EMNLP 2026 Highlights: 12 Long Papers from Meituan on Streaming Voice, RL, Agents & More
Meituan's technical team shares 12 EMNLP 2026 long papers covering streaming voice interaction, reinforcement learning with self-verification, instruction-following evaluation, agent skill internalization, on-policy distillation, Text-to-SQL over unknown schemas, unified memory frameworks, personalized reasoning, tool-use data synthesis, neuron consistency estimation, fast-slow reward models, and fine-grained gated delta networks.
A Listen-Think-Speak Framework for Efficient Streaming Voice Interaction
Paper: A Listen-Think-Speak Framework for Efficient Streaming Voice Interaction via Semantic Triggering and Incremental Reasoning (Main Conference, Long Paper). The authors address the dilemma of real-time voice agents: end-to-end models lack deep reasoning, while cascaded pipelines (ASR → LLM → TTS) suffer high latency due to strict serial execution. Existing streaming strategies use mechanical segmentation or speculative generation, which either break semantic units or waste compute on rollbacks.
They propose LTS-Agent , a "Listen-Think-Speak" framework that explicitly separates "when to think" from "how to reason incrementally." It introduces a Dynamic Semantic Trigger that activates at semantically saturated prefixes, and a Dual-Role Stream Orchestrator leveraging the reasoning engine's batching capability to coordinate a background "Thinker" (state maintenance) and a foreground "Speaker" (speculative solving). This parallel design enables "speak while thinking" without blocking immediate responses.
A Pause-and-Repair benchmark injects hesitations and self-corrections into complex reasoning tasks to stress-test streaming robustness. Experiments on VERA, Spoken-MQA, BigBenchAudio, and the new benchmark show LTS-Agent achieves superior accuracy-latency-efficiency trade-offs versus serial cascaded baselines and existing streaming strategies.
Silence the Judge: Reinforcement Learning with Self-Verifier via Latent Geometric Clustering
Paper: Silence the Judge: Reinforcement Learning with Self-Verifier via Latent Geometric Clustering (Main Conference, Long Paper). Group Relative Policy Optimization (GRPO) boosts LLM reasoning but relies on sparse discrete rewards and expensive external verifiers.
The authors propose Latent-GRPO , using latent-space geometry to obtain intrinsic rewards, eliminating external verification. Empirical analysis reveals a key geometric property: terminal token representations of correct reasoning trajectories exhibit high intra-class cosine similarity, forming tight clusters, while incorrect trajectories disperse.
Based on this, they introduce Iterative Robust Centroid Estimation (IRCE) : spherical projection removes magnitude fluctuations, and iterative soft-weighted aggregation estimates a "ground-truth centroid" to produce dense continuous reward signals. Experiments across model sizes show Latent-GRPO matches external-verifier reasoning performance while significantly reducing training latency and memory overhead, demonstrating efficient, scalable, training-free reward modeling directly in latent space.
One Instruction Is a Benchmark: End-to-End Instruction-Following Evaluation with FlexBench
Paper: One Instruction Is a Benchmark: End-to-End Instruction-Following Evaluation with FlexBench (Main Conference, Long Paper). LLMs increasingly rely on understanding long, precise, workflow-tailored user instructions; generic leaderboard aggregates poorly reflect faithful execution of specific deployment instructions.
The authors present FlexBench/FlexEval , an instruction-conditioned evaluation framework that dynamically generates specialized test suites from a target instruction. FlexBench extracts verifiable evaluation dimensions and generates an instruction-conditioned dialogue corpus. FlexEval aggregates dimension-level judgments into two complementary metrics: COVERAGE (workflow progress) and ACHIEVEMENT (conditional correctness). Evaluation of nine contemporary LLMs on nine multi-turn target instructions across diverse scenario categories and workflow topologies (3,888 generated dialogues) shows model rankings and failure patterns vary significantly across target instructions, highlighting the need for instruction-conditioned evaluation beyond fixed holistic benchmarks.
SIRI: Self-Internalizing Reinforcement Learning with Intrinsic Skills for LLM Agent Training
Paper: SIRI: Self-Internalizing Reinforcement Learning with Intrinsic Skills for LLM Agent Training (Main Conference, Long Paper). Agent skills compress successful experience into reusable policy knowledge, but existing methods depend on external models or separate modules for skill generation and require continuous retrieval/injection at inference, adding complexity, context overhead, and latency.
The authors propose SIRI , a three-stage RL framework for autonomous skill discovery, verification, and internalization. Stage 1: GiGPO warm-up acquires basic interaction ability and accumulates skill-free success trajectories. Stage 2: the model distills candidate skills from its own success trajectories and verifies skill utility via paired Skill-Augmented vs. Skill-Free rollouts. Stage 3: combining trajectory-level skill gain with action-level advantage, only truly beneficial skill-guided actions are distilled into the skill-free policy. After training, the model performs inference with only the original task prompt, no skill retrieval or injection.
On Qwen2.5-7B-Instruct, SIRI lifts GiGPO's ALFWorld success rate from 90.8% to 93.0% (+2.2 pp) and WebShop from 72.8% to 81.3% (+8.5 pp), outperforming prompt learning, standard RL, and memory-augmented methods. Further experiments show autonomously mined skills, after online verification and internalization, match the effect of skills generated by closed-source large models.
SCOPE: Signal-Calibrated On-Policy Distillation Enhancement with Dual-Path Adaptive Weighting
Paper: SCOPE: Signal-Calibrated On-Policy Distillation Enhancement with Dual-Path Adaptive Weighting (Main Conference, Long Paper). In LLM reasoning alignment, On-Policy Distillation (OPD) uses teacher token-level KL supervision to mitigate credit assignment from outcome-level rewards, but existing methods apply uniform supervision to all sampled trajectories, ignoring signal quality differences.
SCOPE routes rollouts by trajectory correctness into two complementary supervision paths: for incorrect trajectories, teacher perplexity-weighted KL distillation prioritizes instances where the teacher genuinely corrects errors while suppressing "confidently wrong" misleading signals; for correct trajectories, student perplexity-weighted MLE focuses reinforcement on low-confidence samples at the capability boundary, avoiding over-reinforcement of mastered paths. Both paths use group-level normalization to adaptively calibrate weight distributions across prompt difficulty variations.
On MATH500, AIME24, AIME25, AMC23, Minerva, and OlympiadBench, SCOPE achieves consistent gains: Avg@32 average relative improvement 11.42%, Pass@32 improvement 7.30% over all competitive baselines, validating signal-quality-aware adaptive supervision in OPD.
TRUST-SQL: Tool-Integrated Multi-Turn Reinforcement Learning for Text-to-SQL over Unknown Schemas
Paper: TRUST-SQL: Tool-Integrated Multi-Turn Reinforcement Learning for Text-to-SQL over Unknown Schemas (Main Conference, Long Paper). Existing Text-to-SQL methods assume pre-loaded full database schemas, which exceeds context windows in large dynamic enterprise databases and introduces noise; without schemas, models hallucinate tables/columns.
The authors propose TRUST-SQL (Truthful Reasoning with Unknown Schema via Tools), a tool-integrated multi-turn RL framework for unknown schemas. The task is modeled as a POMDP with an Explore–Propose–Generate–Confirm four-stage protocol: the Propose stage requires explicit confirmation of verified tables/columns, and execution feedback supports backtracking corrections. A Phase-Aware GRPO computes separate advantages for schema exploration (up to Propose) and full trajectory, with token-level masking to confine schema rewards to corresponding stages, alleviating multi-turn credit assignment.
On BIRD-Dev, Spider-Test, and three Spider robustness benchmarks, without schema pre-loading and using greedy decoding, 4B and 8B models improve average execution accuracy by 30.6 and 16.6 percentage points over base models; a 14B model reaches 68.1% on BIRD-Dev (70.1% with majority voting), matching or surpassing strong baselines that rely on full schema pre-loading.
UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task Streams
Paper: UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task Streams (Main Conference, Long Paper). LLM agents need to accumulate and reuse task experience over long interactions, but real-world tasks evolve continuously without clear boundaries. Relying solely on external retrieval memory or parametric memory creates a stability-plasticity dilemma: external memory absorbs new information quickly but struggles to internalize recurring patterns and adds retrieval overhead; parametric memory is stable and efficient but depends on predefined task boundaries and fixed parameter capacity.
Inspired by Complementary Learning Systems theory, the authors propose UniMem , a unified memory framework using learnable routing tokens to coordinate episodic and parametric memory, decoupling task identification from execution. For known tasks, the model activates corresponding parametric memory units; for novel, sparse, or uncertain tasks, it stores them in an episodic buffer and processes via retrieval augmentation. When recurring reliable patterns emerge in the buffer, UniMem automatically discovers latent task structures without task labels and dynamically extends new routing tokens and parametric memory modules, achieving gradual consolidation from short-term episodic to long-term parametric memory.
Experiments show UniMem maintains stable routing and task execution performance across diverse long-horizon streaming tasks and backbone models, while autonomously expanding memory and controlling parameter growth, outperforming external memory, parameter-efficient fine-tuning, and token memory baselines.
TagPR: Tag-Guided Process Supervision for Personalization Reasoning in Large Language Models
Paper: TagPR: Tag-Guided Process Supervision for Personalization Reasoning in Large Language Models (Main Conference, Long Paper). LLMs have strong general reasoning but in personalized tasks tend to follow generic logic, ignoring unique preferences in user history; outcome-only RL cannot constrain the reasoning process, leading to "correct answer, wrong process."
The authors propose TagPR , introducing semantic tags into personalized reasoning chains to explicitly decompose steps like input analysis, user example checking, preference pattern identification, object comparison, and conclusion formation into supervisable steps. TagPR first automatically constructs high-quality tagged reasoning data for supervised fine-tuning (SFT), then optimizes via multi-stage RL. Its composite reward combines tag-level process supervision with a personalized reward model with user embeddings (PRMU), aligning the model's reasoning process with user-specific logic.
Experiments on LaMP, LongLaMP, PGraphRAG, and a custom dataset show an average 32.65% relative improvement over the base model across all LaMP tasks, with more accurate, interpretable, and transferable personalized reasoning.
ToolForge: A Data Synthesis Pipeline for Multi-Hop, Multi-Turn, and Self-Reflective Tool-Use Data
Paper: ToolForge: A Data Synthesis Pipeline for Multi-Hop, Multi-Turn, and Self-Reflective Tool-Use Data (Main Conference, Long Paper). Training tool-use LLMs typically requires massive real API calls, costly and hard to cover complex scenarios like multi-hop reasoning, multi-turn interaction, and error reflection.
ToolForge synthesizes tool-use data without large-scale online tool execution. Starting from QA seeds, it builds an extensible tool space of 19 basic virtual tools and their auto-extended variants. Through single/multi-turn, single/multi-tool interaction paradigms, it simulates correct retrieval, tool mis-selection, parameter mis-selection, and tool switching, generating tool-call trajectories with multi-hop evidence integration and self-reflection correction. A multi-layer verification mechanism combines rule-based checks and model-based semantic evaluation to filter structurally or logically inconsistent samples, and MCTS mines hard negatives to verify reliability.
Fine-tuning Qwen3-8B on only 4,250 ToolForge synthetic samples yields a 21.6% relative improvement over TIR on five out-of-domain QA benchmarks, and shows strong generalization on BFCL and API-Bank tool-calling benchmarks.
RACE: Scalable Statistical Estimation of Functional Consistency in LLM Neurons
Paper: RACE: Scalable Statistical Estimation of Functional Consistency in LLM Neurons (Main Conference, Long Paper). In mechanistic interpretability, discovering stable neuron behavior across the entire domain remains challenging. Existing methods rely on instance-level point estimates or computationally expensive processes, either masking population-level variability or limiting scalable cross-domain analysis.
The authors propose RACE (Residual Alignment for Consistency Estimation), a forward-pass statistical framework for evaluating Transformer neuron functional consistency across domains. Perturbation experiments show RACE has superior domain specificity compared to gradient-based point estimates. Token-distribution-level results validate the association between selected neurons and target domains. Computational overhead is two orders of magnitude lower than gradient-based methods.
Fast-Slow Thinking RM: Efficient Integration of Scalar and Generative Reward Models
Paper: Fast-Slow Thinking RM: Efficient Integration of Scalar and Generative Reward Models (Main Conference, Long Paper). In content moderation, customer service evaluation, etc., LLM-as-a-Judge is becoming a key automated judgment paradigm. Judgment difficulty varies greatly across samples: many simple samples allow quick decisions, while complex, ambiguous, or low-confidence samples need deeper reasoning.
The authors propose Fast-Slow Thinking Reward Models (F/S-RM) , a hybrid judgment architecture inspired by dual-process theory. F/S-RM unifies "fast thinking" and "slow thinking" in a single model: for simpler, clearer samples, the model makes fast pairwise judgments via the first token; for complex or low-confidence samples, it automatically activates chain-of-thought slow thinking for further reasoning. A dual-confidence activation mechanism dynamically decides whether to enter slow thinking, forming an adaptive "fast judgment — difficult escalation" decision flow.
Under mixed reasoning mode, F/S-RM achieves an average score of 84.3 on multiple reward model benchmarks (state-of-the-art), while reducing token consumption by 22.5%. The recently popular Jev fast decision model shares similar design ideas in probability and confidence utilization; F/S-RM further integrates first-token fast judgment and CoT slow thinking within the same model.
FG²-GDN: Enhancing Long-Context Gated Delta Networks with Doubly Fine-Grained Control
Paper: FG²-GDN: Enhancing Long-Context Gated Delta Networks with Doubly Fine-Grained Control (Main Conference, Long Paper). As LLM applications extend to long-text understanding and complex interactions, balancing long-context modeling capacity with computational efficiency is crucial. Linear attention via recursive fixed-size memory state updates offers a path for efficient long-context modeling. However, Kimi Delta Attention (KDA), while achieving per-channel forgetting control, still uses a shared scalar write rate across channels, unable to flexibly adjust update strength per feature needs.
The authors propose FG²-GDN , expanding the write rate from a scalar to a per-channel vector, enabling simultaneous fine-grained control of "how much old information to retain" and "how much new information to write." FG²-GDN+ further separates key and value scaling control on top of per-channel writing, enhancing memory update flexibility. A symmetric square-root scaling design preserves the update structure required for KDA's chunked parallel algorithm, reusing its efficient compute path. Experiments at 340M and 1.3B model scales cover language modeling, long-context retrieval, and real long-text understanding tasks.
Results show fine-grained write control improves average language modeling performance; FG²-GDN+ with key-value separation achieves overall better results than GDN and KDA on RULER and LongBench, and maintains good retrieval ability beyond training context length. The method adds minimal extra parameters, with prefill efficiency close to KDA. The study demonstrates that refining write control beyond fine-grained forgetting is an effective path to improving linear attention long-context memory capacity.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
