Reasoning Collapse in Machine Learning

• Reasoning collapse is the breakdown of a model’s ability to maintain structured, input-dependent reasoning, revealed as abrupt accuracy cliffs.
• It manifests in diverse forms such as prediction bias, reasoning-trace, template, and evidence collapse across unimodal and multimodal tasks.
• Mitigation strategies include adding structural loss constraints, enhancing causal supervision, and applying task-aware monitoring to preserve reasoning fidelity.

Reasoning collapse is a term used in recent machine-learning literature for a family of failures in which a model’s apparent reasoning ceases to track task structure, input variation, or evidential requirements. The literature does not treat it as a single canonical pathology. Instead, it appears as abrupt accuracy cliffs under increasing complexity, degenerate constant-answer policies after fine-tuning, loss of explicit reasoning traces despite plausible final answers, collapse from interventional to associational reasoning, input-agnostic template reuse in agentic RL, progressive loss of visual grounding during multimodal generation, and semantic narrowing in self-evolving curricula. A recurring theme is that aggregate answer accuracy often obscures the failure, so collapse is typically diagnosed through structural, causal, or trajectory-level analyses rather than answer-only metrics (Deshmukh et al., 6 May 2026, Twist et al., 20 May 2026, Wang et al., 7 Apr 2026, Raghu et al., 5 Apr 2026).

1. Conceptual scope and terminology

Recent work uses “reasoning collapse” as an umbrella label for related but distinct breakdowns. In causal fine-tuning, it can mean prediction bias collapse, where outputs become almost independent of the input graph structure and converge to a fixed class such as always “Yes” or always “No” (Deshmukh et al., 6 May 2026). In explicit reasoning models, it can mean reasoning-trace collapse, defined as “the progressive loss of a model's ability to produce structurally valid reasoning traces during fine-tuning” (Twist et al., 20 May 2026). In causal inference tasks framed by Pearl’s hierarchy, it can mean Rung Collapse, where a higher-rung query is answered using lower-rung reasoning (Chang, 12 Feb 2026). In agentic RL, it can mean template collapse, where reasoning remains superficially diverse within a prompt but becomes input-agnostic across prompts (Wang et al., 7 Apr 2026). In multimodal reasoning, the term broadens to evidence collapse, a progressive loss of attention to the image regions supporting the answer (Raghu et al., 5 Apr 2026). In self-evolving systems, the analogous phenomenon is curriculum collapse, where the question generator narrows into semantically redundant problem families (Mishra, 3 Mar 2026).

Mode	Defining symptom	Representative source
Prediction bias collapse	$>95\%$ of predictions are one class	(Deshmukh et al., 6 May 2026)
Reasoning-trace collapse	Valid traces fall while answers remain plausible	(Twist et al., 20 May 2026)
Rung Collapse	Interventional query answered with associational evidence	(Chang, 12 Feb 2026)
Template collapse	Stable entropy, low input dependence	(Wang et al., 7 Apr 2026)
Evidence collapse	Visual evidence attention decays during reasoning	(Raghu et al., 5 Apr 2026)
Curriculum collapse	Question distribution narrows semantically across iterations	(Mishra, 3 Mar 2026)

A more abstract formulation treats language generation itself as a lossy projection from an internal intention space $\mathcal{I}$ into an external language space $\mathcal{L}$, calling this intention collapse. In that framing, reasoning collapse is the special case in which useful intermediate structure is available pre-verbalization but is compressed, hidden, or lost in the emitted sequence (Vera, 3 Jan 2026).

2. Empirical manifestations across domains

One major manifestation is an abrupt complexity threshold. On controllable puzzle environments, Large Reasoning Models exhibit three regimes: low complexity where standard models can match or outperform them, medium complexity where explicit thinking helps, and high complexity where both thinking and non-thinking models collapse to near-zero accuracy (Shojaee et al., 7 Jun 2025). The same pattern appears in DeepRD, where performance on graph connectivity and natural-language proof planning drops abruptly once lookahead exceeds the training-like regime; even chain graphs with $B=1$ eventually collapse, indicating that branching ambiguity is not the only source of failure (Rameshkumar et al., 25 Oct 2025). In symbolic logic, “Logical Phase Transitions” describes a stable pre-transition plateau followed by an abrupt collapse within critical LoCM intervals, with performance then stabilizing near the random-guess baseline (Zhang et al., 6 Jan 2026). In interactive scientific reasoning, XDomainBench reports a systematic decline as composition order $k$ rises from 1 to 4, with average turn-level Recall dropping from 38.7% to 27.1% and session-level average Recall from 0.205 to 0.114 (Zhiren et al., 14 May 2026).

A second manifestation is degenerate supervised adaptation. Fine-tuning Gemma 3 270M Instruct on causal reasoning without semantic loss yields a 100% collapse rate across five model variants and 200,000+ evaluation samples: the transitivity baseline outputs “Yes” for all 10,000 samples and averages 27.7% accuracy, while the d-separation baseline outputs almost always “No,” attains 73.9% accuracy because the dataset is No-heavy, but has F1 $=7.6\%$ and very poor recall (Deshmukh et al., 6 May 2026). In explicit reasoning models, standard fine-tuning on science QA can suppress reasoning structure while leaving answer accuracy apparently healthy: on Chemistry, Qwen3-8B and Nemotron-7B reach 0% valid reasoning in some settings while pass@1 still rises to roughly 50–57%, and for Qwen3-8B on GSM8K, Rpass@1 remains very high at about 98% even while VR falls to 58% and pass@1 drops to 77% (Twist et al., 20 May 2026).

A third manifestation is distributional drift in the reasoning medium itself. Under GRPO on translated reasoning datasets, multilingual models can revert their chain-of-thought to the dominant pre-training language. On translated GSM8K with Llama, Ukrainian accuracy rises to 70.9% while the target-language word ratio falls to 0.3%; Korean improves to 61.3% with a target-language word ratio of 82.4%; Chinese drift is milder at 94.1% target-language word ratio (Park et al., 6 Jun 2025). In multimodal reasoning, evidence collapse is pervasive across all 9 model–dataset cells studied: $\Delta V>0$ for 94.3%–100% of samples, $\Delta A_{\text{bbox}}>0$ in all 9 cells, and the relative loss of evidence attention ranges from 53.3% or 56.8% on the low end up to 90.8% for GLM-4.6V-Flash on HallusionBench (Raghu et al., 5 Apr 2026). In self-evolving reasoning, curriculum collapse appears as a “diversity illusion”: under R-Zero, active clusters drop from 89 to 65, normalized entropy falls to 0.53, Gini rises to 0.90, and one cluster accumulates 951 questions (Mishra, 3 Mar 2026).

3. Proposed mechanisms

A central explanation is objective mismatch. In causal reasoning fine-tuning, standard cross-entropy permits shortcut solutions aligned with label imbalance, and without structural penalties the optimizer can settle into degenerate minima that ignore causal topology (Deshmukh et al., 6 May 2026). In explicit reasoning models, the mismatch is format-level: when downstream targets contain explanations but no explicit reasoning delimiters, the model can minimize loss by emitting the answer directly, creating pressure toward response-only behavior (Twist et al., 20 May 2026). In agentic VLM training, outcome-only RL fails for the same reason: rewards supervise final actions but not thought quality, so reasoning becomes state-irrelevant, incomplete, and eventually action-invalid (Wei et al., 11 Mar 2025).

A second mechanism is causal supervision failure. “Right for the Wrong Reasons” argues that autoregressive training provides no gradient signal distinguishing $P(Y\mid X)$ from $P(Y\mid do(X))$, formalizing the resulting pathology as Rung Collapse. When outcome-based learning rewards correct answers obtained through incorrect causal models, the agent becomes entrenched in flawed reasoning, termed Aleatoric Entrenchment (Chang, 12 Feb 2026). CausalT5K operationalizes the same family of failures benchmark-wise: models answer higher-rung queries with associational evidence, drift under adversarial pressure, or fail to produce Wise Refusals when evidence is underdetermined (Geng et al., 9 Feb 2026).

A third mechanism is signal degradation during RL. RAGEN-2 decomposes reasoning quality into within-input diversity and cross-input distinguishability, then argues that low reward variance weakens task gradients while KL and entropy regularizers remain prompt-agnostic. The result is template collapse: reasoning can retain stable entropy yet lose input dependence because regularization dominates when the signal-to-noise ratio is low (Wang et al., 7 Apr 2026). A closely related process appears in multimodal RL agents, where GTR identifies rapid loss of thought diversity and state relevance when RL is driven solely by action outcomes (Wei et al., 11 Mar 2025).

A fourth mechanism is geometric or representational degeneration. Seq-VCR attributes failures on multi-step arithmetic partly to intermediate-layer representation collapse, measured by a sharp entropy drop and concentration of variance into a few directions (Arefin et al., 2024). RED identifies an analogous phenomenon in efficiently distilled LLMs: width-reducing projection matrices initialized randomly cause eRank collapse, with uneven singular values leading to token indistinguishability and severe loss of multi-step reasoning despite strong general-benchmark scores (He et al., 28 May 2026).

A fifth mechanism concerns interface and resource constraints, though the literature is divided on how far these explain collapse. The “agentic gap” commentary argues that apparent reasoning cliffs in puzzles are confounded by static text-only evaluation, tool-use restrictions, context-window recall issues, and output-generation limits; for Tower of Hanoi, it gives token cost $\mathcal{I}$0 and argues that a 64k-token cap makes failure near $\mathcal{I}$1 mathematically expected (Khan et al., 23 Jun 2025). By contrast, an environment-interface study on Tower of Hanoi reports that collapse is not delayed or eradicated by external state management and can even occur at lower complexity, with growing divergence from both optimal and uniformly random policies (Su et al., 12 Oct 2025). This suggests that interface constraints are important in some settings, but do not exhaust the phenomenon.

4. Diagnostics and measurement

A consistent finding is that accuracy alone is inadequate. In causal fine-tuning, prediction bias collapse is defined by the existence of a fixed prediction $\mathcal{I}$2 such that

$\mathcal{I}$3

This definition makes collapse a distributional property of outputs rather than a raw error rate, which is why the d-separation baseline can report 73.9% accuracy while being almost always “No” (Deshmukh et al., 6 May 2026).

For explicit reasoning models, structural metrics separate answer correctness from reasoning validity. The core quantities are VR (valid reasoning rate), ER (empty reasoning rate), MR (missing reasoning rate), TR (truncated reasoning rate), and Rpass@1, defined as pass@1 computed only over generations that contain valid reasoning traces. The diagnostic signature of reasoning-trace collapse is precisely the case where VR falls sharply while Rpass@1 stays high (Twist et al., 20 May 2026).

For agentic RL, entropy is no longer sufficient. RAGEN-2 formalizes the decomposition

$\mathcal{I}$4

where $\mathcal{I}$5 captures within-input diversity and $\mathcal{I}$6 captures cross-input dependence. Template collapse is the regime with high $\mathcal{I}$7 but low $\mathcal{I}$8: reasoning looks diverse but is not actually conditioned on the prompt (Wang et al., 7 Apr 2026).

For logical reasoning, LoCM provides a complexity coordinate: $\mathcal{I}$9 This makes phase-transition boundaries measurable and distinguishes logical complexity from surface length or premise count (Zhang et al., 6 Jan 2026).

Multimodal settings require trajectory-level grounding metrics. Evidence-collapse work identifies architecture-specific grounding layers via AUROC against bounding-box annotations, then tracks total visual attention $\mathcal{L}$0, evidence attention $\mathcal{L}$1, and full-response entropy $\mathcal{L}$2. Under cross-dataset transfer, full-response entropy is the strongest text-only baseline: it is negative in all 9 cells, significant in all 9, and outperforms answer-time entropy in 8/9 cells, whereas answer-time entropy is significant only for GLM × HallusionBench (Raghu et al., 5 Apr 2026).

In trustworthy causal reasoning, CausalT5K extends diagnosis beyond accuracy by decomposing performance into Utility and Safety, and by measuring Bad Flip Rate and Dissonance Rate. Its reported detection rates around 77–91% alongside dissonance around 48–55% show that a model can detect a trap yet still fail to correct its answer, a collapse mode invisible to aggregate correctness (Geng et al., 9 Feb 2026).

5. Mitigation strategies

One major line of mitigation introduces structural constraints into the loss. In causal reasoning, semantic loss augments cross-entropy with graph-based logical consistency: $\mathcal{L}$3 with dynamic scheduling

$\mathcal{L}$4

where $\mathcal{L}$5 and $\mathcal{L}$6. On Gemma 3 270M Instruct, this yields 70.4% accuracy on transitivity and 68.6% on d-separation, with a 42.7% average improvement over the collapsed transitivity baseline and balanced performance on 1,000 adversarial structural samples (Deshmukh et al., 6 May 2026). In causal belief revision, ERM penalizes epistemic error independently of task success and reportedly recovers 53–59% of entrenched errors where outcome-level feedback fails (Chang, 12 Feb 2026).

A second line preserves reasoning structure during adaptation. For reasoning-trace collapse, simple loss masking can substantially mitigate failure without teacher-generated reasoning traces, while teacher distillation with GPT-5-mini works very well for Qwen3-8B, Llama-R1-8B, and often Nemotron-7B, though not uniformly for Olmo-3-7B (Twist et al., 20 May 2026). In logical reasoning, Neuro-Symbolic Curriculum Tuning combines adaptive NL–FOL alignment with complexity-aware curriculum optimization and reports average gains of +1.26 under naive prompting and +3.95 under CoT across five benchmarks (Zhang et al., 6 Jan 2026).

A third line targets representation geometry. Seq-VCR regularizes intermediate representations with variance and covariance penalties and, together with dummy pause tokens, reaches 99.5% exact match on $\mathcal{L}$7 integer multiplication, where same-size baselines and pause-only or Seq-VCR-only variants remain at 0% (Arefin et al., 2024). RED stabilizes projection-matrix geometry in efficient distillation by activation-aware initialization and substantially improves reasoning averages while preserving general ability; for example, at the ~4B scale, RED-4B reaches reasoning average about 0.33 versus LRC-4B about 0.18, and GSM8K about 0.49 versus about 0.34 (He et al., 28 May 2026).

A fourth line improves training dynamics in RL and self-evolution. GTR couples PPO on actions with SFT on corrected thoughts from an automated corrector, raising Points24 success from 2.5% for RL4VLM to 17.5% and delivering 3–5 times higher task success rates than SoTA models on the hardest visual tasks reported (Wei et al., 11 Mar 2025). RAGEN-2 proposes SNR-Aware Filtering, selecting high-variance prompts per iteration; MI-family metrics correlate with final performance much more strongly than entropy, with Trajectory MI-ZScore about $\mathcal{L}$8 versus entropy metrics around $\mathcal{L}$9 to $B=1$0 in Spearman correlation (Wang et al., 7 Apr 2026). Prism prevents curriculum collapse by persistent cross-iteration semantic coverage over $B=1$1 embedding clusters and a ZPD gate; it activates 107 of 128 clusters, raises normalized entropy to 0.83, lowers Gini to 0.66, and achieves the highest Pass@1 on six of seven math benchmarks, including gains of +3.98 over R-Zero on AMC and +3.68 on Minerva Math (Mishra, 3 Mar 2026).

A fifth line uses task-aware monitoring rather than blanket fusion rules. In multimodal reasoning, a targeted vision veto applied at 90% coverage and 5% veto rate reduces selective risk by up to 1.94 percentage points on MMMU_Pro × GLM and helps other high-risk transfers, while harming MathVista transfers where visual disengagement can be benign. The sign of the entropy–vision interaction term $B=1$2 predicts this direction exactly across the reported transfers (Raghu et al., 5 Apr 2026).

6. Interpretation, controversies, and broader significance

A central controversy concerns whether reasoning collapse reflects an intrinsic limit or an evaluation artifact. Shojaee et al.’s “reasoning cliff” was interpreted as an intrinsic scaling limitation of CoT reasoning, but the subsequent commentary reframed it as an agentic gap produced by a static, text-only interface lacking tools, verification, and procedural offloading (Khan et al., 23 Jun 2025). The environment-interface study on deterministic games directly challenged the sufficiency of that reframing by showing that access to move_disk(from_peg, to_peg) and end_game() does not delay or eliminate collapse, and that looping, repeated continuations, and mode-like policy behavior persist (Su et al., 12 Oct 2025). The literature therefore does not support a single verdict; rather, it indicates that some apparent cliffs are amplified by interface design, while others survive stronger agency and externalized state.

A second controversy concerns what exactly has collapsed. Some papers study answer accuracy under controlled complexity, others the validity of explicit traces, others causal rung fidelity, evidence grounding, reasoning language, or cross-iteration curriculum diversity. This suggests that “reasoning collapse” is best treated as a family resemblance term. The common denominator is structural decoupling: the model’s internal or external reasoning no longer remains appropriately coupled to the task’s latent constraints (Twist et al., 20 May 2026, Geng et al., 9 Feb 2026).

The practical significance is that models can remain useful over the head of real-world distributions while failing sharply in the tail. DeepRD reports that most real-world graph and proof examples fall within today’s success regime, yet long tails in lookahead and branch count expose substantial failure potential (Rameshkumar et al., 25 Oct 2025). XDomainBench makes the same point for interactive scientific workflows: increasing composition order and trajectory volatility produce error accumulation, reasoning breaks, and domain confusion that do not appear in single-turn, single-domain settings (Zhiren et al., 14 May 2026).

The strongest general implication is methodological. Static audit policies, single-score benchmarks, and answer-only evaluation are repeatedly shown to be insufficient. CausalT5K describes a Four-Quadrant Control Landscape in which static audit policies universally fail, while multiple other papers show that confusion matrices, prediction distributions, reasoning-conditioned accuracy, mutual-information proxies, grounding trajectories, and coverage dynamics are necessary to diagnose failure modes that aggregate accuracy misses (Geng et al., 9 Feb 2026). In contemporary usage, reasoning collapse therefore names not a singular bug, but a broad research program: identifying where reasoning degrades, determining whether the degradation is objective-, representation-, or interface-induced, and designing training and evaluation procedures that preserve input-dependent, structure-faithful, and evidence-grounded inference.