. On the MATH and WildChat benchmarks, we show that EvalTree outperforms baseline weakness profiling methods by identifying weaknesses more precisely and comprehensively. Weakness profiling further enables weakness-guided data collection, and training data collection guided by EvalTree-identified weaknesses improves LM performance more than other data collection strategies. We also show how EvalTree exposes flaws in Chatbot Arena’s human-voter-based evaluation practice. To facilitate future work, we provide an interface that allows practitioners to interactively explore the capability trees built by EvalTree.