NCP Nearly Halves Training Tokens as T1 Reaches 64%

Featured

01 Training: Gate the Tokens, Preserve the Language

Negative training faces a token-selection problem: faulty reasoning and ordinary language share many tokens, so applying an unlearning objective indiscriminately could damage foundational language capabilities. NSD uses dynamic adaptive gating to identify reasoning-critical tokens automatically and confine gradient updates to behavioral flaws while preserving linguistic priors.

The negative signal comes from the model itself. NSD asks it to generate a problem-specific flawed persona, such as a “careless reasoner,” then pushes the student’s distribution away from that self-generated negative teacher. This requires neither reference answers nor external supervision.

The central design choice is where the negative signal changes the model. By separating the generation of flawed behavior from the selection of tokens that receive gradient updates, NSD aims to suppress reasoning errors without broadly penalizing the language used to express them. The abstract reports consistent gains over conventional on-policy self-distillation and other label-free, self-bootstrapping reinforcement learning baselines, but full-task results and language retention still require confirmation from the paper.

Key takeaways:

• Models can create conditions such as a “careless reasoner” for each problem without external negative examples.
• Adaptive gating concentrates updates on reasoning-critical tokens, reducing damage to language priors.
• Teams should test both negative-condition quality and language retention before adopting negative training.

Source: Negative Self-Distillation: Learning to Reason by Avoiding Flaws

02 Retrieval: Four Vectors, Full Pages on Demand

Visual document retrieval systems may not need to store thousands of vectors for every page. Those vectors occupy a high-dimensional space but cluster near a manifold with only five to six intrinsic dimensions.

GLIE stores a small set of representative vectors for initial retrieval. It then reconstructs full vector sets for top-ranked pages and performs exact reranking. Compression becomes a question of when to reconstruct evidence, rather than which evidence to discard.

Four vectors per page retain nearly 80% of the uncompressed system’s nDCG@5 on ViDoRe v1. The previous best post-processing method retained about 70% at the same budget. Its 415,000-parameter decoder trains on 1,000 pages in under three GPU minutes without retraining the encoder.

Key takeaways:

• Generative reconstruction can replace blunt vector pruning when page embeddings lie near a low-dimensional manifold.
• Four vectors per page retain nearly 80% of the original retrieval quality, beating prior post-processing at the same budget.
• Storage-constrained teams should measure reconstruction latency and recall on production data before adoption.

Source: Generative Late-Interaction Embeddings for Visual Document Retrieval

03 Architecture: Beyond the Next Token

Next Concept Prediction adds explicit supervision across multiple tokens without abandoning token-by-token generation. It builds a concept vocabulary from hidden states, predicts those concepts, then feeds them back into the token layer.

An 8.9B model reaches OLMo-3-7B’s final pretraining loss with 51.3% of the training tokens. After full training, its average downstream score is 2.45 points higher. GSM8K improves by 5.99 points.

Compatibility matters here. NCP preserves standard token-level autoregressive generation, although training adds a concept vocabulary, a dedicated Concept Module, and a joint objective. Its concept space can also support lightweight domain adaptation and assisted speculative decoding. Teams still need to assess concept-vocabulary construction, extra modules, joint-training complexity, and stability across domains.

Key takeaways:

• Dual-granularity prediction adds concept-level supervision while preserving autoregressive generation.
• Early results suggest better training efficiency and reasoning, but cross-scale model comparisons are not conclusive.
• Teams upgrading existing pretraining stacks should assess integration cost and cross-domain stability.

Source: NCP-ArchPreview Technical Report: Moving Towards Latent Space Language Models Through Next Concept Prediction

04 Agent: Exact Replay Narrows Training Drift

Small discrepancies between training and sampling accumulate across hundreds of agent steps. T1 addresses this with TITO, which trains on the exact token IDs produced during sampling. It also repairs drift at turn boundaries.

For MoE models, T1 records each expert selected for every token at every MoE layer. Training then replays those routes exactly. TITO and route replay, called R3, cut the train-inference log-probability gap from 0.021 to 0.013. They also eliminate token drift within loss-bearing regions.

T1 runs task-provided validators in real shells and assigns dense rewards based on passed checks. Terminal-Bench 2.1 success rises from 43.8% to 64.0%. It reaches 27.9% on Long-Horizon Terminal Bench, ahead of GPT-5.4 and GLM-5.1. Reproduction will be expensive given the 122B model, cloud sandboxes, and complex training infrastructure.

Key takeaways:

• T1 suggests that exact replay of sampled token IDs and MoE routes can reduce train-inference drift in long-horizon agent training.
• TITO and R3 reduce the train-inference log-probability gap from 0.021 to 0.013 and eliminate drift in loss-bearing regions.
• Isolated training data still yields 64.0% on Terminal-Bench 2.1 and 27.9% on Long-Horizon Terminal Bench.

Source: T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks