← All model news

StartLux-Decision lets decision models decide whether to reason, but probabilities distort after selection

According to a November 4, 2026 report from Jiqizhixin, StartLux, together with the University of Hong Kong and the Southern University of Science and Technology, proposed StartLux-Decision: Instant and AutoThink are two sets of weights that share an option-probability readout, and AutoThink generates reasoning only when the highest option probability is below 0.9. The paper also presents RLCD² training and the DecisionPivot evaluation, along with a code repository and Hugging Face weight addresses.

Published Updated

What changed

This is not a chat model. StartLux-Decision formats a request as 'input state + question + candidate options' and takes the positions corresponding to letters such as A, B, and C from the model's final-layer logits, applying temperature scaling and softmax to obtain a probability for each option, without adding a separate classification head; Instant performs only one forward pass, while AutoThink reads once before and once after reasoning. The two differ in attention direction: Instant removes the causal mask in Full Attention layers so tokens within the prompt interact bidirectionally, while Linear Attention layers still retain causal computation; AutoThink keeps causal attention to support generating reasoning traces. The paper explicitly states that the two are not the same set of released weights and cannot substitute for each other. Under the paper's default evaluation configuration, AutoThink first computes an initial distribution p₀; if the highest option probability reaches 0.9, it returns directly; only below 0.9 does it generate a reasoning trace, read p₁ from the end of the trace, and weight the final output about 75% toward the post-reasoning distribution. Training has three stages: full-parameter supervised fine-tuning; adapter training that completes the divergence in attention direction; and then reasoning supervision and RLCD² reinforcement learning for AutoThink. RLCD²'s reward consists of the probability-score increment given by a frozen scoring model plus an overlong penalty, with a soft length budget of 3072 tokens and a generation cap of 4096 tokens. The accompanying DecisionPivot evaluation contains 13 categories of program-generated tasks, 600 questions each, for a total of 7,800 questions. Most form contrastive question pairs that change only one detail sufficient to change the answer; the core metric is pairwise accuracy, where both questions in a pair are answered correctly. The paper reports 59.9%, 59.9%, and 86.9% for 4B, 9B, and 27B AutoThink, respectively, versus 41.5%, 47.5%, and 75.4% for the corresponding Instant models; in the same table, Decision 3.0 d3 scores 71.3%, Perplexity Decider v1.1 scores 70.1%, and GLM-5.3 with a low-reasoning-effort configuration scores 82.2%. These are all figures and run settings reported by the paper itself.

What it means for developers

The impact on integration is more direct than the scores. First, the output is a probability rather than text: Instant does not do autoregressive generation and cannot be used as a chat or summarization model; if the deployment environment cannot obtain per-option logits or probabilities, the core interface of this model family is unusable. Second, latency becomes bimodal: the paper measured, under a single H200, BF16, 256-token prompt, and single-request setting, a median latency of 24–52 milliseconds for a single Instant readout, but once AutoThink triggers reasoning, end-to-end latency also includes trace generation and a second readout, and p50 will hide the long tail. Third, the semantics of the probability depend on the downstream policy: using the 9B as a verifier, the paper found that answers given 80%–90% confidence in a single prediction were actually 92% correct, but after selecting the highest score among 32 candidates, the same confidence interval dropped to 78%, because the selection operation changes the conditional distribution of the returned predictions. Fourth, local deployment has a memory threshold: the BF16 version of 27B exceeds available memory on an M5 Pro with 64GB unified memory and requires 8-bit or 4-bit quantization to run. Fifth, task gains are uneven: the improvements for 4B and 9B are concentrated in tasks requiring item-by-item computation, such as filtering, target-state verification, and scheduling conflicts, while a strong single-pass prediction model still has the advantage in policy interpretation, clause changes, and dependency handling.

Integration and evaluation

The following is conditional analysis based on the paper's report, not advice given by the paper. Judgment one: whether to choose Instant or AutoThink, and where to set the threshold. The change that can be grounded in the report is AutoThink's confidence-based gating—on DecisionPivot, the 27B answers about 39% of questions directly with about 96% single-question accuracy, while the remaining about 61% enter reasoning and accuracy rises from about 57% to about 93%. If your workload consists of judgments with few options and fully given context, such as request routing or judging retrieval results, Instant's value is eliminating generation; if the rules often contain exception clauses or threshold boundaries, choose AutoThink and sweep the threshold yourself. The prerequisite is that the deployment side can read per-option probabilities. A testable criterion: measure pairwise accuracy on your own contrastive question pairs, while also tracking the share of requests that enter reasoning and p95 latency, sweep the threshold from 0 to 1, and find the segment where accuracy stops rising while token use continues to increase. Judgment two: how large to set the reasoning length budget. The change that can be grounded in the report is RLCD²'s length penalty—the 4B pairwise accuracy drops from 60.4% to 52.9% when the penalty takes effect from 512 tokens onward, and returns to 59.9% when relaxed to 3072 tokens; removing the penalty entirely yields 61.7%, but about one-third of reasoning in this evaluation cannot terminate on its own within the generation cap. The choice is to align the soft budget with the upper edge of your own reasoning-length distribution, rather than copying 3072. The prerequisite is being able to record trace length and whether it terminates on its own. A testable criterion: the self-termination rate and token consumption per question, not accuracy alone; if the self-termination rate is markedly low, first check whether questions that genuinely need long computation are being suppressed along with the rest. Judgment three: whether probability calibration should be redone per policy. The change that can be grounded in the report is policy-aware calibration—across 184 settings covering best-of-N, reasoning gating, cascades, and filter-then-reason, the signed confidence–accuracy gap shrank from -11.7 to +20.5 percentage points to -2.7 to +2.1 percentage points. The choice is to fit calibration on the prediction that is actually returned, after the actual filtering or routing, rather than calibrating individual candidates and reusing that directly. The prerequisite is labeled calibration data and the ability to record the number of candidates, the average candidate score, or the confidence that triggered routing. A testable criterion: within fixed confidence intervals, compare predicted confidence with actual accuracy, and count the share of erroneous passes at the automatic-execution threshold you set; once the policy changes, both the calibration mapping and the acceptance threshold must be re-measured.

Limits and open questions

The Instant and AutoThink models of StartLux-Decision (0.8B to 27B) are currently not in KimiSeek's model catalog; the report also does not mention a hosted API, and the availability of code and weights should be based on the project's GitHub repository and Hugging Face page, which are project channels rather than something provided by this site. The source does not state under what license the weights are released: downloadable weights do not equal an open-source license, and the permitted scope for commercial use, redistribution, and modification depends on the model card and license file; it cannot be inferred from 'the weights have been released.' The numbers themselves also have boundaries. The public portion of DI 0.3 corresponds to only 20% of the full leaderboard score, and the paper states that it was used for this study's data recipe and model selection, not as an independent test set; AutoThink's score on that leaderboard cannot be taken as evidence that it entered the single-prediction leaderboard that excludes reasoning traces. On 2,986 held-out samples, the change intervals for final decision accuracy at all three scales include zero; this comparison best supports improvements in reasoning termination and generation efficiency, not a broad increase in accuracy. The paper itself also notes: 48.0% is AutoThink's score with reasoning turned off and cannot be conflated with Instant's 75.4%; after the calibration mapping is transferred to unseen datasets, there remains residual overconfidence of up to about 6.1 percentage points; a fixed 0.9 threshold does not equal per-question optimal allocation of computation; latency figures from a single H200 and single request cannot be extrapolated to concurrent serving; and evidence for cross-task-family transfer remains limited.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

决策模型何时需要思考?StartLux-Decision 的按需推理与概率校准

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.