Providing external information beyond the query is a common way to improve large language model (LLM) inference, as seen in in-context learning (ICL), retrieval-augmented generation (RAG), and memory-based methods (Mem). Existing work increasingly suggests that useful external information should compensate for the gap between LLMs and user queries, implicitly assuming that stronger gap-correction can lead to better performance. In this paper, we challenge this assumption and show that more gap-corrective external information can instead harm performance. We analyze this phenomenon in Transformers through the lens of reasoning error, defined as the difference between the predicted answer vector and the ground-truth answer vector. Our analysis reveals that external information can be not only under-corrective but also over-corrective, where excessively strong correction increases reasoning error. We further show that the induced correction vector is jointly determined by attention and the complementarity between the external information and the query, and derive conditions under which information is properly corrective in both direction and magnitude. Experiments on four mainstream LLMs and seven reasoning benchmarks across ICL, RAG, and Mem validate our theory and support a theory-guided information selection method.
Figures & tables
Figure 1: An illustration of different types of information, where green denotes properly corrective information, and red denotes under-corrective or over-corrective information. In the lower-right corner of each part, the blue vector represents the baseline error vector, while the green and red vectors represent the information-induced correction vectors. The cyan and purple vectors denote the corresponding composed vectors, respectively. The composed vector for properly corrective information is shorter than the baseline error vector, showing lower error. In contrast, the composed vectors for under-corrective or over-corrective information are longer than the baseline error vector.
Finding
Evidence
Experiment
Over-corrective external information could harm performance.
Theorem 1
§ 4.2
Corrective strength of external information is governed by both attention and complementarity between query and information.
Corollary 2 , Corollary 4
§ 4.3
The proper-corrective external information should satisfy a joint direction-magnitude requirement ρ(t,x)>2∥e(x)∥∥g(t,x)∥ .
Corollary 3 , Corollary 5
§ 4.4
Table 1: The main findings and corresponding evidence of this paper.
Figure 3
Figure 4: Error change with respect to the information-corrective vector g(t,x) , where each point in the figure represents one instance. The length from the axis origin to each point represents 2∥e(x)∥∥g(t,x)∥ , and the angle with the positive x -axis represents arccosρ(t,x) . Blue and red indicate that the error change is smaller or larger than 0 , and darker colors indicate larger absolute values. The gray region corresponds to 2∥e(x)∥∥g(t,x)∥<ρ(t,x) .
Figure 5Table 6
Model
Method
MATH
CHAMP
TQA
MMLU
GPQA
NQ
FinQA
Llama-3.1-8B
Zero
46.0
14.1
26.8
63.8
31.3
38.5
48.6
BM25
46.4
14.4
27.0
64.1
31.8
38.9
49.0
DC Suzgun et al. (2025)
47.0
14.8
27.6
64.6
32.3
39.3
49.5
ExpRAG Wei et al. (2025a)
47.0
14.8
27.6
64.7
31.8
39.4
49.7
ReMem Wei et al. (2025a)
47.2
15.2
27.7
64.9
32.3
39.8
49.6
Ours
49.6±0.7
18.1±2.3
30.4±0.5
67.1±0.4
34.8±3.1
41.7±0.9
52.1±1.5
Table 4: Performance of Mem with top- 1 selection of five runs. Zero denotes inference without information. TQA denotes TheoremQA, MMLU denotes MMLU-Redux, and NQ denotes NaturalQuestions. The best performance of each setting is marked in bold .
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Prompt of Inference
{task_definition} --- Here are some information to help you solve the given task: ‘‘‘ {information} ‘‘‘ --- Based on the above information, solve the following question: {question}
Appendix
Table 5: The prompt of inference during our experiments. {information} can be replaced with demonstrations, retrieved knowledge, and memory.
Prompt of Thinker using Memory Evolution
You are an expert meta-reasoning coach for AI agents. You will receive a SINGLE data sample from an agent’s past experience. This sample may include: a task description, the agent’s thoughts/actions/observations, and whether the attempt succeeded or failed. Your job is to distill this ONE example into ONE HIGH-LEVEL INSIGHT: a general rule or strategy that would be useful for similar future tasks. ================ INPUT SAMPLE ================ [Task description] {TASK_DESCRIPTION} [Episode log / trajectory] {EPISODE_LOG} ============================================= Please now produce exactly ONE INSIGHT that: - Is HIGH-LEVEL and GENERAL. - It should be a reusable principle that could help on future, similar tasks. - It should focus on better thinking and acting (e.g., how to search, plan, verify, explore). - Is NOT tied to superficial details of this specific sample. - Do NOT mention specific names, URLs, IDs, or exact numbers from the log unless they are essential to the principle. - Do NOT describe the episode step-by-step. - Do NOT say “In this trial…” or “In this example…”. Instead, phrase it as “When solving tasks like this, you should…”. - Captures what the agent should DO DIFFERENTLY or KEEP DOING. - If the attempt failed: focus on what change of strategy would most likely prevent this failure. - If the attempt succeeded: focus on what strategy made it robust and should be repeated. Output format (VERY IMPORTANT): INSIGHT: - <one single sentence (or two short sentences) stating the general rule> Do NOT output anything else. No explanations, no justification, no bullet list of multiple rules. Just the single “INSIGHT:” line as specified above.
Appendix
Table 6: The prompt of Thinker during memory evolution.
Dataset
Method
Llama-3.1-8B
Qwen3-8B
MATH
ICL
33.8%
39.4%
Mem
38.3%
25.5%
CHAMP
RAG
27.0%
23.9%
Mem
34.4%
36.7%
TheoremQA
RAG
26.6%
26.0%
Mem
28.8%
29.5%
Appendix
Table 7: The relative error ( gold_value∣gold_value - pred_value∣ ) of the trained predictor under different settings.
Predictor Source
Target Model
ICL
RAG
Mem
Avg.
Llama-3.1-8B
Llama-3.1-8B
49.7
49.8
42.0
45.6
Qwen3-8B
Llama-3.1-8B
48.8
48.6
40.9
44.5
Qwen3-8B
Qwen3-8B
75.1
64.6
62.9
66.1
Llama-3.1-8B
Qwen3-8B
74.5
63.3
62.0
65.2
Appendix
Table 8: Cross-model transfer of the learned correction-direction predictor. Values are average downstream accuracy (%) for each external-information setting.
Target Dataset
In-Domain
Other Datasets
MATH
69.6
68.3
CHAMP
34.1
33.1
TQA
49.7
49.0
MMLU-Redux
68.5
68.0
GPQA
53.9
53.0
NQ
61.9
60.5
Appendix
Table 9: Cross-dataset generalization of the learned correction-direction predictor. Values are downstream accuracy (%). “Other Datasets” denotes the cross-dataset training setting.
Model
Setting
No-filter
Predicted
Oracle
Llama-3.1-8B
ICL
48.6
49.7
54.0
RAG
48.1
49.8
55.2
Mem
40.8
42.0
46.3
Average
44.3
45.6
50.1
Qwen3-8B
ICL
74.2
75.1
79.0
RAG
63.2
64.6
69.8
Appendix
Table 10: Effect of correction-direction estimation on downstream accuracy (%).
Model
Variant
MATH
GPQA
FinQA
Llama-3.1-8B
Direction Filter Only
49.2
45.6
52.7
Strength Proxy Only
48.5
44.3
52.2
Both
49.6
46.5
53.1
Qwen3-8B
Direction Filter Only
89.6
65.5
69.0
Strength Proxy Only
89.0
65.3
68.2
Both
90.2
65.7
69.5
Appendix
Table 11: Ablation of the two selection components in ICL. Values are accuracy (%).
Model
Variant
CHAMP
TQA
NQ
Llama-3.1-8B
Direction Filter Only
26.6
41.9
78.9
Strength Proxy Only
25.9
40.9
77.8
Both
27.4
42.5
79.4
Qwen3-8B
Direction Filter Only
47.0
66.7
78.2
Strength Proxy Only
45.8
65.5
76.2
Both
47.8
67.2
78.9
Appendix
Table 12: Ablation of the two selection components in RAG. Values are accuracy (%).
Model
Variant
MATH
CHAMP
TQA
MMLU
GPQA
NQ
FinQA
Llama-3.1-8B
Direction Filter Only
48.9
17.6
29.8
66.5
34.3
41.1
51.4
Strength Proxy Only
47.9
15.8
28.2
65.8
32.8
40.3
50.5
Both
49.6
18.1
30.4
67.1
34.8
41.7
52.1
Qwen3-8B
Direction Filter Only
88.7
42.2
57.8
69.1
67.6
47.1
62.7
Strength Proxy Only
88.5
41.3
56.5
67.4
66.7
45.8
61.9
Both
89.0
43.0
58.7
69.9
68.7
47.5
63.4
Appendix
Table 13: Ablation of the two selection components in the memory setting. Values are accuracy (%).
Model
Method
MATH
GPQA
FinQA
Llama-3.1-8B
Linear-Time DS
48.8
44.4
52.4
AutoSelect
49.1
45.9
52.8
DiSP
49.3
46.0
52.7
Ours
49.6
46.5
53.1
Qwen3-8B
Linear-Time DS
89.3
65.5
67.8
AutoSelect
89.8
65.6
69.3
Appendix
Table 14: Comparison with stronger ICL/context-selection baselines under the same candidate pool and computational budget. Values are accuracy (%).
Model
Method
CHAMP
TQA
NQ
Llama-3.1-8B
Qwen3-Embedding-0.6B
25.9
41.2
78.1
Qwen3-Reranker-0.6B
26.6
41.6
78.7
jina-reranker-v3
26.8
42.0
78.9
Ours
27.4
42.5
79.4
Qwen3-8B
Qwen3-Embedding-0.6B
46.7
66.2
77.5
Qwen3-Reranker-0.6B
47.2
66.5
78.1
Appendix
Table 15: Comparison with stronger RAG/context-selection baselines under the same candidate pool and computational budget. Values are accuracy (%).
Model
Method
MATH
CHAMP
TQA
MMLU
GPQA
NQ
FinQA
Llama-3.1-8B
MemOS
48.6
16.6
28.7
65.6
33.7
40.7
51.0
SimpleMem
48.9
17.2
29.6
66.5
34.0
41.6
51.8
MemGuide
49.4
17.7
29.4
66.8
34.0
41.2
51.7
Ours
49.6
18.1
30.4
67.1
34.8
41.7
52.1
Qwen3-8B
MemOS
88.5
41.5
57.2
68.4
67.4
46.5
62.7
SimpleMem
88.9
42.3
57.6
69.2
67.8
46.8
62.3
Appendix
Table 16: Comparison with stronger memory-selection baselines under the same candidate pool and computational budget. Values are accuracy (%).
LLMs often exhibit Aha moments such as self-correction after tokens like "Wait," yet the underlying mechanism remains unclear. Standard LLMs collapse mainly through silent divergence, where trajectories drift from the correct answer yet remain locally coherent, so no explicit error triggers reactive self-correction. We introduce an information-theoretic framework that separates reasoning into procedural advancement and epistemic verbalization, the token-level externalization of uncertainty, and prove that sporadic verbalization restores convergence toward the correct answer even without explicit error triggers. Empirically, a minimal doubt cue recovers failed trajectories, and small-scale SFT suffices to instill or suppress this capability, suggesting that strong reasoning hinges less on an extraordinary inner mechanism than on the linguistic habit of externalizing uncertainty. Our framework recasts reasoning as strategic information allocation under uncertainty, offering a new lens for understanding and advancing LLM reasoning.
Jeonghye Kim, Xufang Luo, Minbeom Kim +3
Microsoft Research · KAIST · Seoul National University
LLMs reliably correct false claims when presented in isolation, yet when the same claims are embedded in task-oriented requests, they often comply rather than correct. We term this failure mode correction suppression and construct a benchmark of 300 false premises to systematically evaluate it across eight models. Suppression rates range from 19% to 90%, with four models exceeding 80%, establishing correction suppression as a prevalent and severe phenomenon. Mechanistic analysis reveals that suppression is not a knowledge failure: the model registers the error internally but task context diverts early-layer attention from the false claim as output intent crystallizes toward compliance at middle layers. We characterize this as knowing but not correcting -- suppression occurs at response selection rather than knowledge encoding. Guided by this mechanism, we propose two training-free interventions. Correction Direction Steering (CDS) estimates a correction-compliance direction from matched pairs and injects it at middle layers before output intent crystallizes. Dynamic Payload Amplification (DPA) localizes payload tokens via attention divergence between early and late layers and amplifies their representation at the final layer, requiring no calibration data. Experiments on Qwen3.5-9B and LLaMA3.1-8B show both methods substantially improve factual strictness. CDS achieves the highest correction rate on Qwen3.5-9B (0%→58.2%). DPA is the only method that preserves or improves reasoning capability on both models. These findings introduce factual strictness -- the willingness to uphold accuracy against contextual pressures -- as a new dimension of model reliability.
Zixuan Chen, Hao Lin, Zizhe Chen +6
1Shanghai Jiao Tong University · 2The Chinese University of Hong Kong · 3Ant Group
Large language models struggle to reliably identify and correct errors in their own reasoning. Unlike prior training-based approaches that fine-tune an LLM to generate revisions or feedback, we learn where to intervene in failed rollouts and determine how to redirect them. To choose where to intervene, we train a 130K-parameter MLP head on a frozen encoder using only final-answer correctness. For redirection, we find that under a fixed generation budget, the tested inline injections fail to reliably redirect reasoning, whereas presenting the same instruction in a new user turn changes the model's course. The resulting correction framework, UPTURN (User-Prompt TURN injection at a learned position), requires no LLM fine-tuning: it retains a selected prefix, injects an instruction as a new user turn, and regenerates the suffix. Across four reasoners and six math benchmarks, UPTURN improves average accuracy in every setting over a 32-rollout baseline. It outperforms self-correction and reflection-token injection baselines under matched correction budgets, transfers across reasoners and to GPQA-Diamond without retraining, and complements GRPO and self-correction methods.
Jiwon Jeon, Myungsik Cho, Jeonghye Kim +1
Korea Advanced Institute of Science and Technology (KAIST)