The Inference Report

Existing scaling strategies for Multimodal Large Language Models (MLLMs) typically expand either model parameters or sequential inference computation, incurring substantial memory or latency overhead. More importantly, most existing methods fail to alter the rigid, fixed computation allocation between the Vision Transformer and the Large Language Model components, limiting task-specific optimization. To address this, we introduce the Parallel Vision-Language (ParVL) scaling framework for MLLMs, which scales parallel computation by reusing the existing ViT and LLM backbone parameters across multiple vision and language branches. This framework raises a central question: given a fixed backbone parameter budget, how should additional shared-backbone computation be allocated between the vision and language modalities? We instantiate each parallel computational stream with branch-specific prefix parameters over a shared backbone, and train the entire model end-to-end via full-parameter supervised fine-tuning on roughly 13B tokens. We systematically study the computation-allocation trade-off between the ViT encoder and LLM decoder. ParVL improves overall multimodal performance over same-recipe single-branch baselines, and the best evaluated vision--language allocation varies across tasks. Code is available at https://github.com/YangYangGirl/ParVL.

Large language models (LLMs), and the agents built on top of them, are now benchmarked heavily on whether they can finish a task -- fix a bug, drive a browser, operate a GUI. A complementary social ability, namely how well a model understands and forecasts the way real social events unfold, has barely been measured. We introduce SocietyBench, an end-to-end benchmark that takes a one-line event topic, collects Web news and social-media posts across five platforms, distills them into a date-indexed timeline that keeps factual events and a public-opinion layer separate, and then turns every cutoff date on that timeline into an audited bank of forecasting questions. Questions are scored on two orthogonal 100-point axes: probability calibration and temporal accuracy. Before any model sees a timeline, a three-phase procedure replaces every named entity and shifts every date by a per-event constant, turning a real arc into a counterfactual social world -- structurally identical to what happened, but stripped of the surface labels a model could match against pre-training memory. On five heterogeneous events and 125 prediction points in Chinese and English editions, the strongest of six frontier LLMs reaches only 75.0 out of 100, against a trivial anchor of 50. The two axes come apart: a model can be calibration-strong but time-weak, or the reverse. Three agent frameworks built on a shared base model fail to improve on that base, and two model-free heuristics trail every LLM. Per-event gaps reach 21.4 points on a single axis, which is our main argument for evaluating on several events rather than one. All anonymized timelines, question banks, ground truth, and scoring code are released.

Benchmarks that measure the forecasting ability of large language models are almost always retrospective: the event has happened, the answer is somewhere on the Web, and the evaluation must defend itself against memorisation. We report the opposite design. Over the 39 days of the 2026 FIFA World Cup, six frontier LLMs -- all with extended thinking and native server-side web search -- were asked before every kickoff, one match at a time, to fill in a seven-market prediction card for all 104 matches, plus 12 group winners and a pre-tournament outright pool; no answer existed when the question was asked, so the evaluation is leakage-free by construction rather than by filtering, and the frozen archive holds 4,494 scored predictions. What the tournament establishes is a set of behaviours the six systems share. On match outcome they average 63.9%, level with backing the bookmaker's favourite -- which is in fact what they usually do. They agree with one another far more often than they are right, so a majority vote adds nothing. They under-commit to draws and to goals, and crowd their scoreline picks onto a single prototypical result. Accuracy tracks how lopsided a fixture is rather than how much is known about it: it collapses in the closest ties, where the dossiers are richest, while questions about the tournament as a whole are answered well. On this task the current generation of frontier systems is not sharply differentiated: the standings hold up at the top and the bottom across the run and churn in the middle, and the margins stay narrow throughout. The briefing dossiers, fixtures and official results are released as a benchmark, together with the scoring code.

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment in long-horizon TIR scenarios. On-policy self-distillation offers denser signals through teacher branches with privileged context, but existing approaches typically derive such context from ground-truth answers or retrieved skills, which may not reflect the states actually visited by the agent. Moreover, token-level supervision fails to capture the turn-level structure of tool interactions. To address this, we propose TurnSight, a turn-level hindsight self-distillation framework that derives supervision directly from execution-conditioned hindsight. It then constructs multiple hindsight views with different lookahead horizons and selects reliable supervision through cross-horizon directional agreement. Finally, the selected hindsight signal is normalized across sibling rollouts and used to adaptively modulate RL advantages while preserving their original optimization direction. Extensive experiments on three benchmarks demonstrate the effectiveness of TurnSight. Our codes are available at https://github.com/quchangle1/TurnSight.

Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines, and learned skills across sessions. Yet whether retained experience actually improves them over time has not been systematically tested. We introduce PAST-Bench, a benchmark designed to isolate this question. Each agent runs through ordered sequences of fresh-session tasks under matched conditions that turn retained experience on and off. It spans 26 scenarios and 204 episodes across memory, procedural reuse, information gathering, and update. We report both later-task gains and whether those gains follow the intended save, retrieve, and update pathway. Across seven base models and four agent frameworks, improvement is real but uneven across capabilities. Agents with the same headline gain can differ markedly in whether that gain is supported by evidence of the intended pathway. Guided by these findings, we develop Hermes+, which extends Hermes with five targeted interventions across stages of the agent loop. Hermes+ raises the average gain from retained experience and provides clearer pathway evidence, with its strongest improvement on tasks requiring outdated state to be replaced, although the effect remains capability- and model-dependent. Together, PAST-Bench and Hermes+ provide an evaluation and diagnostic foundation for studying how persistent agents can progress from retaining experience to systematically improving through it. Code: https://github.com/Gen-Verse/PAST-Bench

Large language models can solve substantially harder reasoning problems with more inference-time compute. The term "test-time scaling," however, now covers diverse inference algorithms that extend deliberation along a single trajectory, sample completed candidates and aggregate them through voting or verification, or search over unfinished partial states. These algorithms differ in their statistical structure, compute accounting, and failure modes. Treating these procedures as interchangeable under a single scalar "budget," or reporting accuracy without the inference protocol that produced it, makes results difficult to compare across studies. We develop a systematic account of test-time scaling along three axes. First, we formalize test-time scaling as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish three structural regimes: single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the evaluated object as the entire inference system and develop evaluation principles that separate end-to-end system performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and prescribe protocol-matched reporting of compute and uncertainty. Third, we specify reproducibility requirements for inference protocols, distinguishing exact replay from distributional reproducibility and identifying the artifacts needed to support each. We also organize the open-weight reasoning ecosystem by model-side and interface mechanisms, apply these principles to broad-knowledge, symbolic-reasoning, and competition-mathematics benchmarks, and assemble over 2 billion full reasoning traces for release with progressively richer verifier and token-level signals.